diff --git a/diffusers/pipelines/pag/pipeline_pag_controlnet_sd.py b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd.py new file mode 100644 index 0000000000000000000000000000000000000000..3a88272f24f43d09762f518de0a63d09797f68fb --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd.py @@ -0,0 +1,1348 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ControlNetModel, ImageProjection, MultiControlNetModel, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, is_compiled_module, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> # !pip install opencv-python transformers accelerate + >>> from diffusers import AutoPipelineForText2Image, ControlNetModel, UniPCMultistepScheduler + >>> from diffusers.utils import load_image + >>> import numpy as np + >>> import torch + + >>> import cv2 + >>> from PIL import Image + + >>> # download an image + >>> image = load_image( + ... "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png" + ... ) + >>> image = np.array(image) + + >>> # get canny image + >>> image = cv2.Canny(image, 100, 200) + >>> image = image[:, :, None] + >>> image = np.concatenate([image, image, image], axis=2) + >>> canny_image = Image.fromarray(image) + + >>> # load control net and stable diffusion v1-5 + >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16) + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", + ... controlnet=controlnet, + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + + >>> # speed up diffusion process with faster scheduler and memory optimization + >>> # remove following line if xformers is not installed + >>> pipe.enable_xformers_memory_efficient_attention() + + >>> pipe.enable_model_cpu_offload() + + >>> # generate image + >>> generator = torch.manual_seed(0) + >>> image = pipe( + ... "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting", + ... guidance_scale=7.5, + ... generator=generator, + ... image=canny_image, + ... pag_scale=10, + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionControlNetPAGPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion with ControlNet guidance. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + controlnet ([`ControlNetModel`] or `list[ControlNetModel]`): + Provides additional conditioning to the `unet` during the denoising process. If you set multiple + ControlNets as a list, the outputs from each ControlNet are added together to create one combined + additional conditioning. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetModel | list[ControlNetModel] | tuple[ControlNetModel] | MultiControlNetModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + pag_applied_layers: str | list[str] = "mid", + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + if isinstance(controlnet, (list, tuple)): + controlnet = MultiControlNetModel(controlnet) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + image, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if not isinstance(image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + + # When `image` is a nested list: + # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]]) + elif any(isinstance(i, list) for i in image): + transposed_image = [list(t) for t in zip(*image)] + if len(transposed_image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: if you pass`image` as a list of list, each sublist must have the same length as the number of controlnets, but the sublists in `image` got {len(transposed_image)} images and {len(self.controlnet.nets)} ControlNets." + ) + for image_ in transposed_image: + self.check_image(image_, prompt, prompt_embeds) + elif len(image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets." + ) + else: + for image_ in image: + self.check_image(image_, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if isinstance(controlnet_conditioning_scale, list): + if any(isinstance(i, list) for i in controlnet_conditioning_scale): + raise ValueError( + "A single batch of varying conditioning scale settings (e.g. [[1.0, 0.5], [0.2, 0.8]]) is not supported at the moment. " + "The conditioning scale must be fixed across the batch." + ) + elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( + self.controlnet.nets + ): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + if isinstance(self.controlnet, MultiControlNetModel): + if len(control_guidance_start) != len(self.controlnet.nets): + raise ValueError( + f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + guess_mode: bool = False, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. When `prompt` is a list, and if a list of images is passed for a single + ControlNet, each will be paired with each prompt in the `prompt` list. This also applies to multiple + ControlNets, where a list of image lists can be passed to batch for each prompt and each ControlNet. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + guess_mode (`bool`, *optional*, defaults to `False`): + The ControlNet encoder tries to recognize the content of the input image even if you remove all + prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + global_pool_conditions = ( + controlnet.config.global_pool_conditions + if isinstance(controlnet, ControlNetModel) + else controlnet.nets[0].config.global_pool_conditions + ) + guess_mode = guess_mode or global_pool_conditions + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare image + if isinstance(controlnet, ControlNetModel): + image = self.prepare_image( + image=image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=guess_mode, + ) + height, width = image.shape[-2:] + elif isinstance(controlnet, MultiControlNetModel): + images = [] + + # Nested lists as ControlNet condition + if isinstance(image[0], list): + # Transpose the nested image list + image = [list(t) for t in zip(*image)] + + for image_ in image: + image_ = self.prepare_image( + image=image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=guess_mode, + ) + + images.append(image_) + + image = images + height, width = image[0].shape[-2:] + else: + assert False + + # 5. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + self._num_timesteps = len(timesteps) + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6.5 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Add image embeds for IP-Adapter + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + added_cond_kwargs = ( + {"image_embeds": ip_adapter_image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + controlnet_prompt_embeds = prompt_embeds + + # 7.2 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + images = image if isinstance(image, list) else [image] + for i, single_image in enumerate(images): + if self.do_classifier_free_guidance: + single_image = single_image.chunk(2)[0] + + if self.do_perturbed_attention_guidance: + single_image = self._prepare_perturbed_attention_guidance( + single_image, single_image, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + single_image = torch.cat([single_image] * 2) + single_image = single_image.to(device) + images[i] = single_image + + image = images if isinstance(image, list) else images[0] + + # 8. Denoising loop + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + is_unet_compiled = is_compiled_module(self.unet) + is_controlnet_compiled = is_compiled_module(self.controlnet) + is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1") + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Relevant thread: + # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428 + if ( + torch.cuda.is_available() + and (is_unet_compiled and is_controlnet_compiled) + and is_torch_higher_equal_2_1 + ): + torch._inductor.cudagraph_mark_step_begin() + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # controlnet(s) inference + control_model_input = latent_model_input + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + if guess_mode and self.do_classifier_free_guidance: + # Inferred ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + empty_device_cache() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_inpaint.py b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..98221e4c30ba6e215b2bfb4395734f8e2e7deccc --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_inpaint.py @@ -0,0 +1,1552 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This model implementation is heavily inspired by https://github.com/haofanwang/ControlNet-for-Diffusers/ + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ControlNetModel, ImageProjection, MultiControlNetModel, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> # !pip install transformers accelerate + >>> import cv2 + >>> from diffusers import AutoPipelineForInpainting, ControlNetModel, DDIMScheduler + >>> from diffusers.utils import load_image + >>> import numpy as np + >>> from PIL import Image + >>> import torch + + >>> init_image = load_image( + ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png" + ... ) + >>> init_image = init_image.resize((512, 512)) + + >>> generator = torch.Generator(device="cpu").manual_seed(1) + + >>> mask_image = load_image( + ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png" + ... ) + >>> mask_image = mask_image.resize((512, 512)) + + + >>> def make_canny_condition(image): + ... image = np.array(image) + ... image = cv2.Canny(image, 100, 200) + ... image = image[:, :, None] + ... image = np.concatenate([image, image, image], axis=2) + ... image = Image.fromarray(image) + ... return image + + + >>> control_image = make_canny_condition(init_image) + + >>> controlnet = ControlNetModel.from_pretrained( + ... "lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16 + ... ) + >>> pipe = AutoPipelineForInpainting.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", + ... controlnet=controlnet, + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + + >>> pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config) + >>> pipe.enable_model_cpu_offload() + + >>> # generate image + >>> image = pipe( + ... "a handsome man with ray-ban sunglasses", + ... num_inference_steps=20, + ... generator=generator, + ... eta=1.0, + ... image=init_image, + ... mask_image=mask_image, + ... control_image=control_image, + ... pag_scale=0.3, + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class StableDiffusionControlNetPAGInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, + PAGMixin, +): + r""" + Pipeline for image inpainting using Stable Diffusion with ControlNet guidance. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + > [!TIP] > This pipeline can be used with checkpoints that have been specifically fine-tuned for inpainting > + ([stable-diffusion-v1-5/stable-diffusion-inpainting](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-inpainting)) + as well as > default text-to-image Stable Diffusion checkpoints > + ([stable-diffusion-v1-5/stable-diffusion-v1-5](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5)). + Default text-to-image > Stable Diffusion checkpoints might be preferable for ControlNets that have been fine-tuned + on those, such as > + [lllyasviel/control_v11p_sd15_inpaint](https://huggingface.co/lllyasviel/control_v11p_sd15_inpaint). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + controlnet ([`ControlNetModel`] or `list[ControlNetModel]`): + Provides additional conditioning to the `unet` during the denoising process. If you set multiple + ControlNets as a list, the outputs from each ControlNet are added together to create one combined + additional conditioning. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetModel | list[ControlNetModel] | tuple[ControlNetModel] | MultiControlNetModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + pag_applied_layers: str | list[str] = "mid", + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + if isinstance(controlnet, (list, tuple)): + controlnet = MultiControlNetModel(controlnet) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + image, + mask_image, + height, + width, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if height is not None and height % 8 != 0 or width is not None and width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + # `prompt` needs more sophisticated handling when there are multiple + # conditionings. + if isinstance(self.controlnet, MultiControlNetModel): + if isinstance(prompt, list): + logger.warning( + f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}" + " prompts. The conditionings will be fixed across the prompts." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if not isinstance(image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + + # When `image` is a nested list: + # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]]) + elif any(isinstance(i, list) for i in image): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif len(image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets." + ) + + for image_ in image: + self.check_image(image_, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if isinstance(controlnet_conditioning_scale, list): + if any(isinstance(i, list) for i in controlnet_conditioning_scale): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( + self.controlnet.nets + ): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + if isinstance(self.controlnet, MultiControlNetModel): + if len(control_guidance_start) != len(self.controlnet.nets): + raise ValueError( + f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_inpaint.StableDiffusionControlNetInpaintPipeline.prepare_control_image + def prepare_control_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + crops_coords, + resize_mode, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + else: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 1.0, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + controlnet_conditioning_scale: float | list[float] = 0.5, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, + `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, NumPy array or tensor representing an image batch to be used as the starting point. For both + NumPy array and PyTorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a + list or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a NumPy array or + a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, + `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, NumPy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a NumPy array or PyTorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for PyTorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for NumPy array, it would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, + W, 1)`, or `(H, W)`. + control_image (`torch.Tensor`, `PIL.Image.Image`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, + `list[list[torch.Tensor]]`, or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 0.5): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + control_image, + mask_image, + height, + width, + output_type, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if padding_mask_crop is not None: + height, width = self.image_processor.get_default_height_width(image, height, width) + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + device = self._execution_device + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare control image + if isinstance(controlnet, ControlNetModel): + control_image = self.prepare_control_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + crops_coords=crops_coords, + resize_mode=resize_mode, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + elif isinstance(controlnet, MultiControlNetModel): + control_images = [] + + for control_image_ in control_image: + control_image_ = self.prepare_control_image( + image=control_image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + crops_coords=crops_coords, + resize_mode=resize_mode, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + + control_images.append(control_image_) + + control_image = control_images + else: + assert False + + # 4.1 Preprocess mask and image - resizes image and mask w.r.t height and width + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + mask = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + masked_image = init_image * (mask < 0.5) + _, _, height, width = init_image.shape + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps=num_inference_steps, strength=strength, device=device + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + self._num_timesteps = len(timesteps) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask, masked_image_latents = self.prepare_mask_latents( + mask, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # 7.1 Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + + # 7.2 Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.3 Prepare embeddings + # ip-adapter + if ip_adapter_image_embeds is not None: + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + added_cond_kwargs = ( + {"image_embeds": ip_adapter_image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # control image + control_images = control_image if isinstance(control_image, list) else [control_image] + for i, single_control_image in enumerate(control_images): + if self.do_classifier_free_guidance: + single_control_image = single_control_image.chunk(2)[0] + + if self.do_perturbed_attention_guidance: + single_control_image = self._prepare_perturbed_attention_guidance( + single_control_image, single_control_image, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + single_control_image = torch.cat([single_control_image] * 2) + single_control_image = single_control_image.to(device) + control_images[i] = single_control_image + + control_image = control_images if isinstance(control_image, list) else control_images[0] + controlnet_prompt_embeds = prompt_embeds + + # 7.4 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + # 7.5 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # controlnet(s) inference + control_model_input = latent_model_input + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=control_image, + conditioning_scale=cond_scale, + guess_mode=False, + return_dict=False, + ) + + # concat latents, mask, masked_image_latents in the channel dimension + if num_channels_unet == 9: + first_dim_size = latent_model_input.shape[0] + # Ensure mask and masked_image_latents have the right dimensions + if mask.shape[0] < first_dim_size: + repeat_factor = (first_dim_size + mask.shape[0] - 1) // mask.shape[0] + mask = mask.repeat(repeat_factor, 1, 1, 1)[:first_dim_size] + if masked_image_latents.shape[0] < first_dim_size: + repeat_factor = ( + first_dim_size + masked_image_latents.shape[0] - 1 + ) // masked_image_latents.shape[0] + masked_image_latents = masked_image_latents.repeat(repeat_factor, 1, 1, 1)[:first_dim_size] + # Perform the concatenation + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # Predict noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + empty_device_cache() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl.py b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..9ba42eb0352f4216bb3c2a89f2d4c8a751126fec --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl.py @@ -0,0 +1,1621 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from diffusers.utils.import_utils import is_invisible_watermark_available + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ControlNetModel, ImageProjection, MultiControlNetModel, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput +from .pag_utils import PAGMixin + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + + +from ...utils import is_torch_xla_available + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> # !pip install opencv-python transformers accelerate + >>> from diffusers import AutoPipelineForText2Image, ControlNetModel, AutoencoderKL + >>> from diffusers.utils import load_image + >>> import numpy as np + >>> import torch + + >>> import cv2 + >>> from PIL import Image + + >>> prompt = "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting" + >>> negative_prompt = "low quality, bad quality, sketches" + + >>> # download an image + >>> image = load_image( + ... "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png" + ... ) + + >>> # initialize the models and pipeline + >>> controlnet_conditioning_scale = 0.5 # recommended for good generalization + >>> controlnet = ControlNetModel.from_pretrained( + ... "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16 + ... ) + >>> vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... controlnet=controlnet, + ... vae=vae, + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + >>> pipe.enable_model_cpu_offload() + + >>> # get canny image + >>> image = np.array(image) + >>> image = cv2.Canny(image, 100, 200) + >>> image = image[:, :, None] + >>> image = np.concatenate([image, image, image], axis=2) + >>> canny_image = Image.fromarray(image) + + >>> # generate image + >>> image = pipe( + ... prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, image=canny_image, pag_scale=0.3 + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLControlNetPAGPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]): + Second frozen text-encoder + ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + tokenizer_2 ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + controlnet ([`ControlNetModel`] or `list[ControlNetModel]`): + Provides additional conditioning to the `unet` during the denoising process. If you set multiple + ControlNets as a list, the outputs from each ControlNet are added together to create one combined + additional conditioning. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings should always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to + watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no + watermarker is used. + """ + + # leave controlnet out on purpose because it iterates with unet + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + "image_encoder", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetModel | list[ControlNetModel] | tuple[ControlNetModel] | MultiControlNetModel, + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + pag_applied_layers: str | list[str] = "mid", # ["down.block_2", "up.block_1.attentions_0"], "mid" + ): + super().__init__() + + if isinstance(controlnet, (list, tuple)): + controlnet = MultiControlNetModel(controlnet) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + image, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + negative_pooled_prompt_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # `prompt` needs more sophisticated handling when there are multiple + # conditionings. + if isinstance(self.controlnet, MultiControlNetModel): + if isinstance(prompt, list): + logger.warning( + f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}" + " prompts. The conditionings will be fixed across the prompts." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if not isinstance(image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + + # When `image` is a nested list: + # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]]) + elif any(isinstance(i, list) for i in image): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif len(image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets." + ) + + for image_ in image: + self.check_image(image_, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if isinstance(controlnet_conditioning_scale, list): + if any(isinstance(i, list) for i in controlnet_conditioning_scale): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( + self.controlnet.nets + ): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + if isinstance(self.controlnet, MultiControlNetModel): + if len(control_guidance_start) != len(self.controlnet.nets): + raise ValueError( + f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2` + and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, pooled text embeddings are generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt + weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input + argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned containing the output images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + image, + None, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + negative_pooled_prompt_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + # 3.1 Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt, + prompt_2, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 3.2 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare image + if isinstance(controlnet, ControlNetModel): + image = self.prepare_image( + image=image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + height, width = image.shape[-2:] + elif isinstance(controlnet, MultiControlNetModel): + images = [] + + for image_ in image: + image_ = self.prepare_image( + image=image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + + images.append(image_) + + image = images + height, width = image[0].shape[-2:] + else: + assert False + + # 5. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + self._num_timesteps = len(timesteps) + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6.1 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + # 7.2 Prepare added time ids & embeddings + if isinstance(image, list): + original_size = original_size or image[0].shape[-2:] + else: + original_size = original_size or image.shape[-2:] + target_size = target_size or (height, width) + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + images = image if isinstance(image, list) else [image] + for i, single_image in enumerate(images): + if self.do_classifier_free_guidance: + single_image = single_image.chunk(2)[0] + + if self.do_perturbed_attention_guidance: + single_image = self._prepare_perturbed_attention_guidance( + single_image, single_image, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + single_image = torch.cat([single_image] * 2) + single_image = single_image.to(device) + images[i] = single_image + + image = images if isinstance(image, list) else images[0] + + if ip_adapter_image_embeds is not None: + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + add_text_embeds = self._prepare_perturbed_attention_guidance( + add_text_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + add_time_ids = self._prepare_perturbed_attention_guidance( + add_time_ids, negative_add_time_ids, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + controlnet_prompt_embeds = prompt_embeds + controlnet_added_cond_kwargs = added_cond_kwargs + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + is_unet_compiled = is_compiled_module(self.unet) + is_controlnet_compiled = is_compiled_module(self.controlnet) + is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1") + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Relevant thread: + # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428 + if ( + torch.cuda.is_available() + and (is_unet_compiled and is_controlnet_compiled) + and is_torch_higher_equal_2_1 + ): + torch._inductor.cudagraph_mark_step_begin() + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # controlnet(s) inference + control_model_input = latent_model_input + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=cond_scale, + guess_mode=False, + added_cond_kwargs=controlnet_added_cond_kwargs, + return_dict=False, + ) + + if ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = ip_adapter_image_embeds + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..bfe45761d5a0a2dc877c5e79c7d428e6ce614389 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_controlnet_sd_xl_img2img.py @@ -0,0 +1,1685 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from diffusers.utils.import_utils import is_invisible_watermark_available + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ControlNetModel, ImageProjection, MultiControlNetModel, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, is_compiled_module, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput +from .pag_utils import PAGMixin + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + + +from ...utils import is_torch_xla_available + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> # pip install accelerate transformers safetensors diffusers + + >>> import torch + >>> import numpy as np + >>> from PIL import Image + + >>> from transformers import DPTFeatureExtractor, DPTForDepthEstimation + >>> from diffusers import ControlNetModel, StableDiffusionXLControlNetPAGImg2ImgPipeline, AutoencoderKL + >>> from diffusers.utils import load_image + + >>> depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda") + >>> feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas") + >>> controlnet = ControlNetModel.from_pretrained( + ... "diffusers/controlnet-depth-sdxl-1.0-small", + ... variant="fp16", + ... use_safetensors="True", + ... torch_dtype=torch.float16, + ... ) + >>> vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16) + >>> pipe = StableDiffusionXLControlNetPAGImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... controlnet=controlnet, + ... vae=vae, + ... variant="fp16", + ... use_safetensors=True, + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + >>> pipe.enable_model_cpu_offload() + + + >>> def get_depth_map(image): + ... image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda") + ... with torch.no_grad(), torch.autocast("cuda"): + ... depth_map = depth_estimator(image).predicted_depth + + ... depth_map = torch.nn.functional.interpolate( + ... depth_map.unsqueeze(1), + ... size=(1024, 1024), + ... mode="bicubic", + ... align_corners=False, + ... ) + ... depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True) + ... depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True) + ... depth_map = (depth_map - depth_min) / (depth_max - depth_min) + ... image = torch.cat([depth_map] * 3, dim=1) + ... image = image.permute(0, 2, 3, 1).cpu().numpy()[0] + ... image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8)) + ... return image + + + >>> prompt = "A robot, 4k photo" + >>> image = load_image( + ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" + ... "/kandinsky/cat.png" + ... ).resize((1024, 1024)) + >>> controlnet_conditioning_scale = 0.5 # recommended for good generalization + >>> depth_image = get_depth_map(image) + + >>> images = pipe( + ... prompt, + ... image=image, + ... control_image=depth_image, + ... strength=0.99, + ... num_inference_steps=50, + ... controlnet_conditioning_scale=controlnet_conditioning_scale, + ... ).images + >>> images[0].save(f"robot_cat.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class StableDiffusionXLControlNetPAGImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, + PAGMixin, +): + r""" + Pipeline for image-to-image generation using Stable Diffusion XL with ControlNet guidance. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + controlnet ([`ControlNetModel`] or `list[ControlNetModel]`): + Provides additional conditioning to the unet during the denoising process. If you set multiple ControlNets + as a list, the outputs from each ControlNet are added together to create one combined additional + conditioning. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the + config of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + "image_encoder", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "add_neg_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + controlnet: ControlNetModel | list[ControlNetModel] | tuple[ControlNetModel] | MultiControlNetModel, + scheduler: KarrasDiffusionSchedulers, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + pag_applied_layers: str | list[str] = "mid", # ["mid"], ["down.block_1", "up.block_0.attentions_0"] + ): + super().__init__() + + if isinstance(controlnet, (list, tuple)): + controlnet = MultiControlNetModel(controlnet) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False + ) + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl_img2img.StableDiffusionXLControlNetImg2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + image, + strength, + num_inference_steps, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + if num_inference_steps is None: + raise ValueError("`num_inference_steps` cannot be None.") + elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # `prompt` needs more sophisticated handling when there are multiple + # conditionings. + if isinstance(self.controlnet, MultiControlNetModel): + if isinstance(prompt, list): + logger.warning( + f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}" + " prompts. The conditionings will be fixed across the prompts." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if not isinstance(image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + + # When `image` is a nested list: + # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]]) + elif any(isinstance(i, list) for i in image): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif len(image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets." + ) + + for image_ in image: + self.check_image(image_, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if isinstance(controlnet_conditioning_scale, list): + if any(isinstance(i, list) for i in controlnet_conditioning_scale): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( + self.controlnet.nets + ): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + if isinstance(self.controlnet, MultiControlNetModel): + if len(control_guidance_start) != len(self.controlnet.nets): + raise ValueError( + f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_image + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.prepare_image + def prepare_control_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.prepare_latents + def prepare_latents( + self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + latents_mean = latents_std = None + if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None: + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1) + if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None: + latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + empty_device_cache() + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + if latents_mean is not None and latents_std is not None: + latents_mean = latents_mean.to(device=device, dtype=dtype) + latents_std = latents_std.to(device=device, dtype=dtype) + init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std + else: + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.8, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + controlnet_conditioning_scale: float | list[float] = 0.8, + guess_mode: bool = False, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The initial image will be used as the starting point for the image generation process. Can also accept + image latents as `image`, if passing latents directly, it will not be encoded again. + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If + the type is specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also + be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height + and/or width are passed, `image` is resized according to them. If multiple ControlNets are specified in + init, images must be passed as a list such that each element of the list can be correctly batched for + input to a single controlnet. + height (`int`, *optional*, defaults to the size of control_image): + The height in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to the size of control_image): + The width in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original unet. If multiple ControlNets are specified in init, you can set the + corresponding scale as a list. + guess_mode (`bool`, *optional*, defaults to `False`): + In this mode, the ControlNet encoder will try best to recognize the content of the input image even if + you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the controlnet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the controlnet stops applying. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple` containing the output images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + control_image, + strength, + num_inference_steps, + None, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + # 3.1 Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt, + prompt_2, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 3.2 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare image and controlnet_conditioning_image + image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + + if isinstance(controlnet, ControlNetModel): + control_image = self.prepare_control_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + height, width = control_image.shape[-2:] + elif isinstance(controlnet, MultiControlNetModel): + control_images = [] + + for control_image_ in control_image: + control_image_ = self.prepare_control_image( + image=control_image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + + control_images.append(control_image_) + + control_image = control_images + height, width = control_image[0].shape[-2:] + else: + assert False + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + self._num_timesteps = len(timesteps) + + # 6. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + True, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Create tensor stating which controlnets to keep + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + # 7.2 Prepare added time ids & embeddings + if isinstance(control_image, list): + original_size = original_size or control_image[0].shape[-2:] + else: + original_size = original_size or control_image.shape[-2:] + target_size = target_size or (height, width) + + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + add_text_embeds = pooled_prompt_embeds + + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + control_images = control_image if isinstance(control_image, list) else [control_image] + for i, single_image in enumerate(control_images): + if self.do_classifier_free_guidance: + single_image = single_image.chunk(2)[0] + + if self.do_perturbed_attention_guidance: + single_image = self._prepare_perturbed_attention_guidance( + single_image, single_image, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + single_image = torch.cat([single_image] * 2) + single_image = single_image.to(device) + control_images[i] = single_image + + control_image = control_images if isinstance(control_image, list) else control_images[0] + + if ip_adapter_image_embeds is not None: + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + add_text_embeds = self._prepare_perturbed_attention_guidance( + add_text_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + add_time_ids = self._prepare_perturbed_attention_guidance( + add_time_ids, add_neg_time_ids, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + controlnet_prompt_embeds = prompt_embeds + controlnet_added_cond_kwargs = added_cond_kwargs + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # controlnet(s) inference + control_model_input = latent_model_input + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=control_image, + conditioning_scale=cond_scale, + guess_mode=False, + added_cond_kwargs=controlnet_added_cond_kwargs, + return_dict=False, + ) + + if ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = ip_adapter_image_embeds + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # If we do sequential model offloading, let's offload unet and controlnet + # manually for max memory savings + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.unet.to("cpu") + self.controlnet.to("cpu") + empty_device_cache() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_hunyuandit.py b/diffusers/pipelines/pag/pipeline_pag_hunyuandit.py new file mode 100644 index 0000000000000000000000000000000000000000..a443a19bd9523cfea67346c9088f5fbf3ab518e6 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_hunyuandit.py @@ -0,0 +1,968 @@ +# Copyright 2025 HunyuanDiT Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import torch +from transformers import BertModel, BertTokenizer, CLIPImageProcessor, T5EncoderModel, T5Tokenizer + +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKL, HunyuanDiT2DModel +from ...models.attention_processor import PAGCFGHunyuanAttnProcessor2_0, PAGHunyuanAttnProcessor2_0 +from ...models.embeddings import get_2d_rotary_pos_embed +from ...pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from ...schedulers import DDPMScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import AutoPipelineForText2Image + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "Tencent-Hunyuan/HunyuanDiT-v1.2-Diffusers", + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... pag_applied_layers=[14], + ... ).to("cuda") + + >>> # prompt = "an astronaut riding a horse" + >>> prompt = "一个宇航员在骑马" + >>> image = pipe(prompt, guidance_scale=4, pag_scale=3).images[0] + ``` +""" + +STANDARD_RATIO = np.array( + [ + 1.0, # 1:1 + 4.0 / 3.0, # 4:3 + 3.0 / 4.0, # 3:4 + 16.0 / 9.0, # 16:9 + 9.0 / 16.0, # 9:16 + ] +) +STANDARD_SHAPE = [ + [(1024, 1024), (1280, 1280)], # 1:1 + [(1024, 768), (1152, 864), (1280, 960)], # 4:3 + [(768, 1024), (864, 1152), (960, 1280)], # 3:4 + [(1280, 768)], # 16:9 + [(768, 1280)], # 9:16 +] +STANDARD_AREA = [np.array([w * h for w, h in shapes]) for shapes in STANDARD_SHAPE] +SUPPORTED_SHAPE = [ + (1024, 1024), + (1280, 1280), # 1:1 + (1024, 768), + (1152, 864), + (1280, 960), # 4:3 + (768, 1024), + (864, 1152), + (960, 1280), # 3:4 + (1280, 768), # 16:9 + (768, 1280), # 9:16 +] + + +def map_to_standard_shapes(target_width, target_height): + target_ratio = target_width / target_height + closest_ratio_idx = np.argmin(np.abs(STANDARD_RATIO - target_ratio)) + closest_area_idx = np.argmin(np.abs(STANDARD_AREA[closest_ratio_idx] - target_width * target_height)) + width, height = STANDARD_SHAPE[closest_ratio_idx][closest_area_idx] + return width, height + + +def get_resize_crop_region_for_grid(src, tgt_size): + th = tw = tgt_size + h, w = src + + r = h / w + + # resize + if r > 1: + resize_height = th + resize_width = int(round(th / h * w)) + else: + resize_width = tw + resize_height = int(round(tw / w * h)) + + crop_top = int(round((th - resize_height) / 2.0)) + crop_left = int(round((tw - resize_width) / 2.0)) + + return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width) + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class HunyuanDiTPAGPipeline(DiffusionPipeline, PAGMixin): + r""" + Pipeline for English/Chinese-to-image generation using HunyuanDiT and [Perturbed Attention + Guidance](https://huggingface.co/docs/diffusers/en/using-diffusers/pag). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + HunyuanDiT uses two text encoders: [mT5](https://huggingface.co/google/mt5-base) and [bilingual CLIP](fine-tuned by + ourselves) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. We use + `sdxl-vae-fp16-fix`. + text_encoder (`~transformers.BertModel`, `~transformers.CLIPTextModel` | None): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + HunyuanDiT uses a fine-tuned [bilingual CLIP]. + tokenizer (`~transformers.BertTokenizer`, `~transformers.CLIPTokenizer` | None): + A `BertTokenizer` or `CLIPTokenizer` to tokenize text. + transformer ([`HunyuanDiT2DModel`]): + The HunyuanDiT model designed by Tencent Hunyuan. + text_encoder_2 (`T5EncoderModel`): + The mT5 embedder. Specifically, it is 't5-v1_1-xxl'. + tokenizer_2 (`T5Tokenizer`): + The tokenizer for the mT5 embedder. + scheduler ([`DDPMScheduler`]): + A scheduler to be used in combination with HunyuanDiT to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [ + "safety_checker", + "feature_extractor", + "text_encoder_2", + "tokenizer_2", + "text_encoder", + "tokenizer", + ] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "prompt_embeds_2", + "negative_prompt_embeds_2", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: BertModel, + tokenizer: BertTokenizer, + transformer: HunyuanDiT2DModel, + scheduler: DDPMScheduler, + safety_checker: StableDiffusionSafetyChecker | None = None, + feature_extractor: CLIPImageProcessor | None = None, + requires_safety_checker: bool = True, + text_encoder_2: T5EncoderModel | None = None, + tokenizer_2: T5Tokenizer | None = None, + pag_applied_layers: str | list[str] = "blocks.1", # "blocks.16.attn1", "blocks.16", "16", 16 + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + text_encoder_2=text_encoder_2, + ) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + + self.set_pag_applied_layers( + pag_applied_layers, pag_attn_processors=(PAGCFGHunyuanAttnProcessor2_0(), PAGHunyuanAttnProcessor2_0()) + ) + + # Copied from diffusers.pipelines.hunyuandit.pipeline_hunyuandit.HunyuanDiTPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + device: torch.device = None, + dtype: torch.dtype = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + max_sequence_length: int | None = None, + text_encoder_index: int = 0, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + dtype (`torch.dtype`): + torch dtype + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + max_sequence_length (`int`, *optional*): maximum sequence length to use for the prompt. + text_encoder_index (`int`, *optional*): + Index of the text encoder to use. `0` for clip and `1` for T5. + """ + if dtype is None: + if self.text_encoder_2 is not None: + dtype = self.text_encoder_2.dtype + elif self.transformer is not None: + dtype = self.transformer.dtype + else: + dtype = None + + if device is None: + device = self._execution_device + + tokenizers = [self.tokenizer, self.tokenizer_2] + text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = tokenizers[text_encoder_index] + text_encoder = text_encoders[text_encoder_index] + + if max_sequence_length is None: + if text_encoder_index == 0: + max_length = 77 + if text_encoder_index == 1: + max_length = 256 + else: + max_length = max_sequence_length + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_attention_mask = text_inputs.attention_mask.to(device) + prompt_embeds = text_encoder( + text_input_ids.to(device), + attention_mask=prompt_attention_mask, + ) + prompt_embeds = prompt_embeds[0] + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_attention_mask = uncond_input.attention_mask.to(device) + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + attention_mask=negative_prompt_attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds, negative_prompt_embeds, prompt_attention_mask, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.hunyuandit.pipeline_hunyuandit.HunyuanDiTPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + prompt_embeds_2=None, + negative_prompt_embeds_2=None, + prompt_attention_mask_2=None, + negative_prompt_attention_mask_2=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is None and prompt_embeds_2 is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds_2`. Cannot leave both `prompt` and `prompt_embeds_2` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if prompt_embeds_2 is not None and prompt_attention_mask_2 is None: + raise ValueError("Must provide `prompt_attention_mask_2` when specifying `prompt_embeds_2`.") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if negative_prompt_embeds_2 is not None and negative_prompt_attention_mask_2 is None: + raise ValueError( + "Must provide `negative_prompt_attention_mask_2` when specifying `negative_prompt_embeds_2`." + ) + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_embeds_2 is not None and negative_prompt_embeds_2 is not None: + if prompt_embeds_2.shape != negative_prompt_embeds_2.shape: + raise ValueError( + "`prompt_embeds_2` and `negative_prompt_embeds_2` must have the same shape when passed directly, but" + f" got: `prompt_embeds_2` {prompt_embeds_2.shape} != `negative_prompt_embeds_2`" + f" {negative_prompt_embeds_2.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_2: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_2: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + prompt_attention_mask_2: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask_2: torch.Tensor | None = None, + output_type: str = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int, dict], None] + | PipelineCallback + | MultiPipelineCallbacks + | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + guidance_rescale: float = 0.0, + original_size: tuple[int, int] = (1024, 1024), + target_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + use_resolution_binning: bool = True, + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation with HunyuanDiT. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`): + The height in pixels of the generated image. + width (`int`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + negative_prompt_embeds_2 (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds` is passed directly. + prompt_attention_mask_2 (`torch.Tensor`, *optional*): + Attention mask for the prompt. Required when `prompt_embeds_2` is passed directly. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds` is passed directly. + negative_prompt_attention_mask_2 (`torch.Tensor`, *optional*): + Attention mask for the negative prompt. Required when `negative_prompt_embeds_2` is passed directly. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback_on_step_end (`Callable[[int, int], None]`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A callback function or a list of callback functions to be called at the end of each denoising step. + callback_on_step_end_tensor_inputs (`list[str]`, *optional*): + A list of tensor inputs that should be passed to the callback function. If not defined, all tensor + inputs will be passed. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Rescale the noise_cfg according to `guidance_rescale`. Based on findings of [Common Diffusion Noise + Schedules and Sample Steps are Flawed](https://huggingface.co/papers/2305.08891). See Section 3.4 + original_size (`tuple[int, int]`, *optional*, defaults to `(1024, 1024)`): + The original size of the image. Used to calculate the time ids. + target_size (`tuple[int, int]`, *optional*): + The target size of the image. Used to calculate the time ids. + crops_coords_top_left (`tuple[int, int]`, *optional*, defaults to `(0, 0)`): + The top left coordinates of the crop. Used to calculate the time ids. + use_resolution_binning (`bool`, *optional*, defaults to `True`): + Whether to use resolution binning or not. If `True`, the input resolution will be mapped to the closest + standard resolution. Supported resolutions are 1024x1024, 1280x1280, 1024x768, 1152x864, 1280x960, + 768x1024, 864x1152, 960x1280, 1280x768, and 768x1280. It is recommended to set this to `True`. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + height = int((height // 16) * 16) + width = int((width // 16) * 16) + + if use_resolution_binning and (height, width) not in SUPPORTED_SHAPE: + width, height = map_to_standard_shapes(width, height) + height = int(height) + width = int(width) + logger.warning(f"Reshaped to (height, width)=({height}, {width}), Supported shapes are {SUPPORTED_SHAPE}") + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + prompt_embeds_2, + negative_prompt_embeds_2, + prompt_attention_mask_2, + negative_prompt_attention_mask_2, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=self.transformer.dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + max_sequence_length=77, + text_encoder_index=0, + ) + ( + prompt_embeds_2, + negative_prompt_embeds_2, + prompt_attention_mask_2, + negative_prompt_attention_mask_2, + ) = self.encode_prompt( + prompt=prompt, + device=device, + dtype=self.transformer.dtype, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds_2, + negative_prompt_embeds=negative_prompt_embeds_2, + prompt_attention_mask=prompt_attention_mask_2, + negative_prompt_attention_mask=negative_prompt_attention_mask_2, + max_sequence_length=256, + text_encoder_index=1, + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Create image_rotary_emb, style embedding & time ids + grid_height = height // 8 // self.transformer.config.patch_size + grid_width = width // 8 // self.transformer.config.patch_size + base_size = 512 // 8 // self.transformer.config.patch_size + grid_crops_coords = get_resize_crop_region_for_grid((grid_height, grid_width), base_size) + image_rotary_emb = get_2d_rotary_pos_embed( + self.transformer.inner_dim // self.transformer.num_heads, + grid_crops_coords, + (grid_height, grid_width), + device=device, + output_type="pt", + ) + + style = torch.tensor([0], device=device) + + target_size = target_size or (height, width) + add_time_ids = list(original_size + target_size + crops_coords_top_left) + add_time_ids = torch.tensor([add_time_ids], dtype=prompt_embeds.dtype) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + prompt_attention_mask = self._prepare_perturbed_attention_guidance( + prompt_attention_mask, negative_prompt_attention_mask, self.do_classifier_free_guidance + ) + prompt_embeds_2 = self._prepare_perturbed_attention_guidance( + prompt_embeds_2, negative_prompt_embeds_2, self.do_classifier_free_guidance + ) + prompt_attention_mask_2 = self._prepare_perturbed_attention_guidance( + prompt_attention_mask_2, negative_prompt_attention_mask_2, self.do_classifier_free_guidance + ) + add_time_ids = torch.cat([add_time_ids] * 3, dim=0) + style = torch.cat([style] * 3, dim=0) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask]) + prompt_embeds_2 = torch.cat([negative_prompt_embeds_2, prompt_embeds_2]) + prompt_attention_mask_2 = torch.cat([negative_prompt_attention_mask_2, prompt_attention_mask_2]) + add_time_ids = torch.cat([add_time_ids] * 2, dim=0) + style = torch.cat([style] * 2, dim=0) + + prompt_embeds = prompt_embeds.to(device=device) + prompt_attention_mask = prompt_attention_mask.to(device=device) + prompt_embeds_2 = prompt_embeds_2.to(device=device) + prompt_attention_mask_2 = prompt_attention_mask_2.to(device=device) + add_time_ids = add_time_ids.to(dtype=prompt_embeds.dtype, device=device).repeat( + batch_size * num_images_per_prompt, 1 + ) + style = style.to(device=device).repeat(batch_size * num_images_per_prompt) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.transformer.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # expand scalar t to 1-D tensor to match the 1st dim of latent_model_input + t_expand = torch.tensor([t] * latent_model_input.shape[0], device=device).to( + dtype=latent_model_input.dtype + ) + + # predict the noise residual + noise_pred = self.transformer( + latent_model_input, + t_expand, + encoder_hidden_states=prompt_embeds, + text_embedding_mask=prompt_attention_mask, + encoder_hidden_states_t5=prompt_embeds_2, + text_embedding_mask_t5=prompt_attention_mask_2, + image_meta_size=add_time_ids, + style=style, + image_rotary_emb=image_rotary_emb, + return_dict=False, + )[0] + + noise_pred, _ = noise_pred.chunk(2, dim=1) + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred, noise_pred_text = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t, True + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + prompt_embeds_2 = callback_outputs.pop("prompt_embeds_2", prompt_embeds_2) + negative_prompt_embeds_2 = callback_outputs.pop( + "negative_prompt_embeds_2", negative_prompt_embeds_2 + ) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # 9. Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.transformer.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/pag/pipeline_pag_kolors.py b/diffusers/pipelines/pag/pipeline_pag_kolors.py new file mode 100644 index 0000000000000000000000000000000000000000..4f138d91d9c6d1529e9627925c9def05a3184f38 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_kolors.py @@ -0,0 +1,1129 @@ +# Copyright 2025 Stability AI, Kwai-Kolors Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import IPAdapterMixin, StableDiffusionXLLoraLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import deprecate, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..kolors.pipeline_output import KolorsPipelineOutput +from ..kolors.text_encoder import ChatGLMModel +from ..kolors.tokenizer import ChatGLMTokenizer +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForText2Image + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "Kwai-Kolors/Kolors-diffusers", + ... variant="fp16", + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... pag_applied_layers=["down.block_2.attentions_1", "up.block_0.attentions_1"], + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = ( + ... "A photo of a ladybug, macro, zoom, high quality, film, holding a wooden sign with the text 'KOLORS'" + ... ) + >>> image = pipe(prompt, guidance_scale=5.5, pag_scale=1.5).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class KolorsPAGPipeline( + DiffusionPipeline, StableDiffusionMixin, StableDiffusionXLLoraLoaderMixin, IPAdapterMixin, PAGMixin +): + r""" + Pipeline for text-to-image generation using Kolors. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`ChatGLMModel`]): + Frozen text-encoder. Kolors uses [ChatGLM3-6B](https://huggingface.co/THUDM/chatglm3-6b). + tokenizer (`ChatGLMTokenizer`): + Tokenizer of class + [ChatGLMTokenizer](https://huggingface.co/THUDM/chatglm3-6b/blob/main/tokenization_chatglm.py). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"False"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `Kwai-Kolors/Kolors-diffusers`. + pag_applied_layers (`str` or `list[str]``, *optional*, defaults to `"mid"`): + Set the transformer attention layers where to apply the perturbed attention guidance. Can be a string or a + list of strings with "down", "mid", "up", a whole transformer block or specific transformer block attention + layers, e.g.: + ["mid"] ["down", "mid"] ["down", "mid", "up.block_1"] ["down", "mid", "up.block_1.attentions_0", + "up.block_1.attentions_1"] + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = [ + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: ChatGLMModel, + tokenizer: ChatGLMTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = False, + pag_applied_layers: str | list[str] = "mid", + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.kolors.pipeline_kolors.KolorsPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt=None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 256, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + """ + # from IPython import embed; embed(); exit() + device = device or self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer] + text_encoders = [self.text_encoder] + + if prompt_embeds is None: + prompt_embeds_list = [] + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ).to(device) + output = text_encoder( + input_ids=text_inputs["input_ids"], + attention_mask=text_inputs["attention_mask"], + position_ids=text_inputs["position_ids"], + output_hidden_states=True, + ) + + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() + # [max_sequence_length, batch, hidden_size] -> [batch, hidden_size] + pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = prompt_embeds_list[0] + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + negative_prompt_embeds_list = [] + + for tokenizer, text_encoder in zip(tokenizers, text_encoders): + uncond_input = tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ).to(device) + output = text_encoder( + input_ids=uncond_input["input_ids"], + attention_mask=uncond_input["attention_mask"], + position_ids=uncond_input["position_ids"], + output_hidden_states=True, + ) + + # [max_sequence_length, batch, hidden_size] -> [batch, max_sequence_length, hidden_size] + # clone to have a contiguous tensor + negative_prompt_embeds = output.hidden_states[-2].permute(1, 0, 2).clone() + # [max_sequence_length, batch, hidden_size] -> [batch, hidden_size] + negative_pooled_prompt_embeds = output.hidden_states[-1][-1, :, :].clone() + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=text_encoder.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = negative_prompt_embeds_list[0] + + bs_embed = pooled_prompt_embeds.shape[0] + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.kolors.pipeline_kolors.KolorsPipeline.check_inputs + def check_inputs( + self, + prompt, + num_inference_steps, + height, + width, + negative_prompt=None, + prompt_embeds=None, + pooled_prompt_embeds=None, + negative_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + if max_sequence_length is not None and max_sequence_length > 256: + raise ValueError(f"`max_sequence_length` cannot be greater than 256 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + max_sequence_length: int = 256, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers) and checkpoints + that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [Kwai-Kolors/Kolors-diffusers](https://huggingface.co/Kwai-Kolors/Kolors-diffusers) and checkpoints + that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.kolors.KolorsPipelineOutput`] instead of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.kolors.KolorsPipelineOutput`] or `tuple`: [`~pipelines.kolors.KolorsPipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + num_inference_steps, + height, + width, + negative_prompt, + prompt_embeds, + pooled_prompt_embeds, + negative_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + add_text_embeds = self._prepare_perturbed_attention_guidance( + add_text_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + add_time_ids = self._prepare_perturbed_attention_guidance( + add_time_ids, negative_add_time_ids, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return KolorsPipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_pixart_sigma.py b/diffusers/pipelines/pag/pipeline_pag_pixart_sigma.py new file mode 100644 index 0000000000000000000000000000000000000000..2e7e8ab9814f890b6e3e2b8c0306da87c8a8d53a --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_pixart_sigma.py @@ -0,0 +1,882 @@ +# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +from typing import Callable + +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from ...image_processor import PixArtImageProcessor +from ...models import AutoencoderKL, PixArtTransformer2DModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + BACKENDS_MAPPING, + deprecate, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from ..pixart_alpha.pipeline_pixart_alpha import ( + ASPECT_RATIO_256_BIN, + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, +) +from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForText2Image + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS", + ... torch_dtype=torch.float16, + ... pag_applied_layers=["blocks.14"], + ... enable_pag=True, + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A small cactus with a happy face in the Sahara desert" + >>> image = pipe(prompt, pag_scale=4.0, guidance_scale=1.0).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class PixArtSigmaPAGPipeline(DiffusionPipeline, PAGMixin): + r""" + [PAG pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag) for text-to-image generation + using PixArt-Sigma. + """ + + bad_punct_regex = re.compile( + r"[" + + "#®•©™&@·º½¾¿¡§~" + + r"\)" + + r"\(" + + r"\]" + + r"\[" + + r"\}" + + r"\{" + + r"\|" + + "\\" + + r"\/" + + r"\*" + + r"]{1,}" + ) # noqa + + _optional_components = ["tokenizer", "text_encoder"] + model_cpu_offload_seq = "text_encoder->transformer->vae" + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + vae: AutoencoderKL, + transformer: PixArtTransformer2DModel, + scheduler: KarrasDiffusionSchedulers, + pag_applied_layers: str | list[str] = "blocks.1", # 1st transformer block + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.PixArtAlphaPipeline.encode_prompt with 120->300 + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + **kwargs, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Alpha, it's should be the embeddings of the "" + string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + """ + + if "mask_feature" in kwargs: + deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version." + deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False) + + if device is None: + device = self._execution_device + + # See Section 3.1. of the paper. + max_length = max_sequence_length + + if prompt_embeds is None: + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because T5 can only handle sequences up to" + f" {max_length} tokens: {removed_text}" + ) + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + elif self.transformer is not None: + dtype = self.transformer.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.repeat(1, num_images_per_prompt) + prompt_attention_mask = prompt_attention_mask.view(bs_embed * num_images_per_prompt, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens = [negative_prompt] * bs_embed if isinstance(negative_prompt, str) else negative_prompt + uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption) + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + negative_prompt_attention_mask = uncond_input.attention_mask + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(1, num_images_per_prompt) + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed * num_images_per_prompt, -1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.PixArtAlphaPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + negative_prompt, + callback_steps, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + height: int | None = None, + width: int | None = None, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + clean_caption: bool = True, + use_resolution_binning: bool = True, + max_sequence_length: int = 300, + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ) -> ImagePipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + max_sequence_length (`int` defaults to 300): Maximum sequence length to use with the `prompt`. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images + """ + # 1. Check inputs. Raise error if not correct + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + if use_resolution_binning: + if self.transformer.config.sample_size == 256: + aspect_ratio_bin = ASPECT_RATIO_2048_BIN + elif self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_256_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + negative_prompt, + callback_steps, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + ) + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, do_classifier_free_guidance + ) + prompt_attention_mask = self._prepare_perturbed_attention_guidance( + prompt_attention_mask, negative_prompt_attention_mask, do_classifier_free_guidance + ) + elif do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + if self.do_perturbed_attention_guidance: + original_attn_proc = self.transformer.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=do_classifier_free_guidance, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Prepare micro-conditions. + added_cond_kwargs = {"resolution": None, "aspect_ratio": None} + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + current_timestep = t + if not torch.is_tensor(current_timestep): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = latent_model_input.device.type == "mps" + is_npu = latent_model_input.device.type == "npu" + if isinstance(current_timestep, float): + dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + else: + dtype = torch.int32 if (is_mps or is_npu) else torch.int64 + current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device) + elif len(current_timestep.shape) == 0: + current_timestep = current_timestep[None].to(latent_model_input.device) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + current_timestep = current_timestep.expand(latent_model_input.shape[0]) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + timestep=current_timestep, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, do_classifier_free_guidance, guidance_scale, current_timestep + ) + elif do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + else: + noise_pred = noise_pred + + # compute previous image: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.transformer.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_sana.py b/diffusers/pipelines/pag/pipeline_pag_sana.py new file mode 100644 index 0000000000000000000000000000000000000000..f621255325a08d155875ff9f76e88d361694f277 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sana.py @@ -0,0 +1,924 @@ +# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Callable + +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PixArtImageProcessor +from ...models import AutoencoderDC, SanaTransformer2DModel +from ...models.attention_processor import PAGCFGSanaLinearAttnProcessor2_0, PAGIdentitySanaLinearAttnProcessor2_0 +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + BACKENDS_MAPPING, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from ..pixart_alpha.pipeline_pixart_alpha import ( + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, +) +from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN +from ..sana.pipeline_sana import ASPECT_RATIO_4096_BIN +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaPAGPipeline + + >>> pipe = SanaPAGPipeline.from_pretrained( + ... "Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers", + ... pag_applied_layers=["transformer_blocks.8"], + ... torch_dtype=torch.float32, + ... ) + >>> pipe.to("cuda") + >>> pipe.text_encoder.to(torch.bfloat16) + >>> pipe.transformer = pipe.transformer.to(torch.bfloat16) + + >>> image = pipe(prompt='a cyberpunk cat with a neon sign that says "Sana"')[0] + >>> image[0].save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaPAGPipeline(DiffusionPipeline, PAGMixin): + r""" + Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629). This pipeline + supports the use of [Perturbed Attention Guidance + (PAG)](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag). + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC, + transformer: SanaTransformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + pag_applied_layers: str | list[str] = "transformer_blocks.0", + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + if hasattr(self, "vae") and self.vae is not None + else 8 + ) + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.set_pag_applied_layers( + pag_applied_layers, + pag_attn_processors=(PAGCFGSanaLinearAttnProcessor2_0(), PAGIdentitySanaLinearAttnProcessor2_0()), + ) + + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0][:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + if self.transformer is not None: + dtype = self.transformer.dtype + elif self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption) + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + negative_prompt_attention_mask = uncond_input.attention_mask + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + height: int = 1024, + width: int = 1024, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ) -> ImagePipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 300): Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + if use_resolution_binning: + if self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_4096_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_2048_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 16: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + self._guidance_scale = guidance_scale + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + prompt_attention_mask = self._prepare_perturbed_attention_guidance( + prompt_attention_mask, negative_prompt_attention_mask, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + torch.float32, + device, + generator, + latents, + ) + if self.do_perturbed_attention_guidance: + original_attn_proc = self.transformer.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = latent_model_input.to(prompt_embeds.dtype) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]).to(latents.dtype) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute previous image: x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + try: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.transformer.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd.py b/diffusers/pipelines/pag/pipeline_pag_sd.py new file mode 100644 index 0000000000000000000000000000000000000000..b12597460f653b9fa675b5f64a321adbdc83e6f4 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd.py @@ -0,0 +1,1082 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +from typing import Any, Callable + +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForText2Image + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, enable_pag=True + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, pag_scale=0.3).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPAGPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + pag_applied_layers: str | list[str] = "mid", + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + None, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": ip_adapter_image_embeds} + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) + else None + ) + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred, noise_pred_text = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t, True + ) + + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_3.py b/diffusers/pipelines/pag/pipeline_pag_sd_3.py new file mode 100644 index 0000000000000000000000000000000000000000..d86adccc2ccfc520ca08b7b4db38966793cfbf0a --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_3.py @@ -0,0 +1,1001 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3LoraLoaderMixin +from ...models.attention_processor import PAGCFGJointAttnProcessor2_0, PAGJointAttnProcessor2_0 +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForText2Image + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... pag_applied_layers=["blocks.13"], + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> image = pipe(prompt, guidance_scale=5.0, pag_scale=0.7).images[0] + >>> image.save("sd3_pag.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3PAGPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, PAGMixin): + r""" + [PAG pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag) for text-to-image generation + using Stable Diffusion 3. + + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + pag_applied_layers: str | list[str] = "blocks.1", # 1st transformer block + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + self.set_pag_applied_layers( + pag_applied_layers, pag_attn_processors=(PAGCFGJointAttnProcessor2_0(), PAGJointAttnProcessor2_0()) + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + max_sequence_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + clip_skip: int | None = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str], + prompt_3: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + clip_skip: int | None = None, + max_sequence_length: int = 256, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale # + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + pooled_prompt_embeds = self._prepare_perturbed_attention_guidance( + pooled_prompt_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, sigmas=sigmas + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.transformer.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.transformer.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py b/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..24f3d828bd815f7e18f2f0f9926ed652be5f0345 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_3_img2img.py @@ -0,0 +1,1069 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3LoraLoaderMixin +from ...models.attention_processor import PAGCFGJointAttnProcessor2_0, PAGJointAttnProcessor2_0 +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..stable_diffusion_3.pipeline_output import StableDiffusion3PipelineOutput +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3PAGImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusion3PAGImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", + ... torch_dtype=torch.float16, + ... pag_applied_layers=["blocks.13"], + ... ) + >>> pipe.to("cuda") + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png" + >>> init_image = load_image(url).convert("RGB") + >>> image = pipe(prompt, image=init_image, guidance_scale=5.0, pag_scale=0.7).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3PAGImg2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, PAGMixin): + r""" + [PAG pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/pag) for image-to-image generation + using Stable Diffusion 3. + + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + pag_applied_layers: str | list[str] = "blocks.1", # 1st transformer block + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + self.set_pag_applied_layers( + pag_applied_layers, pag_attn_processors=(PAGCFGJointAttnProcessor2_0(), PAGJointAttnProcessor2_0()) + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + max_sequence_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + clip_skip: int | None = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str], + prompt_3: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + clip_skip: int | None = None, + max_sequence_length: int = 256, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + if image.shape[1] == self.vae.config.latent_channels: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = (init_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.scale_noise(init_latents, timestep, noise) + latents = init_latents.to(device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + image: PipelineImageInput = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to `self.default_sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to `self.default_sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + pooled_prompt_embeds = self._prepare_perturbed_attention_guidance( + pooled_prompt_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Preprocess image + image = self.image_processor.preprocess(image, height=height, width=width) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, sigmas=sigmas + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.transformer.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + # 6. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.transformer.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_animatediff.py b/diffusers/pipelines/pag/pipeline_pag_sd_animatediff.py new file mode 100644 index 0000000000000000000000000000000000000000..2baeda5649ade3bf56bf011a497c0b37ebc87405 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_animatediff.py @@ -0,0 +1,882 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput +from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...models.unets.unet_motion_model import MotionAdapter +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..animatediff.pipeline_output import AnimateDiffPipelineOutput +from ..free_init_utils import FreeInitMixin +from ..free_noise_utils import AnimateDiffFreeNoiseMixin +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AnimateDiffPAGPipeline, MotionAdapter, DDIMScheduler + >>> from diffusers.utils import export_to_gif + + >>> model_id = "SG161222/Realistic_Vision_V5.1_noVAE" + >>> motion_adapter_id = "guoyww/animatediff-motion-adapter-v1-5-2" + >>> motion_adapter = MotionAdapter.from_pretrained(motion_adapter_id) + >>> scheduler = DDIMScheduler.from_pretrained( + ... model_id, subfolder="scheduler", beta_schedule="linear", steps_offset=1, clip_sample=False + ... ) + >>> pipe = AnimateDiffPAGPipeline.from_pretrained( + ... model_id, + ... motion_adapter=motion_adapter, + ... scheduler=scheduler, + ... pag_applied_layers=["mid"], + ... torch_dtype=torch.float16, + ... ).to("cuda") + + >>> video = pipe( + ... prompt="car, futuristic cityscape with neon lights, street, no human", + ... negative_prompt="low quality, bad quality", + ... num_inference_steps=25, + ... guidance_scale=6.0, + ... pag_scale=3.0, + ... generator=torch.Generator().manual_seed(42), + ... ).frames[0] + + >>> export_to_gif(video, "animatediff_pag.gif") + ``` +""" + + +class AnimateDiffPAGPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FreeInitMixin, + AnimateDiffFreeNoiseMixin, + PAGMixin, +): + r""" + Pipeline for text-to-video generation using + [AnimateDiff](https://huggingface.co/docs/diffusers/en/api/pipelines/animatediff) and [Perturbed Attention + Guidance](https://huggingface.co/docs/diffusers/en/using-diffusers/pag). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`] to tokenize text. + unet ([`UNet2DConditionModel`]): + A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents. + motion_adapter ([`MotionAdapter`]): + A [`MotionAdapter`] to be used in combination with `unet` to denoise the encoded video latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel | UNetMotionModel, + motion_adapter: MotionAdapter, + scheduler: KarrasDiffusionSchedulers, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + pag_applied_layers: str | list[str] = "mid_block.*attn1", # ["mid"], ["down_blocks.1"] + ): + super().__init__() + if isinstance(unet, UNet2DConditionModel): + unet = UNetMotionModel.from_unet2d(unet, motion_adapter) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + motion_adapter=motion_adapter, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.decode_latents + def decode_latents(self, latents, decode_chunk_size: int = 16): + latents = 1 / self.vae.config.scaling_factor * latents + + batch_size, channels, num_frames, height, width = latents.shape + latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width) + + video = [] + for i in range(0, latents.shape[0], decode_chunk_size): + batch_latents = latents[i : i + decode_chunk_size] + batch_latents = self.vae.decode(batch_latents).sample + video.append(batch_latents) + + video = torch.cat(video) + video = video[None, :].reshape((batch_size, num_frames, -1) + video.shape[2:]).permute(0, 2, 1, 3, 4) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + video = video.float() + return video + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.deprecated.pia.pipeline_pia.PIAPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.animatediff.pipeline_animatediff.AnimateDiffPipeline.prepare_latents + def prepare_latents( + self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None + ): + # If FreeNoise is enabled, generate latents as described in Equation (7) of [FreeNoise](https://huggingface.co/papers/2310.15169) + if self.free_noise_enabled: + latents = self._prepare_latents_free_noise( + batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + shape = ( + batch_size, + num_channels_latents, + num_frames, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + num_frames: int | None = 16, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_videos_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + decode_chunk_size: int = 16, + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated video. + num_frames (`int`, *optional*, defaults to 16): + The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds + amounts to 2 seconds of video. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality videos at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. Latents should be of shape + `(batch_size, num_channel, num_frames, height, width)`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated video. Choose between `torch.Tensor`, `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead + of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + decode_chunk_size (`int`, *optional*, defaults to 16): + The number of frames to decode at a time when calling `decode_latents` method. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated frames. + """ + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + num_videos_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_videos_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + prompt_embeds = prompt_embeds.repeat_interleave(repeats=num_frames, dim=0) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_videos_per_prompt, + self.do_classifier_free_guidance, + ) + + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + num_frames, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": ip_adapter_image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + num_free_init_iters = self._free_init_num_iters if self.free_init_enabled else 1 + for free_init_iter in range(num_free_init_iters): + if self.free_init_enabled: + latents, timesteps = self._apply_free_init( + latents, free_init_iter, num_inference_steps, device, latents.dtype, generator + ) + + self._num_timesteps = len(timesteps) + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + # 8. Denoising loop + with self.progress_bar(total=self._num_timesteps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat( + [latents] * (prompt_embeds.shape[0] // num_frames // latents.shape[0]) + ) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + ).sample + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post processing + if output_type == "latent": + video = latents + else: + video_tensor = self.decode_latents(latents, decode_chunk_size) + video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type) + + # 10. Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (video,) + + return AnimateDiffPipelineOutput(frames=video) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_img2img.py b/diffusers/pipelines/pag/pipeline_pag_sd_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..de6dfbc585fa0f72ca1ef125ef8a848a6a482228 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_img2img.py @@ -0,0 +1,1112 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForImage2Image + >>> from diffusers.utils import load_image + + >>> pipe = AutoPipelineForImage2Image.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + >>> pipe = pipe.to("cuda") + >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png" + + >>> init_image = load_image(url).convert("RGB") + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, image=init_image, pag_scale=0.3).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPAGImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, + PAGMixin, +): + r""" + Pipeline for text-guided image-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + pag_applied_layers: str | list[str] = "mid", # ["mid"], ["down.block_1", "up.block_0.attentions_0"] + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. set timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + self._num_timesteps = len(timesteps) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + if ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = ip_adapter_image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_inpaint.py b/diffusers/pipelines/pag/pipeline_pag_sd_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..426419f12f730abfa11d44ebeb833c24162dd4ea --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_inpaint.py @@ -0,0 +1,1401 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker +from .pag_utils import PAGMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForInpainting + + >>> pipe = AutoPipelineForInpainting.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, enable_pag=True + ... ) + >>> pipe = pipe.to("cuda") + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> init_image = load_image(img_url).convert("RGB") + >>> mask_image = load_image(mask_url).convert("RGB") + >>> prompt = "A majestic tiger sitting on a bench" + >>> image = pipe( + ... prompt=prompt, + ... image=init_image, + ... mask_image=mask_image, + ... strength=0.8, + ... num_inference_steps=50, + ... guidance_scale=guidance_scale, + ... generator=generator, + ... pag_scale=pag_scale, + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPAGInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + pag_applied_layers: str | list[str] = "mid", + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.check_inputs + def check_inputs( + self, + prompt, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + else: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.Tensor = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.9999, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be inpainted (which parts of the image to + be masked out with `mask_image` and repainted according to `prompt`). + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, *optional*): + Pre-encoded latent of the masked image (for inpainting). + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 0.9999): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + mask_image, + height, + width, + strength, + None, + None, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. set timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps=num_inference_steps, strength=strength, device=device + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + if self.do_perturbed_attention_guidance: + if self.do_classifier_free_guidance: + mask, _ = mask.chunk(2) + masked_image_latents, _ = masked_image_latents.chunk(2) + mask = self._prepare_perturbed_attention_guidance(mask, mask, self.do_classifier_free_guidance) + masked_image_latents = self._prepare_perturbed_attention_guidance( + masked_image_latents, masked_image_latents, self.do_classifier_free_guidance + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + # 9 Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 9.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": ip_adapter_image_embeds} + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) + else None + ) + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 10. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if num_channels_unet == 9: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t + ) + + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_perturbed_attention_guidance: + init_mask, *_ = mask.chunk(3) if self.do_classifier_free_guidance else mask.chunk(2) + else: + init_mask, *_ = mask.chunk(2) if self.do_classifier_free_guidance else mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + condition_kwargs = {} + if isinstance(self.vae, AsymmetricAutoencoderKL): + init_image = init_image.to(device=device, dtype=masked_image_latents.dtype) + init_image_condition = init_image.clone() + init_image = self._encode_vae_image(init_image, generator=generator) + mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype) + condition_kwargs = {"image": init_image_condition, "mask": mask_condition} + image = self.vae.decode( + latents / self.vae.config.scaling_factor, return_dict=False, generator=generator, **condition_kwargs + )[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_xl.py b/diffusers/pipelines/pag/pipeline_pag_sd_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..1de57a92261694abedb4dcbe78c96c2d3e9b9156 --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_xl.py @@ -0,0 +1,1347 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput +from .pag_utils import PAGMixin + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForText2Image + + >>> pipe = AutoPipelineForText2Image.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, pag_scale=0.3).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLPAGPipeline( + DiffusionPipeline, + StableDiffusionMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "negative_add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + pag_applied_layers: str | list[str] = "mid", # ["mid"],["down.block_1"],["up.block_0.attentions_0"] + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + None, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + add_text_embeds = self._prepare_perturbed_attention_guidance( + add_text_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + add_time_ids = self._prepare_perturbed_attention_guidance( + add_time_ids, negative_add_time_ids, self.do_classifier_free_guidance + ) + + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance, perturbed-attention guidance, or both + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = ip_adapter_image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred, noise_pred_text = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t, True + ) + + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py b/diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..201d16a86f8a2da38d2847982192a3c73aacb7cd --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_xl_img2img.py @@ -0,0 +1,1539 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput +from .pag_utils import PAGMixin + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForImage2Image + >>> from diffusers.utils import load_image + + >>> pipe = AutoPipelineForImage2Image.from_pretrained( + ... "stabilityai/stable-diffusion-xl-refiner-1.0", + ... torch_dtype=torch.float16, + ... enable_pag=True, + ... ) + >>> pipe = pipe.to("cuda") + >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png" + + >>> init_image = load_image(url).convert("RGB") + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, image=init_image, pag_scale=0.3).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLPAGImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the + config of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "add_neg_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + pag_applied_layers: str | list[str] = "mid", # ["mid"], ["down.block_1", "up.block_0.attentions_0"] + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + strength, + num_inference_steps, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + if num_inference_steps is None: + raise ValueError("`num_inference_steps` cannot be None.") + elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.prepare_latents + def prepare_latents( + self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + latents_mean = latents_std = None + if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None: + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1) + if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None: + latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + empty_device_cache() + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + if latents_mean is not None and latents_std is not None: + latents_mean = latents_mean.to(device=device, dtype=dtype) + latents_std = latents_std.to(device=device, dtype=dtype) + init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std + else: + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + strength: float = 0.3, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_start: float | None = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `list[torch.Tensor]` or `list[PIL.Image.Image]` or `list[np.ndarray]`): + The image(s) to modify with the pipeline. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of + `denoising_start` being declared as an integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + num_inference_steps, + None, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. Prepare timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + add_noise = True if self.denoising_start is None else False + + # 6. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + add_noise, + ) + # 7. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 8. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + add_text_embeds = self._prepare_perturbed_attention_guidance( + add_text_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + add_time_ids = self._prepare_perturbed_attention_guidance( + add_time_ids, add_neg_time_ids, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 9. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 9.1 Apply denoising_end + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = ip_adapter_image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred, noise_pred_text = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t, True + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/pag/pipeline_pag_sd_xl_inpaint.py b/diffusers/pipelines/pag/pipeline_pag_sd_xl_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..9e70a7779f1e73467c2784c5893a08da63a637ca --- /dev/null +++ b/diffusers/pipelines/pag/pipeline_pag_sd_xl_inpaint.py @@ -0,0 +1,1775 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput +from .pag_utils import PAGMixin + + +if is_invisible_watermark_available(): + from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import AutoPipelineForInpainting + >>> from diffusers.utils import load_image + + >>> pipe = AutoPipelineForInpainting.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... torch_dtype=torch.float16, + ... variant="fp16", + ... enable_pag=True, + ... ) + >>> pipe.to("cuda") + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = load_image(img_url).convert("RGB") + >>> mask_image = load_image(mask_url).convert("RGB") + + >>> prompt = "A majestic tiger sitting on a bench" + >>> image = pipe( + ... prompt=prompt, + ... image=init_image, + ... mask_image=mask_image, + ... num_inference_steps=50, + ... strength=0.80, + ... pag_scale=0.3, + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLPAGInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, + PAGMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config + of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "negative_prompt_embeds", + "add_text_embeds", + "add_time_ids", + "negative_pooled_prompt_embeds", + "add_neg_time_ids", + "mask", + "masked_image_latents", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + pag_applied_layers: str | list[str] = "mid", # ["mid"], ["down.block_1", "up.block_0.attentions_0"] + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + self.set_pag_applied_layers(pag_applied_layers) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + add_noise=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if image.shape[1] == 4: + image_latents = image.to(device=device, dtype=dtype) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + elif return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None and add_noise: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + elif add_noise: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + else: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = image_latents.to(device) + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + dtype = image.dtype + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + image_latents = image_latents.to(dtype) + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_inpaint.StableDiffusionXLInpaintPipeline.prepare_mask_latents + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + + if masked_image is not None and masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = None + + if masked_image is not None: + if masked_image_latents is None: + masked_image = masked_image.to(device=device, dtype=dtype) + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) + + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.Tensor = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.9999, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_start: float | None = None, + denoising_end: float | None = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + pag_scale: float = 3.0, + pag_adaptive_scale: float = 0.0, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + masked_image_latents (`torch.Tensor`, *optional*): + Pre-encoded latent of the masked image (for inpainting). + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 0.9999): + Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be + between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the + `strength`. The number of denoising steps depends on the amount of noise initially added. When + `strength` is 1, added noise will be maximum and the denoising process will run for the full number of + iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked + portion of the reference `image`. Note that in the case of `denoising_start` being declared as an + integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + pag_scale (`float`, *optional*, defaults to 3.0): + The scale factor for the perturbed attention guidance. If it is set to 0.0, the perturbed attention + guidance will not be used. + pag_adaptive_scale (`float`, *optional*, defaults to 0.0): + The adaptive scale factor for the perturbed attention guidance. If it is set to 0.0, `pag_scale` is + used. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + None, + output_type, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + self._pag_scale = pag_scale + self._pag_adaptive_scale = pag_adaptive_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. set timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + mask = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is not None: + masked_image = masked_image_latents + elif init_image.shape[1] == 4: + # if images are in latent space, we can't mask it + masked_image = None + else: + masked_image = init_image * (mask < 0.5) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + add_noise = True if self.denoising_start is None else False + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + add_noise=add_noise, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask, masked_image_latents = self.prepare_mask_latents( + mask, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + if self.do_perturbed_attention_guidance: + if self.do_classifier_free_guidance: + mask, _ = mask.chunk(2) + masked_image_latents, _ = masked_image_latents.chunk(2) + mask = self._prepare_perturbed_attention_guidance(mask, mask, self.do_classifier_free_guidance) + masked_image_latents = self._prepare_perturbed_attention_guidance( + masked_image_latents, masked_image_latents, self.do_classifier_free_guidance + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + # 8.1 Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 10. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_perturbed_attention_guidance: + prompt_embeds = self._prepare_perturbed_attention_guidance( + prompt_embeds, negative_prompt_embeds, self.do_classifier_free_guidance + ) + add_text_embeds = self._prepare_perturbed_attention_guidance( + add_text_embeds, negative_pooled_prompt_embeds, self.do_classifier_free_guidance + ) + add_time_ids = self._prepare_perturbed_attention_guidance( + add_time_ids, add_neg_time_ids, self.do_classifier_free_guidance + ) + + elif self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + for i, image_embeds in enumerate(ip_adapter_image_embeds): + negative_image_embeds = None + if self.do_classifier_free_guidance: + negative_image_embeds, image_embeds = image_embeds.chunk(2) + + if self.do_perturbed_attention_guidance: + image_embeds = self._prepare_perturbed_attention_guidance( + image_embeds, negative_image_embeds, self.do_classifier_free_guidance + ) + elif self.do_classifier_free_guidance: + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + image_embeds = image_embeds.to(device) + ip_adapter_image_embeds[i] = image_embeds + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 11.1 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + if self.do_perturbed_attention_guidance: + original_attn_proc = self.unet.attn_processors + self._set_pag_attn_processor( + pag_applied_layers=self.pag_applied_layers, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * (prompt_embeds.shape[0] // latents.shape[0])) + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if num_channels_unet == 9: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = ip_adapter_image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_perturbed_attention_guidance: + noise_pred, noise_pred_text = self._apply_perturbed_attention_guidance( + noise_pred, self.do_classifier_free_guidance, self.guidance_scale, t, True + ) + elif self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_perturbed_attention_guidance: + init_mask, *_ = mask.chunk(3) if self.do_classifier_free_guidance else mask.chunk(2) + else: + init_mask, *_ = mask.chunk(2) if self.do_classifier_free_guidance else mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + return StableDiffusionXLPipelineOutput(images=latents) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if self.do_perturbed_attention_guidance: + self.unet.set_attn_processor(original_attn_proc) + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/pipeline_loading_utils.py b/diffusers/pipelines/pipeline_loading_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..67f055fb62e61b378f661b761b4f933d1ed0b70d --- /dev/null +++ b/diffusers/pipelines/pipeline_loading_utils.py @@ -0,0 +1,1237 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import importlib +import os +import re +import warnings +from pathlib import Path +from typing import Any, Callable + +import httpx +import requests +import torch +from huggingface_hub import DDUFEntry, ModelCard, model_info, snapshot_download +from huggingface_hub.utils import HfHubHTTPError, OfflineModeIsEnabled, validate_hf_hub_args +from packaging import version + +from .. import __version__ +from ..utils import ( + FLASHPACK_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + WEIGHTS_NAME, + _maybe_remap_transformers_class, + deprecate, + get_class_from_dynamic_module, + is_accelerate_available, + is_peft_available, + is_sdnq_available, + is_transformers_available, + is_transformers_version, + logging, +) +from ..utils.constants import DIFFUSERS_SDNQ_TRANSFORMERS +from ..utils.torch_utils import is_compiled_module +from .transformers_loading_utils import _load_tokenizer_from_dduf, _load_transformers_model_from_dduf + + +if is_transformers_available(): + import transformers + from transformers import PreTrainedModel, PreTrainedTokenizerBase + from transformers.utils import SAFE_WEIGHTS_NAME as TRANSFORMERS_SAFE_WEIGHTS_NAME + from transformers.utils import WEIGHTS_NAME as TRANSFORMERS_WEIGHTS_NAME + + +if is_accelerate_available(): + import accelerate + from accelerate import dispatch_model + from accelerate.hooks import remove_hook_from_module + from accelerate.utils import compute_module_sizes, get_max_memory + + +INDEX_FILE = "diffusion_pytorch_model.bin" +CUSTOM_PIPELINE_FILE_NAME = "pipeline.py" +DUMMY_MODULES_FOLDER = "diffusers.utils" +TRANSFORMERS_DUMMY_MODULES_FOLDER = "transformers.utils" +CONNECTED_PIPES_KEYS = ["prior"] + +# Auxiliary (non-weight) files a transformers component saves next to its weights. Repos with a flat, +# transformers-style layout host a component's files at the repo root instead of in a subfolder, where the +# folder-based allow patterns of `DiffusionPipeline.download` would miss them. Root-hosted weights and +# `config.json` are matched by their own patterns, so only these auxiliary filenames need listing. +# Currently the set needed by DiffusionGemma — extend as new flat-layout pipelines require it. +TRANSFORMERS_COMPONENT_AUX_FILES = [ + "chat_template.jinja", + "generation_config.json", + "processor_config.json", + "tokenizer.json", + "tokenizer_config.json", +] + +logger = logging.get_logger(__name__) + +LOADABLE_CLASSES = { + "diffusers": { + "ModelMixin": ["save_pretrained", "from_pretrained"], + "SchedulerMixin": ["save_pretrained", "from_pretrained"], + "DiffusionPipeline": ["save_pretrained", "from_pretrained"], + "OnnxRuntimeModel": ["save_pretrained", "from_pretrained"], + "BaseGuidance": ["save_pretrained", "from_pretrained"], + }, + "transformers": { + "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"], + "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"], + "PreTrainedModel": ["save_pretrained", "from_pretrained"], + "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"], + "ProcessorMixin": ["save_pretrained", "from_pretrained"], + "ImageProcessingMixin": ["save_pretrained", "from_pretrained"], + }, + "onnxruntime.training": { + "ORTModule": ["save_pretrained", "from_pretrained"], + }, +} + +ALL_IMPORTABLE_CLASSES = {} +for library in LOADABLE_CLASSES: + ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library]) + + +def is_safetensors_compatible(filenames, passed_components=None, folder_names=None, variant=None) -> bool: + """ + Checking for safetensors compatibility: + - The model is safetensors compatible only if there is a safetensors file for each model component present in + filenames. + + Converting default pytorch serialized filenames to safetensors serialized filenames: + - For models from the diffusers library, just replace the ".bin" extension with ".safetensors" + - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin" + extension is replaced with ".safetensors" + """ + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME] + + # model_pytorch, diffusion_model_pytorch, ... + weight_prefixes = [w.split(".")[0] for w in weight_names] + # .bin, .safetensors, ... + weight_suffixs = [w.split(".")[-1] for w in weight_names] + # -00001-of-00002 + transformers_index_format = r"\d{5}-of-\d{5}" + # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors` + variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$" + ) + non_variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$" + ) + + passed_components = passed_components or [] + # only weight files matter for safetensors compatibility + filenames = filter_model_files(filenames) + if folder_names: + filenames = {f for f in filenames if os.path.split(f)[0] in folder_names} + + # extract all components of the pipeline and their associated files + components = {} + for filename in filenames: + if not len(filename.split("/")) == 2: + continue + + component, component_filename = filename.split("/") + if component in passed_components: + continue + + components.setdefault(component, []) + components[component].append(component_filename) + + # If there are no component folders check the main directory for safetensors files + filtered_filenames = set() + if not components: + if variant is not None: + filtered_filenames = filter_with_regex(filenames, variant_file_re) + + # If no variant filenames exist check if non-variant files are available + if not filtered_filenames: + filtered_filenames = filter_with_regex(filenames, non_variant_file_re) + return any(".safetensors" in filename for filename in filtered_filenames) + + # iterate over all files of a component + # check if safetensor files exist for that component + for component, component_filenames in components.items(): + matches = [] + filtered_component_filenames = set() + # if variant is provided check if the variant of the safetensors exists + if variant is not None: + filtered_component_filenames = filter_with_regex(component_filenames, variant_file_re) + + # if variant safetensor files do not exist check for non-variants + if not filtered_component_filenames: + filtered_component_filenames = filter_with_regex(component_filenames, non_variant_file_re) + for component_filename in filtered_component_filenames: + filename, extension = os.path.splitext(component_filename) + + match_exists = extension == ".safetensors" + matches.append(match_exists) + + if not any(matches): + return False + + return True + + +def filter_model_files(filenames): + """Filter model repo files for just files/folders that contain model weights""" + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + FLASHPACK_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME] + + allowed_extensions = [wn.split(".")[-1] for wn in weight_names] + + return [f for f in filenames if any(f.endswith(extension) for extension in allowed_extensions)] + + +def filter_with_regex(filenames, pattern_re): + return {f for f in filenames if pattern_re.match(f.split("/")[-1]) is not None} + + +def variant_compatible_siblings(filenames, variant=None, ignore_patterns=None) -> list[os.PathLike] | str: + weight_names = [ + WEIGHTS_NAME, + SAFETENSORS_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + ONNX_EXTERNAL_WEIGHTS_NAME, + ] + + if is_transformers_available(): + weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME] + + # model_pytorch, diffusion_model_pytorch, ... + weight_prefixes = [w.split(".")[0] for w in weight_names] + # .bin, .safetensors, ... + weight_suffixs = [w.split(".")[-1] for w in weight_names] + # -00001-of-00002 + transformers_index_format = r"\d{5}-of-\d{5}" + + if variant is not None: + # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors` + variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$" + ) + # `text_encoder/pytorch_model.bin.index.fp16.json` + variant_index_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$" + ) + legacy_variant_file_re = re.compile(rf".*-{transformers_index_format}\.{variant}\.[a-z]+$") + legacy_variant_index_re = re.compile( + rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.{variant}\.index\.json$" + ) + + # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors` + non_variant_file_re = re.compile( + rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$" + ) + # `text_encoder/pytorch_model.bin.index.json` + non_variant_index_re = re.compile(rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.json") + + def filter_for_compatible_extensions(filenames, ignore_patterns=None): + if not ignore_patterns: + return filenames + + # ignore patterns uses glob style patterns e.g *.safetensors but we're only + # interested in the extension name + return {f for f in filenames if not any(f.endswith(pat.lstrip("*.")) for pat in ignore_patterns)} + + # Group files by component + components = {} + for filename in filenames: + if not len(filename.split("/")) == 2: + components.setdefault("", []).append(filename) + continue + + component, _ = filename.split("/") + components.setdefault(component, []).append(filename) + + usable_filenames = set() + variant_filenames = set() + for component, component_filenames in components.items(): + component_filenames = filter_for_compatible_extensions(component_filenames, ignore_patterns=ignore_patterns) + + component_variants = set() + component_legacy_variants = set() + component_non_variants = set() + if variant is not None: + component_variants = filter_with_regex(component_filenames, variant_file_re) + component_variant_index_files = filter_with_regex(component_filenames, variant_index_re) + + component_legacy_variants = filter_with_regex(component_filenames, legacy_variant_file_re) + component_legacy_variant_index_files = filter_with_regex(component_filenames, legacy_variant_index_re) + + if component_variants or component_legacy_variants: + variant_filenames.update( + component_variants | component_variant_index_files + if component_variants + else component_legacy_variants | component_legacy_variant_index_files + ) + + else: + component_non_variants = filter_with_regex(component_filenames, non_variant_file_re) + component_variant_index_files = filter_with_regex(component_filenames, non_variant_index_re) + + usable_filenames.update(component_non_variants | component_variant_index_files) + + usable_filenames.update(variant_filenames) + + if len(variant_filenames) == 0 and variant is not None: + error_message = f"You are trying to load model files of the `variant={variant}`, but no such modeling files are available. " + raise ValueError(error_message) + + if len(variant_filenames) > 0 and usable_filenames != variant_filenames: + logger.warning( + f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n" + f"[{', '.join(variant_filenames)}]\nLoaded non-{variant} filenames:\n" + f"[{', '.join(usable_filenames - variant_filenames)}\nIf this behavior is not " + f"expected, please check your folder structure." + ) + + return usable_filenames, variant_filenames + + +@validate_hf_hub_args +def warn_deprecated_model_variant(pretrained_model_name_or_path, token, variant, revision, model_filenames): + info = model_info( + pretrained_model_name_or_path, + token=token, + revision=None, + ) + filenames = {sibling.rfilename for sibling in info.siblings} + comp_model_filenames, _ = variant_compatible_siblings(filenames, variant=revision) + comp_model_filenames = [".".join(f.split(".")[:1] + f.split(".")[2:]) for f in comp_model_filenames] + + if set(model_filenames).issubset(set(comp_model_filenames)): + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` even though you can load it via `variant=`{revision}`. Loading model variants via `revision='{revision}'` is deprecated and will be removed in diffusers v1. Please use `variant='{revision}'` instead.", + FutureWarning, + ) + else: + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have the required variant filenames in the 'main' branch. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {revision} files' so that the correct variant file can be added.", + FutureWarning, + ) + + +def _unwrap_model(model): + """Unwraps a model.""" + if is_compiled_module(model): + model = model._orig_mod + + if is_peft_available(): + from peft import PeftModel + + if isinstance(model, PeftModel): + model = model.base_model.model + + return model + + +def maybe_raise_or_warn( + library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module +): + """Simple helper method to raise or warn in case incorrect module has been passed""" + if not is_pipeline_module: + library = importlib.import_module(library_name) + + # Handle deprecated Transformers classes + if library_name == "transformers": + class_name = _maybe_remap_transformers_class(class_name) or class_name + + class_obj = getattr(library, class_name) + class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()} + + expected_class_obj = None + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + expected_class_obj = class_candidate + + # Dynamo wraps the original model in a private class. + # I didn't find a public API to get the original class. + sub_model = passed_class_obj[name] + unwrapped_sub_model = _unwrap_model(sub_model) + model_cls = unwrapped_sub_model.__class__ + + if not issubclass(model_cls, expected_class_obj): + raise ValueError(f"{passed_class_obj[name]} is of type: {model_cls}, but should be {expected_class_obj}") + else: + logger.warning( + f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it" + " has the correct type" + ) + + +# a simpler version of get_class_obj_and_candidates, it won't work with custom code +def simple_get_class_obj(library_name, class_name): + from diffusers import pipelines + + is_pipeline_module = hasattr(pipelines, library_name) + + if is_pipeline_module: + pipeline_module = getattr(pipelines, library_name) + class_obj = getattr(pipeline_module, class_name) + else: + library = importlib.import_module(library_name) + + # Handle deprecated Transformers classes + if library_name == "transformers": + class_name = _maybe_remap_transformers_class(class_name) or class_name + + class_obj = getattr(library, class_name) + + return class_obj + + +def get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=None, + cache_dir=None, + trust_remote_code: bool = False, +): + """Simple helper method to retrieve class object of module as well as potential parent class objects""" + component_folder = os.path.join(cache_dir, component_name) if component_name and cache_dir else None + + if class_name.startswith("FlashPack"): + class_name = class_name.removeprefix("FlashPack") + + if is_pipeline_module: + pipeline_module = getattr(pipelines, library_name) + + class_obj = getattr(pipeline_module, class_name) + class_candidates = dict.fromkeys(importable_classes.keys(), class_obj) + elif component_folder and os.path.isfile(os.path.join(component_folder, library_name + ".py")): + # load custom component + class_obj = get_class_from_dynamic_module( + component_folder, + module_file=library_name + ".py", + class_name=class_name, + trust_remote_code=trust_remote_code, + ) + class_candidates = dict.fromkeys(importable_classes.keys(), class_obj) + else: + # else we just import it from the library. + library = importlib.import_module(library_name) + + # Handle deprecated Transformers classes + if library_name == "transformers": + class_name = _maybe_remap_transformers_class(class_name) or class_name + + class_obj = getattr(library, class_name) + class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()} + + return class_obj, class_candidates + + +def _get_custom_pipeline_class( + custom_pipeline, + repo_id=None, + hub_revision=None, + class_name=None, + cache_dir=None, + revision=None, + trust_remote_code: bool = False, +): + if custom_pipeline.endswith(".py"): + path = Path(custom_pipeline) + # decompose into folder & file + file_name = path.name + custom_pipeline = path.parent.absolute() + elif repo_id is not None: + file_name = f"{custom_pipeline}.py" + custom_pipeline = repo_id + else: + file_name = CUSTOM_PIPELINE_FILE_NAME + + if repo_id is not None and hub_revision is not None: + # if we load the pipeline code from the Hub + # make sure to overwrite the `revision` + revision = hub_revision + + return get_class_from_dynamic_module( + custom_pipeline, + module_file=file_name, + class_name=class_name, + cache_dir=cache_dir, + revision=revision, + trust_remote_code=trust_remote_code, + ) + + +def _get_pipeline_class( + class_obj, + config=None, + load_connected_pipeline=False, + custom_pipeline=None, + repo_id=None, + hub_revision=None, + class_name=None, + cache_dir=None, + revision=None, + trust_remote_code: bool = False, +): + if custom_pipeline is not None: + return _get_custom_pipeline_class( + custom_pipeline, + repo_id=repo_id, + hub_revision=hub_revision, + class_name=class_name, + cache_dir=cache_dir, + revision=revision, + trust_remote_code=trust_remote_code, + ) + + if class_obj.__name__ != "DiffusionPipeline" and class_obj.__name__ != "ModularPipeline": + return class_obj + + diffusers_module = importlib.import_module(class_obj.__module__.split(".")[0]) + class_name = class_name or config["_class_name"] + if not class_name: + raise ValueError( + "The class name could not be found in the configuration file. Please make sure to pass the correct `class_name`." + ) + + pipeline_cls = getattr(diffusers_module, class_name) + + if load_connected_pipeline: + from .auto_pipeline import _get_connected_pipeline + + connected_pipeline_cls = _get_connected_pipeline(pipeline_cls) + if connected_pipeline_cls is not None: + logger.info( + f"Loading connected pipeline {connected_pipeline_cls.__name__} instead of {pipeline_cls.__name__} as specified via `load_connected_pipeline=True`" + ) + else: + logger.info(f"{pipeline_cls.__name__} has no connected pipeline class. Loading {pipeline_cls.__name__}.") + + pipeline_cls = connected_pipeline_cls or pipeline_cls + + return pipeline_cls + + +def _load_empty_model( + library_name: str, + class_name: str, + importable_classes: list[Any], + pipelines: Any, + is_pipeline_module: bool, + name: str, + dtype: str | torch.dtype, + cached_folder: str | os.PathLike, + **kwargs, +): + # retrieve class objects. + class_obj, _ = get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=name, + cache_dir=cached_folder, + ) + + if is_transformers_available(): + transformers_version = version.parse(version.parse(transformers.__version__).base_version) + else: + transformers_version = "N/A" + + # Determine library. + is_transformers_model = ( + is_transformers_available() + and issubclass(class_obj, PreTrainedModel) + and transformers_version >= version.parse("4.20.0") + ) + diffusers_module = importlib.import_module(__name__.split(".")[0]) + is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin) + + model = None + config_path = cached_folder + user_agent = { + "diffusers": __version__, + "file_type": "model", + "framework": "pytorch", + } + + if is_diffusers_model: + # Load config and then the model on meta. + config, unused_kwargs, commit_hash = class_obj.load_config( + os.path.join(config_path, name), + cache_dir=cached_folder, + return_unused_kwargs=True, + return_commit_hash=True, + force_download=kwargs.pop("force_download", False), + proxies=kwargs.pop("proxies", None), + local_files_only=kwargs.pop("local_files_only", False), + token=kwargs.pop("token", None), + revision=kwargs.pop("revision", None), + subfolder=kwargs.pop("subfolder", None), + user_agent=user_agent, + ) + with accelerate.init_empty_weights(): + model = class_obj.from_config(config, **unused_kwargs) + elif is_transformers_model: + config_class = getattr(class_obj, "config_class", None) + if config_class is None: + raise ValueError("`config_class` cannot be None. Please double-check the model.") + + config = config_class.from_pretrained( + cached_folder, + subfolder=name, + force_download=kwargs.pop("force_download", False), + proxies=kwargs.pop("proxies", None), + local_files_only=kwargs.pop("local_files_only", False), + token=kwargs.pop("token", None), + revision=kwargs.pop("revision", None), + user_agent=user_agent, + ) + with accelerate.init_empty_weights(): + model = class_obj(config) + + if model is not None: + model = model.to(dtype=dtype) + return model + + +def _assign_components_to_devices( + module_sizes: dict[str, float], device_memory: dict[str, float], device_mapping_strategy: str = "balanced" +): + device_ids = list(device_memory.keys()) + device_cycle = device_ids + device_ids[::-1] + device_memory = device_memory.copy() + + device_id_component_mapping = {} + current_device_index = 0 + for component in module_sizes: + device_id = device_cycle[current_device_index % len(device_cycle)] + component_memory = module_sizes[component] + curr_device_memory = device_memory[device_id] + + # If the GPU doesn't fit the current component offload to the CPU. + if component_memory > curr_device_memory: + device_id_component_mapping["cpu"] = [component] + else: + if device_id not in device_id_component_mapping: + device_id_component_mapping[device_id] = [component] + else: + device_id_component_mapping[device_id].append(component) + + # Update the device memory. + device_memory[device_id] -= component_memory + current_device_index += 1 + + return device_id_component_mapping + + +def _get_final_device_map(device_map, pipeline_class, passed_class_obj, init_dict, library, max_memory, **kwargs): + # TODO: separate out different device_map methods when it gets to it. + if device_map != "balanced": + return device_map + # To avoid circular import problem. + from diffusers import pipelines + + dtype = kwargs.get("dtype", torch.float32) + + # Load each module in the pipeline on a meta device so that we can derive the device map. + init_empty_modules = {} + for name, (library_name, class_name) in init_dict.items(): + # Define all importable classes + is_pipeline_module = hasattr(pipelines, library_name) + importable_classes = ALL_IMPORTABLE_CLASSES + loaded_sub_model = None + + # Use passed sub model or load class_name from library_name + if name in passed_class_obj: + # if the model is in a pipeline module, then we load it from the pipeline + # check that passed_class_obj has correct parent class + maybe_raise_or_warn( + library_name, + library, + class_name, + importable_classes, + passed_class_obj, + name, + is_pipeline_module, + ) + with accelerate.init_empty_weights(): + loaded_sub_model = passed_class_obj[name] + + else: + sub_model_dtype = ( + dtype.get(name, dtype.get("default", torch.float32)) if isinstance(dtype, dict) else dtype + ) + loaded_sub_model = _load_empty_model( + library_name=library_name, + class_name=class_name, + importable_classes=importable_classes, + pipelines=pipelines, + is_pipeline_module=is_pipeline_module, + pipeline_class=pipeline_class, + name=name, + dtype=sub_model_dtype, + cached_folder=kwargs.get("cached_folder", None), + force_download=kwargs.get("force_download", None), + proxies=kwargs.get("proxies", None), + local_files_only=kwargs.get("local_files_only", None), + token=kwargs.get("token", None), + revision=kwargs.get("revision", None), + ) + + if loaded_sub_model is not None: + init_empty_modules[name] = loaded_sub_model + + # determine device map + # Obtain a sorted dictionary for mapping the model-level components + # to their sizes. + module_sizes = { + module_name: compute_module_sizes( + module, + dtype=dtype.get(module_name, dtype.get("default", torch.float32)) if isinstance(dtype, dict) else dtype, + )[""] + for module_name, module in init_empty_modules.items() + if isinstance(module, torch.nn.Module) + } + module_sizes = dict(sorted(module_sizes.items(), key=lambda item: item[1], reverse=True)) + + # Obtain maximum memory available per device (GPUs only). + max_memory = get_max_memory(max_memory) + max_memory = dict(sorted(max_memory.items(), key=lambda item: item[1], reverse=True)) + max_memory = {k: v for k, v in max_memory.items() if k != "cpu"} + + # Obtain a dictionary mapping the model-level components to the available + # devices based on the maximum memory and the model sizes. + final_device_map = None + if len(max_memory) > 0: + device_id_component_mapping = _assign_components_to_devices( + module_sizes, max_memory, device_mapping_strategy=device_map + ) + + # Obtain the final device map, e.g., `{"unet": 0, "text_encoder": 1, "vae": 1, ...}` + final_device_map = {} + for device_id, components in device_id_component_mapping.items(): + for component in components: + final_device_map[component] = device_id + + return final_device_map + + +def load_sub_model( + library_name: str, + class_name: str, + importable_classes: list[Any], + pipelines: Any, + is_pipeline_module: bool, + pipeline_class: Any, + dtype: torch.dtype, + provider: Any, + sess_options: Any, + device_map: dict[str, torch.device] | str | None, + max_memory: dict[int | str, int | str] | None, + offload_folder: str | os.PathLike | None, + offload_state_dict: bool, + model_variants: dict[str, str], + name: str, + variant: str, + low_cpu_mem_usage: bool, + cached_folder: str | os.PathLike, + use_safetensors: bool, + dduf_entries: dict[str, DDUFEntry] | None, + provider_options: Any, + disable_mmap: bool, + quantization_config: Any | None = None, + use_flashpack: bool = False, + trust_remote_code: bool = False, +): + """Helper method to load the module `name` from `library_name` and `class_name`""" + from ..quantizers import PipelineQuantizationConfig + + # retrieve class candidates + + class_obj, class_candidates = get_class_obj_and_candidates( + library_name, + class_name, + importable_classes, + pipelines, + is_pipeline_module, + component_name=name, + cache_dir=cached_folder, + trust_remote_code=trust_remote_code, + ) + + load_method_name = None + # retrieve load method name + for class_name, class_candidate in class_candidates.items(): + if class_candidate is not None and issubclass(class_obj, class_candidate): + load_method_name = importable_classes[class_name][1] + + # if load method name is None, then we have a dummy module -> raise Error + if load_method_name is None: + none_module = class_obj.__module__ + is_dummy_path = none_module.startswith(DUMMY_MODULES_FOLDER) or none_module.startswith( + TRANSFORMERS_DUMMY_MODULES_FOLDER + ) + if is_dummy_path and "dummy" in none_module: + # call class_obj for nice error message of missing requirements + class_obj() + + raise ValueError( + f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have" + f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}." + ) + + load_method = _get_load_method(class_obj, load_method_name, is_dduf=dduf_entries is not None) + + # add kwargs to loading method + diffusers_module = importlib.import_module(__name__.split(".")[0]) + loading_kwargs = {} + + is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin) + + if is_transformers_available(): + transformers_version = version.parse(version.parse(transformers.__version__).base_version) + else: + transformers_version = "N/A" + + is_transformers_model = ( + is_transformers_available() + and issubclass(class_obj, PreTrainedModel) + and transformers_version >= version.parse("4.20.0") + ) + + # For transformers models >= 4.56.0, use 'dtype' instead of 'torch_dtype' to avoid deprecation warnings + if issubclass(class_obj, torch.nn.Module): + if is_transformers_model and transformers_version >= version.parse("4.56.0"): + loading_kwargs["dtype"] = dtype + else: + loading_kwargs["torch_dtype"] = dtype + if issubclass(class_obj, diffusers_module.OnnxRuntimeModel): + loading_kwargs["provider"] = provider + loading_kwargs["sess_options"] = sess_options + loading_kwargs["provider_options"] = provider_options + + # When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers. + # To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default. + # This makes sure that the weights won't be initialized which significantly speeds up loading. + if is_diffusers_model or is_transformers_model: + loading_kwargs["device_map"] = device_map + loading_kwargs["max_memory"] = max_memory + loading_kwargs["offload_folder"] = offload_folder + loading_kwargs["offload_state_dict"] = offload_state_dict + loading_kwargs["variant"] = model_variants.pop(name, None) + loading_kwargs["use_safetensors"] = use_safetensors + + if is_diffusers_model: + loading_kwargs["use_flashpack"] = use_flashpack + + # the following can be deleted once the minimum required `transformers` version + # is higher than 4.27 + if ( + is_transformers_model + and loading_kwargs["variant"] is not None + and transformers_version < version.parse("4.27.0") + ): + raise ImportError( + f"When passing `variant='{variant}'`, please make sure to upgrade your `transformers` version to at least 4.27.0.dev0" + ) + elif is_transformers_model and loading_kwargs["variant"] is None: + loading_kwargs.pop("variant") + + loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage + + if is_diffusers_model: + loading_kwargs["disable_mmap"] = disable_mmap + + if is_transformers_model and is_transformers_version(">=", "4.57.0"): + loading_kwargs.pop("offload_state_dict") + + if ( + quantization_config is not None + and isinstance(quantization_config, PipelineQuantizationConfig) + and issubclass(class_obj, torch.nn.Module) + ): + model_quant_config = quantization_config._resolve_quant_config( + is_diffusers=is_diffusers_model, module_name=name + ) + if model_quant_config is not None: + loading_kwargs["quantization_config"] = model_quant_config + + if is_transformers_model and DIFFUSERS_SDNQ_TRANSFORMERS and is_sdnq_available(): + # Opt-in via DIFFUSERS_SDNQ_TRANSFORMERS: import sdnq so it registers with transformers. + from ..quantizers.sdnq.sdnq_quantizer import _ensure_sdnq_registered + + _ensure_sdnq_registered() + + # check if the module is in a subdirectory + if dduf_entries: + loading_kwargs["dduf_entries"] = dduf_entries + loaded_sub_model = load_method(name, **loading_kwargs) + elif os.path.isdir(os.path.join(cached_folder, name)): + loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs) + else: + # else load from the root directory + loaded_sub_model = load_method(cached_folder, **loading_kwargs) + + if isinstance(loaded_sub_model, torch.nn.Module) and isinstance(device_map, dict) and not use_flashpack: + # remove hooks + remove_hook_from_module(loaded_sub_model, recurse=True) + needs_offloading_to_cpu = device_map[""] == "cpu" + skip_keys = None + if hasattr(loaded_sub_model, "_skip_keys") and loaded_sub_model._skip_keys is not None: + skip_keys = loaded_sub_model._skip_keys + + if needs_offloading_to_cpu: + dispatch_model( + loaded_sub_model, + state_dict=loaded_sub_model.state_dict(), + device_map=device_map, + force_hooks=True, + main_device=0, + skip_keys=skip_keys, + ) + else: + dispatch_model(loaded_sub_model, device_map=device_map, force_hooks=True, skip_keys=skip_keys) + + return loaded_sub_model + + +def _get_load_method(class_obj: object, load_method_name: str, is_dduf: bool) -> Callable: + """ + Return the method to load the sub model. + + In practice, this method will return the `"from_pretrained"` (or `load_method_name`) method of the class object + except if loading from a DDUF checkpoint. In that case, transformers models and tokenizers have a specific loading + method that we need to use. + """ + if is_dduf: + if issubclass(class_obj, PreTrainedTokenizerBase): + return lambda *args, **kwargs: _load_tokenizer_from_dduf(class_obj, *args, **kwargs) + if issubclass(class_obj, PreTrainedModel): + return lambda *args, **kwargs: _load_transformers_model_from_dduf(class_obj, *args, **kwargs) + return getattr(class_obj, load_method_name) + + +def _fetch_class_library_tuple(module): + # import it here to avoid circular import + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipelines = getattr(diffusers_module, "pipelines") + + # register the config from the original module, not the dynamo compiled one + not_compiled_module = _unwrap_model(module) + library = not_compiled_module.__module__.split(".")[0] + + # check if the module is a pipeline module + module_path_items = not_compiled_module.__module__.split(".") + pipeline_dir = module_path_items[-2] if len(module_path_items) > 2 else None + + path = not_compiled_module.__module__.split(".") + is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir) + + # if library is not in LOADABLE_CLASSES, then it is a custom module. + # Or if it's a pipeline module, then the module is inside the pipeline + # folder so we set the library to module name. + if is_pipeline_module: + library = pipeline_dir + elif library not in LOADABLE_CLASSES: + library = not_compiled_module.__module__ + + # retrieve class_name + if isinstance(not_compiled_module, type): + class_name = not_compiled_module.__name__ + else: + class_name = not_compiled_module.__class__.__name__ + + return (library, class_name) + + +def _identify_model_variants(folder: str, variant: str, config: dict) -> dict: + model_variants = {} + if variant is not None: + for sub_folder in os.listdir(folder): + folder_path = os.path.join(folder, sub_folder) + is_folder = os.path.isdir(folder_path) and sub_folder in config + variant_exists = is_folder and any(p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)) + if variant_exists: + model_variants[sub_folder] = variant + return model_variants + + +def _resolve_custom_pipeline_and_cls(folder, config, custom_pipeline): + custom_class_name = None + if os.path.isfile(os.path.join(folder, f"{custom_pipeline}.py")): + custom_pipeline = os.path.join(folder, f"{custom_pipeline}.py") + elif isinstance(config["_class_name"], (list, tuple)) and os.path.isfile( + os.path.join(folder, f"{config['_class_name'][0]}.py") + ): + custom_pipeline = os.path.join(folder, f"{config['_class_name'][0]}.py") + custom_class_name = config["_class_name"][1] + + return custom_pipeline, custom_class_name + + +def _maybe_raise_warning_for_inpainting(pipeline_class, pretrained_model_name_or_path: str, config: dict): + if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse( + version.parse(config["_diffusers_version"]).base_version + ) <= version.parse("0.5.1"): + from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy + + pipeline_class = StableDiffusionInpaintPipelineLegacy + + deprecation_message = ( + "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the" + f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For" + " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting" + " checkpoint: https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-inpainting instead or adapting your" + f" checkpoint {pretrained_model_name_or_path} to the format of" + " https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-inpainting. Note that we do not actively maintain" + " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0." + ) + deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False) + + +def _update_init_kwargs_with_connected_pipeline( + init_kwargs: dict, passed_pipe_kwargs: dict, passed_class_objs: dict, folder: str, **pipeline_loading_kwargs +) -> dict: + from .pipeline_utils import DiffusionPipeline + + modelcard = ModelCard.load(os.path.join(folder, "README.md")) + connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS} + + # We don't scheduler argument to match the existing logic: + # https://github.com/huggingface/diffusers/blob/867e0c919e1aa7ef8b03c8eb1460f4f875a683ae/src/diffusers/pipelines/pipeline_utils.py#L906C13-L925C14 + pipeline_loading_kwargs_cp = pipeline_loading_kwargs.copy() + if pipeline_loading_kwargs_cp is not None and len(pipeline_loading_kwargs_cp) >= 1: + for k in pipeline_loading_kwargs: + if "scheduler" in k: + _ = pipeline_loading_kwargs_cp.pop(k) + + def get_connected_passed_kwargs(prefix): + connected_passed_class_obj = { + k.replace(f"{prefix}_", ""): w for k, w in passed_class_objs.items() if k.split("_")[0] == prefix + } + connected_passed_pipe_kwargs = { + k.replace(f"{prefix}_", ""): w for k, w in passed_pipe_kwargs.items() if k.split("_")[0] == prefix + } + + connected_passed_kwargs = {**connected_passed_class_obj, **connected_passed_pipe_kwargs} + return connected_passed_kwargs + + connected_pipes = { + prefix: DiffusionPipeline.from_pretrained( + repo_id, **pipeline_loading_kwargs_cp, **get_connected_passed_kwargs(prefix) + ) + for prefix, repo_id in connected_pipes.items() + if repo_id is not None + } + + for prefix, connected_pipe in connected_pipes.items(): + # add connected pipes to `init_kwargs` with _, e.g. "prior_text_encoder" + init_kwargs.update( + {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()} + ) + + return init_kwargs + + +def _get_custom_components_and_folders( + pretrained_model_name: str, + config_dict: dict[str, Any], + filenames: list[str] | None = None, + variant_filenames: list[str] | None = None, + variant: str | None = None, +): + config_dict = config_dict.copy() + + # retrieve all folder_names that contain relevant files + folder_names = [k for k, v in config_dict.items() if isinstance(v, list) and k != "_class_name"] + + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipelines = getattr(diffusers_module, "pipelines") + + # optionally create a custom component <> custom file mapping + custom_components = {} + for component in folder_names: + module_candidate = config_dict[component][0] + + if module_candidate is None or not isinstance(module_candidate, str): + continue + + # We compute candidate file path on the Hub. Do not use `os.path.join`. + candidate_file = f"{component}/{module_candidate}.py" + + if candidate_file in filenames: + custom_components[component] = module_candidate + elif module_candidate not in LOADABLE_CLASSES and not hasattr(pipelines, module_candidate): + raise ValueError( + f"{candidate_file} as defined in `model_index.json` does not exist in {pretrained_model_name} and is not a module in 'diffusers/pipelines'." + ) + + return custom_components, folder_names + + +def _get_ignore_patterns( + passed_components, + model_folder_names: list[str], + model_filenames: list[str], + use_safetensors: bool, + allow_pickle: bool, + use_onnx: bool, + is_onnx: bool, + use_flashpack: bool, + variant: str | None = None, +) -> list[str]: + if ( + use_safetensors + and not allow_pickle + and not is_safetensors_compatible( + model_filenames, passed_components=passed_components, folder_names=model_folder_names, variant=variant + ) + ): + raise EnvironmentError( + f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})" + ) + + if use_safetensors and is_safetensors_compatible( + model_filenames, passed_components=passed_components, folder_names=model_folder_names, variant=variant + ): + ignore_patterns = ["*.bin", "*.msgpack"] + + use_onnx = use_onnx if use_onnx is not None else is_onnx + if not use_onnx: + ignore_patterns += ["*.onnx", "*.pb"] + + elif use_flashpack: + ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb", "*.msgpack"] + + else: + ignore_patterns = ["*.safetensors", "*.msgpack"] + + use_onnx = use_onnx if use_onnx is not None else is_onnx + if not use_onnx: + ignore_patterns += ["*.onnx", "*.pb"] + + return ignore_patterns + + +def _download_dduf_file( + pretrained_model_name: str, + dduf_file: str, + pipeline_class_name: str, + cache_dir: str, + proxies: str, + local_files_only: bool, + token: str, + revision: str, +): + model_info_call_error = None + if not local_files_only: + try: + info = model_info(pretrained_model_name, token=token, revision=revision) + except (HfHubHTTPError, OfflineModeIsEnabled, requests.ConnectionError, httpx.NetworkError) as e: + logger.warning(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.") + local_files_only = True + model_info_call_error = e # save error to reraise it if model is not cached locally + + if ( + not local_files_only + and dduf_file is not None + and dduf_file not in (sibling.rfilename for sibling in info.siblings) + ): + raise ValueError(f"Requested {dduf_file} file is not available in {pretrained_model_name}.") + + try: + user_agent = {"pipeline_class": pipeline_class_name, "dduf": True} + cached_folder = snapshot_download( + pretrained_model_name, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=[dduf_file], + user_agent=user_agent, + ) + return cached_folder + except FileNotFoundError: + # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache. + # This can happen in two cases: + # 1. If the user passed `local_files_only=True` => we raise the error directly + # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error + if model_info_call_error is None: + # 1. user passed `local_files_only=True` + raise + else: + # 2. we forced `local_files_only=True` when `model_info` failed + raise EnvironmentError( + f"Cannot load model {pretrained_model_name}: model is not cached locally and an error occurred" + " while trying to fetch metadata from the Hub. Please check out the root cause in the stacktrace" + " above." + ) from model_info_call_error + + +def _maybe_raise_error_for_incorrect_transformers(config_dict): + has_transformers_component = False + for k in config_dict: + if isinstance(config_dict[k], list): + has_transformers_component = config_dict[k][0] == "transformers" + if has_transformers_component: + break + if has_transformers_component and not is_transformers_version(">", "4.47.1"): + raise ValueError("Please upgrade your `transformers` installation to the latest version to use DDUF.") + + +def _maybe_warn_for_wrong_component_in_quant_config(pipe_init_dict, quant_config): + if quant_config is None: + return + + actual_pipe_components = set(pipe_init_dict.keys()) + missing = "" + quant_components = None + if getattr(quant_config, "components_to_quantize", None) is not None: + quant_components = set(quant_config.components_to_quantize) + elif getattr(quant_config, "quant_mapping", None) is not None and isinstance(quant_config.quant_mapping, dict): + quant_components = set(quant_config.quant_mapping.keys()) + + if quant_components and not quant_components.issubset(actual_pipe_components): + missing = quant_components - actual_pipe_components + + if missing: + logger.warning( + f"The following components in the quantization config {missing} will be ignored " + "as they do not belong to the underlying pipeline. Acceptable values for the pipeline " + f"components are: {', '.join(actual_pipe_components)}." + ) diff --git a/diffusers/pipelines/pipeline_utils.py b/diffusers/pipelines/pipeline_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..a683973df5d76c800c83f5d3f207cf61ec5057aa --- /dev/null +++ b/diffusers/pipelines/pipeline_utils.py @@ -0,0 +1,2363 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import fnmatch +import importlib +import inspect +import os +import re +import sys +import types +from dataclasses import dataclass +from pathlib import Path +from typing import Any, Callable, Dict, List, Union, get_args, get_origin, get_type_hints + +import httpx +import numpy as np +import PIL.Image +import requests +import torch +from huggingface_hub import ( + DDUFEntry, + ModelCard, + create_repo, + get_cached_repo_tree, + hf_hub_download, + model_info, + read_dduf_file, + snapshot_download, +) +from huggingface_hub.errors import CachedRepoTreeNotFoundError +from huggingface_hub.utils import ( + HfHubHTTPError, + LocalEntryNotFoundError, + OfflineModeIsEnabled, + validate_hf_hub_args, +) +from packaging import version +from tqdm.auto import tqdm +from typing_extensions import Self + +from .. import __version__ +from ..configuration_utils import ConfigMixin +from ..models import AutoencoderKL +from ..models.attention_processor import FusedAttnProcessor2_0 +from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, ModelMixin +from ..quantizers import PipelineQuantizationConfig +from ..quantizers.bitsandbytes.utils import _check_bnb_status +from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME +from ..utils import ( + CONFIG_NAME, + DEPRECATED_REVISION_ARGS, + BaseOutput, + PushToHubMixin, + _get_detailed_type, + _is_valid_type, + _resolve_dtype, + deprecate, + is_accelerate_available, + is_accelerate_version, + is_bitsandbytes_version, + is_hpu_available, + is_torch_npu_available, + is_torch_version, + is_transformers_version, + logging, + numpy_to_pil, +) +from ..utils.distributed_utils import is_torch_dist_rank_zero +from ..utils.hub_utils import _check_legacy_sharding_variant_format, load_or_create_model_card, populate_model_card +from ..utils.torch_utils import empty_device_cache, get_device, is_compiled_module + + +if is_torch_npu_available(): + import torch_npu # noqa: F401 + +from .pipeline_loading_utils import ( + ALL_IMPORTABLE_CLASSES, + CONNECTED_PIPES_KEYS, + CUSTOM_PIPELINE_FILE_NAME, + LOADABLE_CLASSES, + TRANSFORMERS_COMPONENT_AUX_FILES, + _download_dduf_file, + _fetch_class_library_tuple, + _get_custom_components_and_folders, + _get_custom_pipeline_class, + _get_final_device_map, + _get_ignore_patterns, + _get_pipeline_class, + _identify_model_variants, + _maybe_raise_error_for_incorrect_transformers, + _maybe_raise_warning_for_inpainting, + _maybe_warn_for_wrong_component_in_quant_config, + _resolve_custom_pipeline_and_cls, + _unwrap_model, + _update_init_kwargs_with_connected_pipeline, + filter_model_files, + load_sub_model, + maybe_raise_or_warn, + variant_compatible_siblings, + warn_deprecated_model_variant, +) + + +if is_accelerate_available(): + import accelerate + + +LIBRARIES = [] +for library in LOADABLE_CLASSES: + LIBRARIES.append(library) + +SUPPORTED_DEVICE_MAP = ["balanced"] + [get_device(), "cpu"] + +_DDUF_DEPRECATION_MESSAGE = ( + "Loading pipelines from DDUF files is deprecated and DDUF support will be removed entirely. " + "Please save and load your pipelines using the standard Diffusers directory format instead." +) + +logger = logging.get_logger(__name__) + + +@dataclass +class ImagePipelineOutput(BaseOutput): + """ + Output class for image pipelines. + + Args: + images (`List[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + """ + + images: list[PIL.Image.Image] | np.ndarray + + +@dataclass +class AudioPipelineOutput(BaseOutput): + """ + Output class for audio pipelines. + + Args: + audios (`np.ndarray`) + List of denoised audio samples of a NumPy array of shape `(batch_size, num_channels, sample_rate)`. + """ + + audios: np.ndarray + + +class DeprecatedPipelineMixin: + """ + A mixin that can be used to mark a pipeline as deprecated. + + Pipelines inheriting from this mixin will raise a warning when instantiated, indicating that they are deprecated + and won't receive updates past the specified version. Tests will be skipped for pipelines that inherit from this + mixin. + + Example usage: + ```python + class MyDeprecatedPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + _last_supported_version = "0.20.0" + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + ``` + """ + + # Override this in the inheriting class to specify the last version that will support this pipeline + _last_supported_version = None + + def __init__(self, *args, **kwargs): + # Get the class name for the warning message + class_name = self.__class__.__name__ + + # Get the last supported version or use the current version if not specified + version_info = getattr(self.__class__, "_last_supported_version", __version__) + + # Raise a warning that this pipeline is deprecated + logger.warning( + f"The {class_name} has been deprecated and will not receive bug fixes or feature updates after Diffusers version {version_info}. " + ) + + # Call the parent class's __init__ method + super().__init__(*args, **kwargs) + + +class DiffusionPipeline(ConfigMixin, PushToHubMixin): + r""" + Base class for all pipelines. + + [`DiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and + provides methods for loading, downloading and saving models. It also includes methods to: + + - move all PyTorch modules to the device of your choice + - enable/disable the progress bar for the denoising iteration + + Class attributes: + + - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the + diffusion pipeline's components. + - **_optional_components** (`List[str]`) -- List of all optional components that don't have to be passed to the + pipeline to function (should be overridden by subclasses). + """ + + config_name = "model_index.json" + model_cpu_offload_seq = None + hf_device_map = None + _optional_components = [] + _exclude_from_cpu_offload = [] + _load_connected_pipes = False + _is_onnx = False + + def register_modules(self, **kwargs): + for name, module in kwargs.items(): + # retrieve library + if module is None or isinstance(module, (tuple, list)) and module[0] is None: + register_dict = {name: (None, None)} + else: + library, class_name = _fetch_class_library_tuple(module) + register_dict = {name: (library, class_name)} + + # save model index config + self.register_to_config(**register_dict) + + # set models + setattr(self, name, module) + + def __setattr__(self, name: str, value: Any): + if name in self.__dict__ and hasattr(self.config, name): + # We need to overwrite the config if name exists in config + if isinstance(getattr(self.config, name), (tuple, list)): + if value is not None and self.config[name][0] is not None: + class_library_tuple = _fetch_class_library_tuple(value) + else: + class_library_tuple = (None, None) + + self.register_to_config(**{name: class_library_tuple}) + else: + self.register_to_config(**{name: value}) + + super().__setattr__(name, value) + + def save_pretrained( + self, + save_directory: str | os.PathLike, + safe_serialization: bool = True, + variant: str | None = None, + max_shard_size: int | str | None = None, + push_to_hub: bool = False, + use_flashpack: bool = False, + **kwargs, + ): + """ + Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its + class implements both a save and loading method. The pipeline is easily reloaded using the + [`~DiffusionPipeline.from_pretrained`] class method. + + Arguments: + save_directory (`str` or `os.PathLike`): + Directory to save a pipeline to. Will be created if it doesn't exist. + safe_serialization (`bool`, *optional*, defaults to `True`): + Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`. + variant (`str`, *optional*): + If specified, weights are saved in the format `pytorch_model..bin`. + max_shard_size (`int` or `str`, defaults to `None`): + The maximum size for a checkpoint before being sharded. Checkpoints shard will then be each of size + lower than this size. If expressed as a string, needs to be digits followed by a unit (like `"5GB"`). + If expressed as an integer, the unit is bytes. Note that this limit will be decreased after a certain + period of time (starting from Oct 2024) to allow users to upgrade to the latest version of `diffusers`. + This is to establish a common default size for this argument across different libraries in the Hugging + Face ecosystem (`transformers`, and `accelerate`, for example). + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + + kwargs (`Dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + model_index_dict = dict(self.config) + model_index_dict.pop("_class_name", None) + model_index_dict.pop("_diffusers_version", None) + model_index_dict.pop("_module", None) + model_index_dict.pop("_name_or_path", None) + + if push_to_hub: + commit_message = kwargs.pop("commit_message", None) + private = kwargs.pop("private", None) + create_pr = kwargs.pop("create_pr", False) + token = kwargs.pop("token", None) + repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1]) + repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id + + expected_modules, optional_kwargs = self._get_signature_keys(self) + + def is_saveable_module(name, value): + if name not in expected_modules: + return False + if name in self._optional_components and value[0] is None: + return False + return True + + model_index_dict = {k: v for k, v in model_index_dict.items() if is_saveable_module(k, v)} + for pipeline_component_name in model_index_dict.keys(): + sub_model = getattr(self, pipeline_component_name) + model_cls = sub_model.__class__ + + # Dynamo wraps the original model in a private class. + # I didn't find a public API to get the original class. + if is_compiled_module(sub_model): + sub_model = _unwrap_model(sub_model) + model_cls = sub_model.__class__ + + save_method_name = None + # search for the model's base class in LOADABLE_CLASSES + for library_name, library_classes in LOADABLE_CLASSES.items(): + if library_name in sys.modules: + library = importlib.import_module(library_name) + else: + logger.info( + f"{library_name} is not installed. Cannot save {pipeline_component_name} as {library_classes} from {library_name}" + ) + + for base_class, save_load_methods in library_classes.items(): + class_candidate = getattr(library, base_class, None) + if class_candidate is not None and issubclass(model_cls, class_candidate): + # if we found a suitable base class in LOADABLE_CLASSES then grab its save method + save_method_name = save_load_methods[0] + break + if save_method_name is not None: + break + + if save_method_name is None: + logger.warning( + f"self.{pipeline_component_name}={sub_model} of type {type(sub_model)} cannot be saved." + ) + # make sure that unsaveable components are not tried to be loaded afterward + self.register_to_config(**{pipeline_component_name: (None, None)}) + continue + + save_method = getattr(sub_model, save_method_name) + + # Call the save method with the argument safe_serialization only if it's supported + save_method_signature = inspect.signature(save_method) + save_method_accept_safe = "safe_serialization" in save_method_signature.parameters + save_method_accept_variant = "variant" in save_method_signature.parameters + save_method_accept_max_shard_size = "max_shard_size" in save_method_signature.parameters + save_method_accept_flashpack = "use_flashpack" in save_method_signature.parameters + save_method_accept_peft_format = "save_peft_format" in save_method_signature.parameters + + save_kwargs = {} + if save_method_accept_safe: + save_kwargs["safe_serialization"] = safe_serialization + if save_method_accept_variant: + save_kwargs["variant"] = variant + if save_method_accept_max_shard_size and max_shard_size is not None: + # max_shard_size is expected to not be None in ModelMixin + save_kwargs["max_shard_size"] = max_shard_size + if save_method_accept_flashpack: + save_kwargs["use_flashpack"] = use_flashpack + if save_method_accept_peft_format: + # Set save_peft_format=False for transformers>=5.0.0 compatibility + # In transformers 5.0.0+, the default save_peft_format=True adds "base_model.model" prefix + # to adapter keys, but from_pretrained expects keys without this prefix + save_kwargs["save_peft_format"] = False + + save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs) + + # finally save the config + self.save_config(save_directory) + + if push_to_hub: + # Create a new empty model card and eventually tag it + model_card = load_or_create_model_card(repo_id, token=token, is_pipeline=True) + model_card = populate_model_card(model_card) + model_card.save(os.path.join(save_directory, "README.md")) + + self._upload_folder( + save_directory, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + ) + + def to(self, *args, **kwargs) -> Self: + r""" + Performs Pipeline dtype and/or device conversion. A torch.dtype and torch.device are inferred from the + arguments of `self.to(*args, **kwargs).` + + > [!TIP] > If the pipeline already has the correct torch.dtype and torch.device, then it is returned as is. + Otherwise, > the returned pipeline is a copy of self with the desired torch.dtype and torch.device. + + + Here are the ways to call `to`: + + - `to(dtype, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + - `to(device, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified + [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) + - `to(device=None, dtype=None, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the + specified [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) and + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + + Arguments: + dtype (`torch.dtype`, *optional*): + Returns a pipeline with the specified + [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype) + device (`torch.Device`, *optional*): + Returns a pipeline with the specified + [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) + silence_dtype_warnings (`str`, *optional*, defaults to `False`): + Whether to omit warnings if the target `dtype` is not compatible with the target `device`. + + Returns: + [`DiffusionPipeline`]: The pipeline converted to specified `dtype` and/or `dtype`. + """ + dtype = kwargs.pop("dtype", None) + device = kwargs.pop("device", None) + silence_dtype_warnings = kwargs.pop("silence_dtype_warnings", False) + + dtype_arg = None + device_arg = None + if len(args) == 1: + if isinstance(args[0], torch.dtype): + dtype_arg = args[0] + else: + device_arg = torch.device(args[0]) if args[0] is not None else None + elif len(args) == 2: + if isinstance(args[0], torch.dtype): + raise ValueError( + "When passing two arguments, make sure the first corresponds to `device` and the second to `dtype`." + ) + device_arg = torch.device(args[0]) if args[0] is not None else None + dtype_arg = args[1] + elif len(args) > 2: + raise ValueError("Please make sure to pass at most two arguments (`device` and `dtype`) `.to(...)`") + + if dtype is not None and dtype_arg is not None: + raise ValueError( + "You have passed `dtype` both as an argument and as a keyword argument. Please only pass one of the two." + ) + + dtype = dtype or dtype_arg + + if device is not None and device_arg is not None: + raise ValueError( + "You have passed `device` both as an argument and as a keyword argument. Please only pass one of the two." + ) + + device = device or device_arg + device_type = torch.device(device).type if device is not None else None + pipeline_has_bnb = any(any((_check_bnb_status(module))) for _, module in self.components.items()) + + # throw warning if pipeline is in "offloaded"-mode but user tries to manually set to GPU. + def module_is_sequentially_offloaded(module): + if not is_accelerate_available() or is_accelerate_version("<", "0.14.0"): + return False + + _, _, is_loaded_in_8bit_bnb = _check_bnb_status(module) + + # https://github.com/huggingface/accelerate/pull/3907 + if is_loaded_in_8bit_bnb and ( + is_bitsandbytes_version("<", "0.48.0") or is_accelerate_version("<", "1.13.0.dev0") + ): + return False + + return hasattr(module, "_hf_hook") and ( + isinstance(module._hf_hook, accelerate.hooks.AlignDevicesHook) + or hasattr(module._hf_hook, "hooks") + and isinstance(module._hf_hook.hooks[0], accelerate.hooks.AlignDevicesHook) + ) + + def module_is_offloaded(module): + if not is_accelerate_available() or is_accelerate_version("<", "0.17.0.dev0"): + return False + + return hasattr(module, "_hf_hook") and isinstance(module._hf_hook, accelerate.hooks.CpuOffload) + + # .to("cuda") would raise an error if the pipeline is sequentially offloaded, so we raise our own to make it clearer + pipeline_is_sequentially_offloaded = any( + module_is_sequentially_offloaded(module) for _, module in self.components.items() + ) + is_pipeline_device_mapped = self._is_pipeline_device_mapped() + if is_pipeline_device_mapped: + raise ValueError( + "It seems like you have activated a device mapping strategy on the pipeline which doesn't allow explicit device placement using `to()`. You can call `reset_device_map()` to remove the existing device map from the pipeline." + ) + + if device_type in ["cuda", "xpu"]: + if pipeline_is_sequentially_offloaded and not pipeline_has_bnb: + raise ValueError( + "It seems like you have activated sequential model offloading by calling `enable_sequential_cpu_offload`, but are now attempting to move the pipeline to GPU. This is not compatible with offloading. Please, move your pipeline `.to('cpu')` or consider removing the move altogether if you use sequential offloading." + ) + # PR: https://github.com/huggingface/accelerate/pull/3223/ + elif pipeline_has_bnb and is_accelerate_version("<", "1.1.0.dev0"): + raise ValueError( + "You are trying to call `.to('cuda')` on a pipeline that has models quantized with `bitsandbytes`. Your current `accelerate` installation does not support it. Please upgrade the installation." + ) + + # Display a warning in this case (the operation succeeds but the benefits are lost) + pipeline_is_offloaded = any(module_is_offloaded(module) for _, module in self.components.items()) + if pipeline_is_offloaded and device_type in ["cuda", "xpu"]: + logger.warning( + f"It seems like you have activated model offloading by calling `enable_model_cpu_offload`, but are now manually moving the pipeline to GPU. It is strongly recommended against doing so as memory gains from offloading are likely to be lost. Offloading automatically takes care of moving the individual components {', '.join(self.components.keys())} to GPU when needed. To make sure offloading works as expected, you should consider moving the pipeline back to CPU: `pipeline.to('cpu')` or removing the move altogether if you use offloading." + ) + + # Enable generic support for Intel Gaudi accelerator using GPU/HPU migration + if device_type == "hpu" and kwargs.pop("hpu_migration", True) and is_hpu_available(): + os.environ["PT_HPU_GPU_MIGRATION"] = "1" + logger.debug("Environment variable set: PT_HPU_GPU_MIGRATION=1") + + import habana_frameworks.torch # noqa: F401 + + # HPU hardware check + if not (hasattr(torch, "hpu") and torch.hpu.is_available()): + raise ValueError("You are trying to call `.to('hpu')` but HPU device is unavailable.") + + os.environ["PT_HPU_MAX_COMPOUND_OP_SIZE"] = "1" + logger.debug("Environment variable set: PT_HPU_MAX_COMPOUND_OP_SIZE=1") + + if dtype in (torch.bfloat16, None) and kwargs.pop("sdp_on_bf16", True): + if hasattr(torch._C, "_set_math_sdp_allow_fp16_bf16_reduction"): + torch._C._set_math_sdp_allow_fp16_bf16_reduction(True) + logger.warning( + "Enabled SDP with BF16 precision on HPU. To disable, please use `.to('hpu', sdp_on_bf16=False)`" + ) + + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded + for module in modules: + _, is_loaded_in_4bit_bnb, is_loaded_in_8bit_bnb = _check_bnb_status(module) + is_group_offloaded = self._maybe_raise_error_if_group_offload_active(module=module) + + if (is_loaded_in_4bit_bnb or is_loaded_in_8bit_bnb) and dtype is not None: + logger.warning( + f"The module '{module.__class__.__name__}' has been loaded in `bitsandbytes` {'4bit' if is_loaded_in_4bit_bnb else '8bit'} and conversion to {dtype} is not supported. Module is still in {'4bit' if is_loaded_in_4bit_bnb else '8bit'} precision." + ) + + if is_loaded_in_8bit_bnb and device is not None and is_bitsandbytes_version("<", "0.48.0"): + logger.warning( + f"The module '{module.__class__.__name__}' has been loaded in `bitsandbytes` 8bit and moving it to {device} via `.to()` is not supported. Module is still on {module.device}." + "You need to upgrade bitsandbytes to at least 0.48.0" + ) + + # Note: we also handle this at the ModelMixin level. The reason for doing it here too is that modeling + # components can be from outside diffusers too, but still have group offloading enabled. + if ( + self._maybe_raise_error_if_group_offload_active(raise_error=False, module=module) + and device is not None + ): + logger.warning( + f"The module '{module.__class__.__name__}' is group offloaded and moving it to {device} via `.to()` is not supported." + ) + + # This can happen for `transformer` models. CPU placement was added in + # https://github.com/huggingface/transformers/pull/33122. So, we guard this accordingly. + if is_loaded_in_4bit_bnb and device is not None and is_transformers_version(">", "4.44.0"): + module.to(device=device) + # added here https://github.com/huggingface/transformers/pull/43258 + if ( + is_loaded_in_8bit_bnb + and device is not None + and is_transformers_version(">", "4.58.0") + and is_bitsandbytes_version(">=", "0.48.0") + ): + module.to(device=device) + elif not is_loaded_in_4bit_bnb and not is_loaded_in_8bit_bnb and not is_group_offloaded: + module.to(device, dtype) + + if ( + module.dtype == torch.float16 + and str(device) in ["cpu"] + and not silence_dtype_warnings + and not is_offloaded + ): + logger.warning( + "Pipelines loaded with `dtype=torch.float16` cannot run with `cpu` device. It" + " is not recommended to move them to `cpu` as running them will fail. Please make" + " sure to use an accelerator to run the pipeline in inference, due to the lack of" + " support for`float16` operations on this device in PyTorch. Please, remove the" + " `dtype=torch.float16` argument, or use another device for inference." + ) + return self + + @property + def device(self) -> torch.device: + r""" + Returns: + `torch.device`: The torch device on which the pipeline is located. + """ + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + return module.device + + return torch.device("cpu") + + @property + def dtype(self) -> torch.dtype: + r""" + Returns: + `torch.dtype`: The torch dtype on which the pipeline is located. + """ + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + return module.dtype + + return torch.float32 + + @classmethod + @validate_hf_hub_args + def from_pretrained(cls, pretrained_model_name_or_path: str | os.PathLike, **kwargs) -> Self: + r""" + Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights. + + The pipeline is set in evaluation mode (`model.eval()`) by default. + + If you get the error message below, you need to finetune the weights for your downstream task: + + ``` + Some weights of UNet2DConditionModel were not initialized from the model checkpoint at stable-diffusion-v1-5/stable-diffusion-v1-5 and are newly initialized because the shapes did not match: + - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated + You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. + ``` + + Parameters: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights + saved using + [`~DiffusionPipeline.save_pretrained`]. + - A path to a *directory* (for example `./my_pipeline_directory/`) containing a dduf file + dtype (`torch.dtype` or `dict[str, Union[str, torch.dtype]]`, *optional*): + Override the default `torch.dtype` and load the model with another dtype. To load submodels with + different dtype pass a `dict` (for example `{'transformer': torch.bfloat16, 'vae': torch.float16}`). + Set the default dtype for unspecified components with `default` (for example `{'transformer': + torch.bfloat16, 'default': torch.float16}`). If a component is not specified and no default is set, + `torch.float32` is used. + custom_pipeline (`str`, *optional*): + + > [!WARNING] > 🧪 This is an experimental feature and may change in the future. + + Can be either: + + - A string, the *repo id* (for example `hf-internal-testing/diffusers-dummy-pipeline`) of a custom + pipeline hosted on the Hub. The repository must contain a file called pipeline.py that defines + the custom pipeline. + - A string, the *file name* of a community pipeline hosted on GitHub under + [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file + names must match the file name and not the pipeline script (`clip_guided_stable_diffusion` + instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the + current main branch of GitHub. + - A path to a directory (`./my_pipeline_directory/`) containing a custom pipeline. The directory + must contain a file called `pipeline.py` that defines the custom pipeline. + + For more information on how to load and create custom pipelines, please have a look at [Loading and + Adding Custom + Pipelines](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview) + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + cache_dir (`Union[str, os.PathLike]`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. Defaults to the latest stable 🤗 Diffusers + version. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + device_map (`str`, *optional*): + Strategy that dictates how the different components of a pipeline should be placed on available + devices. Currently, only "balanced" `device_map` is supported. Check out + [this](https://huggingface.co/docs/diffusers/main/en/tutorials/inference_with_big_models#device-placement) + to know more. + max_memory (`Dict`, *optional*): + A dictionary device identifier for the maximum memory. Will default to the maximum memory available for + each GPU and the available CPU RAM if unset. + offload_folder (`str` or `os.PathLike`, *optional*): + The path to offload weights if device_map contains the value `"disk"`. + offload_state_dict (`bool`, *optional*): + If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if + the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True` + when there is some disk offload. + low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`): + Speed up model loading only loading the pretrained weights and not initializing the weights. This also + tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model. + Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this + argument to `True` will raise an error. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + use_onnx (`bool`, *optional*, defaults to `None`): + If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights + will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is + `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending + with `.onnx` and `.pb`. + kwargs (remaining dictionary of keyword arguments, *optional*): + Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline + class). The overwritten components are passed directly to the pipelines `__init__` method. See example + below for more information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. + dduf_file(`str`, *optional*): + Load weights from the specified dduf file. This argument is deprecated and will be removed + in version 0.41.0. + disable_mmap ('bool', *optional*, defaults to 'False'): + Whether to disable mmap when loading a Safetensors model. This option can perform better when the model + is on a network mount or hard drive, which may not handle the seeky-ness of mmap very well. + + > [!TIP] > To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in + with `hf > auth login`. + + Examples: + + ```py + >>> from diffusers import DiffusionPipeline + + >>> # Download pipeline from huggingface.co and cache. + >>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256") + + >>> # Download pipeline that requires an authorization token + >>> # For more information on access tokens, please refer to this section + >>> # of the documentation](https://huggingface.co/docs/hub/security-tokens) + >>> pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + + >>> # Use a different scheduler + >>> from diffusers import LMSDiscreteScheduler + + >>> scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config) + >>> pipeline.scheduler = scheduler + ``` + """ + # Copy the kwargs to re-use during loading connected pipeline. + kwargs_copied = kwargs.copy() + + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + torch_dtype = kwargs.pop("torch_dtype", None) + dtype = _resolve_dtype(kwargs.pop("dtype", None), torch_dtype) + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + provider = kwargs.pop("provider", None) + sess_options = kwargs.pop("sess_options", None) + provider_options = kwargs.pop("provider_options", None) + device_map = kwargs.pop("device_map", None) + max_memory = kwargs.pop("max_memory", None) + offload_folder = kwargs.pop("offload_folder", None) + offload_state_dict = kwargs.pop("offload_state_dict", None) + low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT) + variant = kwargs.pop("variant", None) + dduf_file = kwargs.pop("dduf_file", None) + use_safetensors = kwargs.pop("use_safetensors", None) + use_onnx = kwargs.pop("use_onnx", None) + load_connected_pipeline = kwargs.pop("load_connected_pipeline", False) + quantization_config = kwargs.pop("quantization_config", None) + use_flashpack = kwargs.pop("use_flashpack", False) + disable_mmap = kwargs.pop("disable_mmap", False) + trust_remote_code = kwargs.pop("trust_remote_code", False) + + if dtype is not None and not isinstance(dtype, dict) and not isinstance(dtype, torch.dtype): + logger.warning(f"Passed `dtype` {dtype} is not a `torch.dtype`. Defaulting to `torch.float32`.") + dtype = torch.float32 + + if low_cpu_mem_usage and not is_accelerate_available(): + low_cpu_mem_usage = False + logger.warning( + "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the" + " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install" + " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip" + " install accelerate\n```\n." + ) + + if quantization_config is not None and not isinstance(quantization_config, PipelineQuantizationConfig): + raise ValueError("`quantization_config` must be an instance of `PipelineQuantizationConfig`.") + + if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"): + raise NotImplementedError( + "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set" + " `low_cpu_mem_usage=False`." + ) + + if device_map is not None and not is_torch_version(">=", "1.9.0"): + raise NotImplementedError( + "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set" + " `device_map=None`." + ) + + if device_map is not None and not is_accelerate_available(): + raise NotImplementedError( + "Using `device_map` requires the `accelerate` library. Please install it using: `pip install accelerate`." + ) + + if device_map is not None and not isinstance(device_map, str): + raise ValueError("`device_map` must be a string.") + + if device_map is not None and device_map not in SUPPORTED_DEVICE_MAP: + raise NotImplementedError( + f"{device_map} not supported. Supported strategies are: {', '.join(SUPPORTED_DEVICE_MAP)}" + ) + + if device_map is not None and device_map in SUPPORTED_DEVICE_MAP: + if is_accelerate_version("<", "0.28.0"): + raise NotImplementedError("Device placement requires `accelerate` version `0.28.0` or later.") + + if low_cpu_mem_usage is False and device_map is not None: + raise ValueError( + f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and" + " dispatching. Please make sure to set `low_cpu_mem_usage=True`." + ) + + if dduf_file: + deprecate("dduf_file", "0.41.0", _DDUF_DEPRECATION_MESSAGE) + if custom_pipeline: + raise NotImplementedError("Custom pipelines are not supported with DDUF at the moment.") + if load_connected_pipeline: + raise NotImplementedError("Connected pipelines are not supported with DDUF at the moment.") + + # 1. Download the checkpoints and configs + # use snapshot download here to get it working from from_pretrained + if not os.path.isdir(pretrained_model_name_or_path): + if pretrained_model_name_or_path.count("/") > 1: + raise ValueError( + f'The provided pretrained_model_name_or_path "{pretrained_model_name_or_path}"' + " is neither a valid local path nor a valid repo id. Please check the parameter." + ) + cached_folder = cls.download( + pretrained_model_name_or_path, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + use_safetensors=use_safetensors, + use_onnx=use_onnx, + custom_pipeline=custom_pipeline, + custom_revision=custom_revision, + variant=variant, + dduf_file=dduf_file, + load_connected_pipeline=load_connected_pipeline, + trust_remote_code=trust_remote_code, + **kwargs, + ) + else: + cached_folder = pretrained_model_name_or_path + + # The variant filenames can have the legacy sharding checkpoint format that we check and throw + # a warning if detected. + if variant is not None and _check_legacy_sharding_variant_format(folder=cached_folder, variant=variant): + warn_msg = ( + f"Warning: The repository contains sharded checkpoints for variant '{variant}' maybe in a deprecated format. " + "Please check your files carefully:\n\n" + "- Correct format example: diffusion_pytorch_model.fp16-00003-of-00003.safetensors\n" + "- Deprecated format example: diffusion_pytorch_model-00001-of-00002.fp16.safetensors\n\n" + "If you find any files in the deprecated format:\n" + "1. Remove all existing checkpoint files for this variant.\n" + "2. Re-obtain the correct files by running `save_pretrained()`.\n\n" + "This will ensure you're using the most up-to-date and compatible checkpoint format." + ) + logger.warning(warn_msg) + + dduf_entries = None + if dduf_file: + dduf_file_path = os.path.join(cached_folder, dduf_file) + dduf_entries = read_dduf_file(dduf_file_path) + # The reader contains already all the files needed, no need to check it again + cached_folder = "" + + config_dict = cls.load_config(cached_folder, dduf_entries=dduf_entries) + + if dduf_file: + _maybe_raise_error_for_incorrect_transformers(config_dict) + + # pop out "_ignore_files" as it is only needed for download + config_dict.pop("_ignore_files", None) + + # 2. Define which model components should load variants + # We retrieve the information by matching whether variant model checkpoints exist in the subfolders. + # Example: `diffusion_pytorch_model.safetensors` -> `diffusion_pytorch_model.fp16.safetensors` + # with variant being `"fp16"`. + model_variants = _identify_model_variants(folder=cached_folder, variant=variant, config=config_dict) + if len(model_variants) == 0 and variant is not None: + error_message = f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available." + raise ValueError(error_message) + + # 3. Load the pipeline class, if using custom module then load it from the hub + # if we load from explicit class, let's use it + custom_pipeline, custom_class_name = _resolve_custom_pipeline_and_cls( + folder=cached_folder, config=config_dict, custom_pipeline=custom_pipeline + ) + pipeline_class = _get_pipeline_class( + cls, + config=config_dict, + load_connected_pipeline=load_connected_pipeline, + custom_pipeline=custom_pipeline, + class_name=custom_class_name, + cache_dir=cache_dir, + revision=custom_revision, + trust_remote_code=trust_remote_code, + ) + + if device_map is not None and pipeline_class._load_connected_pipes: + raise NotImplementedError("`device_map` is not yet supported for connected pipelines.") + + # DEPRECATED: To be removed in 1.0.0 + # we are deprecating the `StableDiffusionInpaintPipelineLegacy` pipeline which gets loaded + # when a user requests for a `StableDiffusionInpaintPipeline` with `diffusers` version being <= 0.5.1. + _maybe_raise_warning_for_inpainting( + pipeline_class=pipeline_class, + pretrained_model_name_or_path=pretrained_model_name_or_path, + config=config_dict, + ) + + # 4. Define expected modules given pipeline signature + # and define non-None initialized modules (=`init_kwargs`) + + # some modules can be passed directly to the init + # in this case they are already instantiated in `kwargs` + # extract them here + expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class) + expected_types = pipeline_class._get_signature_types() + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs) + + # define init kwargs and make sure that optional component modules are filtered out + init_kwargs = { + k: init_dict.pop(k) + for k in optional_kwargs + if k in init_dict and k not in pipeline_class._optional_components + } + init_kwargs = {**init_kwargs, **passed_pipe_kwargs} + + # remove `null` components + def load_module(name, value): + if value[0] is None: + return False + if name in passed_class_obj and passed_class_obj[name] is None: + return False + return True + + init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)} + + # 5. Throw nice warnings / errors for fast accelerate loading + if len(unused_kwargs) > 0: + logger.warning( + f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored." + ) + + # import it here to avoid circular import + from diffusers import pipelines + + # 6. device map delegation + final_device_map = None + if device_map is not None: + final_device_map = _get_final_device_map( + device_map=device_map, + pipeline_class=pipeline_class, + passed_class_obj=passed_class_obj, + init_dict=init_dict, + library=library, + max_memory=max_memory, + dtype=dtype, + cached_folder=cached_folder, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + ) + + # 7. Load each module in the pipeline + current_device_map = None + _maybe_warn_for_wrong_component_in_quant_config(init_dict, quantization_config) + logging_tqdm_kwargs = {"desc": "Loading pipeline components..."} + if not is_torch_dist_rank_zero(): + logging_tqdm_kwargs["disable"] = True + + for name, (library_name, class_name) in logging.tqdm(init_dict.items(), **logging_tqdm_kwargs): + # 7.1 device_map shenanigans + if final_device_map is not None: + if isinstance(final_device_map, dict) and len(final_device_map) > 0: + component_device = final_device_map.get(name, None) + if component_device is not None: + current_device_map = {"": component_device} + else: + current_device_map = None + elif isinstance(final_device_map, str): + current_device_map = final_device_map + + # 7.2 Define all importable classes + is_pipeline_module = hasattr(pipelines, library_name) + importable_classes = ALL_IMPORTABLE_CLASSES + loaded_sub_model = None + + # 7.3 Use passed sub model or load class_name from library_name + if name in passed_class_obj: + # if the model is in a pipeline module, then we load it from the pipeline + # check that passed_class_obj has correct parent class + maybe_raise_or_warn( + library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module + ) + + loaded_sub_model = passed_class_obj[name] + else: + # load sub model + sub_model_dtype = ( + dtype.get(name, dtype.get("default", torch.float32)) if isinstance(dtype, dict) else dtype + ) + loaded_sub_model = load_sub_model( + library_name=library_name, + class_name=class_name, + importable_classes=importable_classes, + pipelines=pipelines, + is_pipeline_module=is_pipeline_module, + pipeline_class=pipeline_class, + dtype=sub_model_dtype, + provider=provider, + sess_options=sess_options, + device_map=current_device_map, + max_memory=max_memory, + offload_folder=offload_folder, + offload_state_dict=offload_state_dict, + model_variants=model_variants, + name=name, + variant=variant, + low_cpu_mem_usage=low_cpu_mem_usage, + cached_folder=cached_folder, + use_safetensors=use_safetensors, + dduf_entries=dduf_entries, + provider_options=provider_options, + disable_mmap=disable_mmap, + quantization_config=quantization_config, + use_flashpack=use_flashpack, + trust_remote_code=trust_remote_code, + ) + logger.info( + f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}." + ) + + init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...) + + # 8. Handle connected pipelines. + if pipeline_class._load_connected_pipes and os.path.isfile(os.path.join(cached_folder, "README.md")): + init_kwargs = _update_init_kwargs_with_connected_pipeline( + init_kwargs=init_kwargs, + passed_pipe_kwargs=passed_pipe_kwargs, + passed_class_objs=passed_class_obj, + folder=cached_folder, + **kwargs_copied, + ) + + # 9. Potentially add passed objects if expected + missing_modules = set(expected_modules) - set(init_kwargs.keys()) + passed_modules = list(passed_class_obj.keys()) + optional_modules = pipeline_class._optional_components + if len(missing_modules) > 0 and missing_modules <= set(passed_modules + optional_modules): + for module in missing_modules: + init_kwargs[module] = passed_class_obj.get(module, None) + elif len(missing_modules) > 0: + passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - set(optional_kwargs) + raise ValueError( + f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed." + ) + + # 10. Type checking init arguments + for kw, arg in init_kwargs.items(): + # Too complex to validate with type annotation alone + if "scheduler" in kw: + continue + # Many tokenizer annotations don't include its "Fast" variant, so skip this + # e.g T5Tokenizer but not T5TokenizerFast + elif "tokenizer" in kw: + continue + elif ( + arg is not None # Skip if None + and not expected_types[kw] == (inspect.Signature.empty,) # Skip if no type annotations + and not _is_valid_type(arg, expected_types[kw]) # Check type + ): + logger.warning(f"Expected types for {kw}: {expected_types[kw]}, got {_get_detailed_type(arg)}.") + + # 11. Instantiate the pipeline + model = pipeline_class(**init_kwargs) + + # 12. Save where the model was instantiated from + model.register_to_config(_name_or_path=pretrained_model_name_or_path) + if device_map is not None: + setattr(model, "hf_device_map", final_device_map) + if quantization_config is not None: + setattr(model, "quantization_config", quantization_config) + return model + + @property + def name_or_path(self) -> str: + return getattr(self.config, "_name_or_path", None) + + @property + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from + Accelerate's module hooks. + """ + from ..hooks.group_offloading import _get_group_onload_device + + # When apply group offloading at the leaf_level, we're in the same situation as accelerate's sequential + # offloading. We need to return the onload device of the group offloading hooks so that the intermediates + # required for computation (latents, prompt embeddings, etc.) can be created on the correct device. + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module): + continue + try: + return _get_group_onload_device(model) + except ValueError: + pass + + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module) or name in self._exclude_from_cpu_offload: + continue + + if not hasattr(model, "_hf_hook"): + return self.device + for module in model.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + + def remove_all_hooks(self): + r""" + Removes all hooks that were added when using `enable_sequential_cpu_offload` or `enable_model_cpu_offload`. + """ + for _, model in self.components.items(): + if isinstance(model, torch.nn.Module) and hasattr(model, "_hf_hook"): + accelerate.hooks.remove_hook_from_module(model, recurse=True) + self._all_hooks = [] + + def enable_model_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the accelerator when its + `forward` method is called, and the model remains in accelerator until the next model runs. Memory savings are + lower than with `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution + of the `unet`. + + Arguments: + gpu_id (`int`, *optional*): + The ID of the accelerator that shall be used in inference. If not specified, it will default to 0. + device (`torch.Device` or `str`, *optional*, defaults to None): + The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will + automatically detect the available accelerator and use. + """ + self._maybe_raise_error_if_group_offload_active(raise_error=True) + + is_pipeline_device_mapped = self._is_pipeline_device_mapped() + if is_pipeline_device_mapped: + raise ValueError( + "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_model_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_model_cpu_offload()`." + ) + + if self.model_cpu_offload_seq is None: + raise ValueError( + "Model CPU offload cannot be enabled because no `model_cpu_offload_seq` class attribute is set." + ) + + if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): + from accelerate import cpu_offload_with_hook + else: + raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + + self.remove_all_hooks() + + if device is None: + device = get_device() + if device == "cpu": + raise RuntimeError("`enable_model_cpu_offload` requires accelerator, but not found") + + torch_device = torch.device(device) + device_index = torch_device.index + + if gpu_id is not None and device_index is not None: + raise ValueError( + f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" + f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}" + ) + + # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0 + self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) + + device_type = torch_device.type + device = torch.device(f"{device_type}:{self._offload_gpu_id}") + self._offload_device = device + + self.to("cpu", silence_dtype_warnings=True) + empty_device_cache(device.type) + + all_model_components = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)} + + self._all_hooks = [] + hook = None + for model_str in self.model_cpu_offload_seq.split("->"): + model = all_model_components.pop(model_str, None) + + if not isinstance(model, torch.nn.Module): + continue + + # This is because the model would already be placed on a CUDA device. + _, _, is_loaded_in_8bit_bnb = _check_bnb_status(model) + if is_loaded_in_8bit_bnb and ( + is_transformers_version("<", "4.58.0") or is_bitsandbytes_version("<", "0.48.0") + ): + logger.info( + f"Skipping the hook placement for the {model.__class__.__name__} as it is loaded in `bitsandbytes` 8bit." + ) + continue + + _, hook = cpu_offload_with_hook(model, device, prev_module_hook=hook) + self._all_hooks.append(hook) + + # CPU offload models that are not in the seq chain unless they are explicitly excluded + # these models will stay on CPU until maybe_free_model_hooks is called + # some models cannot be in the seq chain because they are iteratively called, such as controlnet + for name, model in all_model_components.items(): + if not isinstance(model, torch.nn.Module): + continue + + if name in self._exclude_from_cpu_offload: + model.to(device) + else: + _, hook = cpu_offload_with_hook(model, device) + self._all_hooks.append(hook) + + def maybe_free_model_hooks(self): + r""" + Method that performs the following: + - Offloads all components. + - Removes all model hooks that were added when using `enable_model_cpu_offload`, and then applies them again. + In case the model has not been offloaded, this function is a no-op. + - Resets stateful diffusers hooks of denoiser components if they were added with + [`~hooks.HookRegistry.register_hook`]. + + Make sure to add this function to the end of the `__call__` function of your pipeline so that it functions + correctly when applying `enable_model_cpu_offload`. + """ + for component in self.components.values(): + if hasattr(component, "_reset_stateful_cache"): + component._reset_stateful_cache() + + if not hasattr(self, "_all_hooks") or len(self._all_hooks) == 0: + # `enable_model_cpu_offload` has not be called, so silently do nothing + return + + # make sure the model is in the same state as before calling it + self.enable_model_cpu_offload(device=getattr(self, "_offload_device", "cuda")) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using 🤗 Accelerate, significantly reducing memory usage. When called, the state + dicts of all `torch.nn.Module` components (except those in `self._exclude_from_cpu_offload`) are saved to CPU + and then moved to `torch.device('meta')` and loaded to accelerator only when their specific submodule has its + `forward` method called. Offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + + Arguments: + gpu_id (`int`, *optional*): + The ID of the accelerator that shall be used in inference. If not specified, it will default to 0. + device (`torch.Device` or `str`, *optional*, defaults to None): + The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will + automatically detect the available accelerator and use. + """ + self._maybe_raise_error_if_group_offload_active(raise_error=True) + + if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"): + from accelerate import cpu_offload + else: + raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher") + self.remove_all_hooks() + + is_pipeline_device_mapped = self._is_pipeline_device_mapped() + if is_pipeline_device_mapped: + raise ValueError( + "It seems like you have activated a device mapping strategy on the pipeline so calling `enable_sequential_cpu_offload() isn't allowed. You can call `reset_device_map()` first and then call `enable_sequential_cpu_offload()`." + ) + + if device is None: + device = get_device() + if device == "cpu": + raise RuntimeError("`enable_sequential_cpu_offload` requires accelerator, but not found") + + torch_device = torch.device(device) + device_index = torch_device.index + + if gpu_id is not None and device_index is not None: + raise ValueError( + f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}" + f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}" + ) + + # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0 + self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0) + + device_type = torch_device.type + device = torch.device(f"{device_type}:{self._offload_gpu_id}") + self._offload_device = device + + if self.device.type != "cpu": + orig_device_type = self.device.type + self.to("cpu", silence_dtype_warnings=True) + empty_device_cache(orig_device_type) + + for name, model in self.components.items(): + if not isinstance(model, torch.nn.Module): + continue + + if name in self._exclude_from_cpu_offload: + model.to(device) + else: + # make sure to offload buffers if not all high level weights + # are of type nn.Module + offload_buffers = len(model._parameters) > 0 + cpu_offload(model, device, offload_buffers=offload_buffers) + + def enable_group_offload( + self, + onload_device: torch.device, + offload_device: torch.device = torch.device("cpu"), + offload_type: str = "block_level", + num_blocks_per_group: int | None = None, + non_blocking: bool = False, + use_stream: bool = False, + record_stream: bool = False, + low_cpu_mem_usage=False, + offload_to_disk_path: str | None = None, + exclude_modules: str | list[str] | None = None, + ) -> None: + r""" + Applies group offloading to the internal layers of a torch.nn.Module. To understand what group offloading is, + and where it is beneficial, we need to first provide some context on how other supported offloading methods + work. + + Typically, offloading is done at two levels: + - Module-level: In Diffusers, this can be enabled using the `ModelMixin::enable_model_cpu_offload()` method. It + works by offloading each component of a pipeline to the CPU for storage, and onloading to the accelerator + device when needed for computation. This method is more memory-efficient than keeping all components on the + accelerator, but the memory requirements are still quite high. For this method to work, one needs memory + equivalent to size of the model in runtime dtype + size of largest intermediate activation tensors to be able + to complete the forward pass. + - Leaf-level: In Diffusers, this can be enabled using the `ModelMixin::enable_sequential_cpu_offload()` method. + It + works by offloading the lowest leaf-level parameters of the computation graph to the CPU for storage, and + onloading only the leafs to the accelerator device for computation. This uses the lowest amount of accelerator + memory, but can be slower due to the excessive number of device synchronizations. + + Group offloading is a middle ground between the two methods. It works by offloading groups of internal layers, + (either `torch.nn.ModuleList` or `torch.nn.Sequential`). This method uses lower memory than module-level + offloading. It is also faster than leaf-level/sequential offloading, as the number of device synchronizations + is reduced. + + Another supported feature (for CUDA devices with support for asynchronous data transfer streams) is the ability + to overlap data transfer and computation to reduce the overall execution time compared to sequential + offloading. This is enabled using layer prefetching with streams, i.e., the layer that is to be executed next + starts onloading to the accelerator device while the current layer is being executed - this increases the + memory requirements slightly. Note that this implementation also supports leaf-level offloading but can be made + much faster when using streams. + + Args: + onload_device (`torch.device`): + The device to which the group of modules are onloaded. + offload_device (`torch.device`, defaults to `torch.device("cpu")`): + The device to which the group of modules are offloaded. This should typically be the CPU. Default is + CPU. + offload_type (`str` or `GroupOffloadingType`, defaults to "block_level"): + The type of offloading to be applied. Can be one of "block_level" or "leaf_level". Default is + "block_level". + offload_to_disk_path (`str`, *optional*, defaults to `None`): + The path to the directory where parameters will be offloaded. Setting this option can be useful in + limited RAM environment settings where a reasonable speed-memory trade-off is desired. + num_blocks_per_group (`int`, *optional*): + The number of blocks per group when using offload_type="block_level". This is required when using + offload_type="block_level". + non_blocking (`bool`, defaults to `False`): + If True, offloading and onloading is done with non-blocking data transfer. + use_stream (`bool`, defaults to `False`): + If True, offloading and onloading is done asynchronously using a CUDA stream. This can be useful for + overlapping computation and data transfer. + record_stream (`bool`, defaults to `False`): When enabled with `use_stream`, it marks the current tensor + as having been used by this stream. It is faster at the expense of slightly more memory usage. Refer to + the [PyTorch official docs](https://pytorch.org/docs/stable/generated/torch.Tensor.record_stream.html) + more details. + low_cpu_mem_usage (`bool`, defaults to `False`): + If True, the CPU memory usage is minimized by pinning tensors on-the-fly instead of pre-pinning them. + This option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be + useful when the CPU memory is a bottleneck but may counteract the benefits of using streams. + exclude_modules (`Union[str, List[str]]`, defaults to `None`): List of modules to exclude from offloading. + + Example: + ```python + >>> from diffusers import DiffusionPipeline + >>> import torch + + >>> pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", dtype=torch.bfloat16) + + >>> pipe.enable_group_offload( + ... onload_device=torch.device("cuda"), + ... offload_device=torch.device("cpu"), + ... offload_type="leaf_level", + ... use_stream=True, + ... ) + >>> image = pipe("a beautiful sunset").images[0] + ``` + """ + from ..hooks import apply_group_offloading + + if isinstance(exclude_modules, str): + exclude_modules = [exclude_modules] + elif exclude_modules is None: + exclude_modules = [] + + unknown = set(exclude_modules) - self.components.keys() + if unknown: + logger.info( + f"The following modules are not present in pipeline: {', '.join(unknown)}. Ignore if this is expected." + ) + + group_offload_kwargs = { + "onload_device": onload_device, + "offload_device": offload_device, + "offload_type": offload_type, + "num_blocks_per_group": num_blocks_per_group, + "non_blocking": non_blocking, + "use_stream": use_stream, + "record_stream": record_stream, + "low_cpu_mem_usage": low_cpu_mem_usage, + "offload_to_disk_path": offload_to_disk_path, + } + for name, component in self.components.items(): + if name not in exclude_modules and isinstance(component, torch.nn.Module): + if hasattr(component, "enable_group_offload"): + component.enable_group_offload(**group_offload_kwargs) + else: + apply_group_offloading(module=component, **group_offload_kwargs) + + if exclude_modules: + for module_name in exclude_modules: + module = getattr(self, module_name, None) + if module is not None and isinstance(module, torch.nn.Module): + module.to(onload_device) + logger.debug(f"Placed `{module_name}` on {onload_device} device as it was in `exclude_modules`.") + + def reset_device_map(self): + r""" + Resets the device maps (if any) to None. + """ + if self.hf_device_map is None: + return + else: + self.remove_all_hooks() + for name, component in self.components.items(): + if isinstance(component, torch.nn.Module): + component.to("cpu") + self.hf_device_map = None + + @classmethod + @validate_hf_hub_args + def download(cls, pretrained_model_name, **kwargs) -> str | os.PathLike: + r""" + Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights. + + Parameters: + pretrained_model_name (`str` or `os.PathLike`, *optional*): + A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline + hosted on the Hub. + custom_pipeline (`str`, *optional*): + Can be either: + + - A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained + pipeline hosted on the Hub. The repository must contain a file called `pipeline.py` that defines + the custom pipeline. + + - A string, the *file name* of a community pipeline hosted on GitHub under + [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file + names must match the file name and not the pipeline script (`clip_guided_stable_diffusion` + instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the + current `main` branch of GitHub. + + - A path to a *directory* (`./my_pipeline_directory/`) containing a custom pipeline. The directory + must contain a file called `pipeline.py` that defines the custom pipeline. + + > [!WARNING] > 🧪 This is an experimental feature and may change in the future. + + For more information on how to load and create custom pipelines, take a look at [How to contribute a + community pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/contribute_pipeline). + + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + + proxies (`Dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info(`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + custom_revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id similar to + `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a + custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub. + mirror (`str`, *optional*): + Mirror source to resolve accessibility issues if you're downloading a model in China. We do not + guarantee the timeliness or safety of the source, and you should refer to the mirror site for more + information. + variant (`str`, *optional*): + Load weights from a specified variant filename such as `"fp16"` or `"ema"`. + dduf_file(`str`, *optional*): + Load weights from the specified DDUF file. This argument is deprecated and will be removed + in version 0.41.0. + use_safetensors (`bool`, *optional*, defaults to `None`): + If set to `None`, the safetensors weights are downloaded if they're available **and** if the + safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors + weights. If set to `False`, safetensors weights are not loaded. + use_onnx (`bool`, *optional*, defaults to `False`): + If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights + will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is + `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending + with `.onnx` and `.pb`. + trust_remote_code (`bool`, *optional*, defaults to `False`): + Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This + option should only be set to `True` for repositories you trust and in which you have read the code, as + it will execute code present on the Hub on your local machine. + use_flashpack (`bool`, *optional*, defaults to `False`): + If set to `True`, FlashPack weights will always be downloaded if present. If set to `False`, FlashPack + weights will never be downloaded. + + Returns: + `os.PathLike`: + A path to the downloaded pipeline. + + > [!TIP] > To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in + with `hf > auth login + + """ + cache_dir = kwargs.pop("cache_dir", None) + force_download = kwargs.pop("force_download", False) + proxies = kwargs.pop("proxies", None) + local_files_only = kwargs.pop("local_files_only", None) + token = kwargs.pop("token", None) + revision = kwargs.pop("revision", None) + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + variant = kwargs.pop("variant", None) + use_safetensors = kwargs.pop("use_safetensors", None) + use_onnx = kwargs.pop("use_onnx", None) + load_connected_pipeline = kwargs.pop("load_connected_pipeline", False) + trust_remote_code = kwargs.pop("trust_remote_code", False) + dduf_file: dict[str, DDUFEntry] | None = kwargs.pop("dduf_file", None) + use_flashpack = kwargs.pop("use_flashpack", False) + + if dduf_file: + deprecate("dduf_file", "0.41.0", _DDUF_DEPRECATION_MESSAGE) + if custom_pipeline: + raise NotImplementedError("Custom pipelines are not supported with DDUF at the moment.") + if load_connected_pipeline: + raise NotImplementedError("Connected pipelines are not supported with DDUF at the moment.") + return _download_dduf_file( + pretrained_model_name=pretrained_model_name, + dduf_file=dduf_file, + pipeline_class_name=cls.__name__, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + ) + + allow_pickle = True if (use_safetensors is None or use_safetensors is False) else False + use_safetensors = use_safetensors if use_safetensors is not None else True + + allow_patterns = None + ignore_patterns = None + + model_info_call_error: Exception | None = None + if not local_files_only: + try: + info = model_info(pretrained_model_name, token=token, revision=revision) + except (HfHubHTTPError, OfflineModeIsEnabled, requests.ConnectionError, httpx.NetworkError) as e: + logger.warning(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.") + local_files_only = True + model_info_call_error = e # save error to reraise it if model is not cached locally + + if not local_files_only: + config_file = hf_hub_download( + pretrained_model_name, + cls.config_name, + cache_dir=cache_dir, + revision=revision, + proxies=proxies, + force_download=force_download, + token=token, + ) + filenames = {sibling.rfilename for sibling in info.siblings} + else: + # We are offline (either the user asked for `local_files_only` or the `model_info` call above + # failed), so the repo's file listing needed to compute the allow/ignore patterns below cannot be + # fetched from the Hub. Read it from the local cache with `get_cached_repo_tree` instead: it returns + # the file listing that a previous online call cached, so both code paths compute the same patterns. + # If nothing is cached for this repo, skip straight to `snapshot_download` below, which raises the + # appropriate offline not-found error. + try: + config_file = hf_hub_download( + pretrained_model_name, + cls.config_name, + cache_dir=cache_dir, + revision=revision, + token=token, + local_files_only=True, + ) + filenames = { + f.path for f in get_cached_repo_tree(pretrained_model_name, revision=revision, cache_dir=cache_dir) + } + except (LocalEntryNotFoundError, CachedRepoTreeNotFoundError): + config_file = None + + if config_file is not None: + snapshot_folder = Path(config_file).parent + + config_dict = cls._dict_from_json_file(config_file) + ignore_filenames = config_dict.pop("_ignore_files", []) + + if variant is not None and _check_legacy_sharding_variant_format(filenames=filenames, variant=variant): + warn_msg = ( + f"Warning: The repository contains sharded checkpoints for variant '{variant}' maybe in a deprecated format. " + "Please check your files carefully:\n\n" + "- Correct format example: diffusion_pytorch_model.fp16-00003-of-00003.safetensors\n" + "- Deprecated format example: diffusion_pytorch_model-00001-of-00002.fp16.safetensors\n\n" + "If you find any files in the deprecated format:\n" + "1. Remove all existing checkpoint files for this variant.\n" + "2. Re-obtain the correct files by running `save_pretrained()`.\n\n" + "This will ensure you're using the most up-to-date and compatible checkpoint format." + ) + logger.warning(warn_msg) + + filenames = set(filenames) - set(ignore_filenames) + if ( + not local_files_only + and revision in DEPRECATED_REVISION_ARGS + and version.parse(version.parse(__version__).base_version) >= version.parse("0.22.0") + ): + warn_deprecated_model_variant(pretrained_model_name, token, variant, revision, filenames) + + custom_components, folder_names = _get_custom_components_and_folders( + pretrained_model_name, config_dict, filenames, variant + ) + custom_class_name = None + if custom_pipeline is None and isinstance(config_dict["_class_name"], (list, tuple)): + custom_pipeline = config_dict["_class_name"][0] + custom_class_name = config_dict["_class_name"][1] + + load_pipe_from_hub = custom_pipeline is not None and f"{custom_pipeline}.py" in filenames + + # retrieve passed components that should not be downloaded + pipeline_class = _get_pipeline_class( + cls, + config_dict, + load_connected_pipeline=load_connected_pipeline, + custom_pipeline=custom_pipeline, + repo_id=pretrained_model_name if load_pipe_from_hub else None, + hub_revision=revision, + class_name=custom_class_name, + cache_dir=cache_dir, + revision=custom_revision, + trust_remote_code=trust_remote_code, + ) + expected_components, _ = cls._get_signature_keys(pipeline_class) + passed_components = [k for k in expected_components if k in kwargs] + + # retrieve the names of the folders containing model weights + model_folder_names = { + os.path.split(f)[0] for f in filter_model_files(filenames) if os.path.split(f)[0] in folder_names + } + # retrieve all patterns that should not be downloaded and error out when needed + ignore_patterns = _get_ignore_patterns( + passed_components, + model_folder_names, + filenames, + use_safetensors, + allow_pickle, + use_onnx, + pipeline_class._is_onnx, + use_flashpack, + variant, + ) + + model_filenames, variant_filenames = variant_compatible_siblings( + filenames, variant=variant, ignore_patterns=ignore_patterns + ) + + # all filenames compatible with variant will be added + allow_patterns = list(model_filenames) + + # allow all patterns from non-model folders + # this enables downloading schedulers, tokenizers, ... + allow_patterns += [f"{k}/*" for k in folder_names if k not in model_folder_names] + # add custom component files + allow_patterns += [f"{k}/{f}.py" for k, f in custom_components.items()] + # add custom pipeline file + allow_patterns += [f"{custom_pipeline}.py"] if f"{custom_pipeline}.py" in filenames else [] + # also allow downloading config.json files with the model + allow_patterns += [os.path.join(k, "config.json") for k in model_folder_names] + allow_patterns += [ + SCHEDULER_CONFIG_NAME, + CONFIG_NAME, + cls.config_name, + CUSTOM_PIPELINE_FILE_NAME, + ] + + # Don't download any objects that are passed + allow_patterns = [ + p for p in allow_patterns if not (len(p.split("/")) == 2 and p.split("/")[0] in passed_components) + ] + + # Repos with a flat, transformers-style layout host a component's files at the repo root instead of + # in a subfolder, where the folder-based allow patterns above miss its auxiliary files (root-hosted + # weights are already included via `model_filenames`, root `config.json` via `CONFIG_NAME`). + allow_patterns += TRANSFORMERS_COMPONENT_AUX_FILES + + if pipeline_class._load_connected_pipes: + allow_patterns.append("README.md") + + # Don't download index files of forbidden patterns either + ignore_patterns = ignore_patterns + [f"{i}.index.*json" for i in ignore_patterns] + re_ignore_pattern = [re.compile(fnmatch.translate(p)) for p in ignore_patterns] + re_allow_pattern = [re.compile(fnmatch.translate(p)) for p in allow_patterns] + + expected_files = [f for f in filenames if not any(p.match(f) for p in re_ignore_pattern)] + expected_files = [f for f in expected_files if any(p.match(f) for p in re_allow_pattern)] + + pipeline_is_cached = all((snapshot_folder / f).is_file() for f in expected_files) + + if pipeline_is_cached and not force_download: + # if the pipeline is cached, we can directly return it + # else call snapshot_download + return snapshot_folder + + user_agent = {"pipeline_class": cls.__name__} + if custom_pipeline is not None and not custom_pipeline.endswith(".py"): + user_agent["custom_pipeline"] = custom_pipeline + + # download all allow_patterns - ignore_patterns + try: + cached_folder = snapshot_download( + pretrained_model_name, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=allow_patterns, + ignore_patterns=ignore_patterns, + user_agent=user_agent, + ) + + cls_name = cls.load_config(os.path.join(cached_folder, "model_index.json")).get("_class_name", None) + + diffusers_module = importlib.import_module(__name__.split(".")[0]) + pipeline_class = getattr(diffusers_module, cls_name, None) if isinstance(cls_name, str) else None + + if pipeline_class is not None and pipeline_class._load_connected_pipes: + modelcard = ModelCard.load(os.path.join(cached_folder, "README.md")) + connected_pipes = sum([getattr(modelcard.data, k, []) for k in CONNECTED_PIPES_KEYS], []) + for connected_pipe_repo_id in connected_pipes: + download_kwargs = { + "cache_dir": cache_dir, + "force_download": force_download, + "proxies": proxies, + "local_files_only": local_files_only, + "token": token, + "variant": variant, + "use_safetensors": use_safetensors, + } + DiffusionPipeline.download(connected_pipe_repo_id, **download_kwargs) + + return cached_folder + + except FileNotFoundError as e: + # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache. + # This can happen in two cases: + # 1. If the user passed `local_files_only=True` => we raise the error directly + # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error + if model_info_call_error is None: + # 1. user passed `local_files_only=True` + raise + else: + # 2. we forced `local_files_only=True` when `model_info` failed + raise EnvironmentError( + f"Cannot load model {pretrained_model_name}: the model is not fully cached locally ({e}) and an" + " error occurred while trying to fetch metadata from the Hub. Please check out the root cause in" + " the stacktrace above." + ) from model_info_call_error + + @classmethod + def _get_signature_keys(cls, obj): + parameters = inspect.signature(obj.__init__).parameters + required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty} + optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty}) + expected_modules = set(required_parameters.keys()) - {"self"} + + optional_names = list(optional_parameters) + for name in optional_names: + if name in cls._optional_components: + expected_modules.add(name) + optional_parameters.remove(name) + + return sorted(expected_modules), sorted(optional_parameters) + + @classmethod + def _get_signature_types(cls): + signature_types = {} + # Use get_type_hints to properly resolve string annotations (from __future__ import annotations) + try: + type_hints = get_type_hints(cls.__init__) + except Exception: + # Fallback to direct annotation access if get_type_hints fails + type_hints = {} + for k, v in inspect.signature(cls.__init__).parameters.items(): + if v.annotation != inspect.Parameter.empty: + type_hints[k] = v.annotation + + # Get all parameters from the signature to ensure we don't miss any + all_params = inspect.signature(cls.__init__).parameters + + for param_name, param in all_params.items(): + # Skip 'self' parameter + if param_name == "self": + continue + + # If we have type hints, use them + if param_name in type_hints: + annotation = type_hints[param_name] + if inspect.isclass(annotation): + signature_types[param_name] = (annotation,) + elif get_origin(annotation) == Union: + signature_types[param_name] = get_args(annotation) + elif isinstance(annotation, types.UnionType): + # Handle PEP 604 union syntax (X | Y) introduced in Python 3.10+ + signature_types[param_name] = get_args(annotation) + elif get_origin(annotation) in [List, Dict, list, dict]: + signature_types[param_name] = (annotation,) + else: + logger.warning(f"cannot get type annotation for Parameter {param_name} of {cls}.") + # Still add it with empty signature so it's in expected_types + signature_types[param_name] = (inspect.Signature.empty,) + else: + # No type annotation found - add with empty signature + signature_types[param_name] = (inspect.Signature.empty,) + + return signature_types + + @property + def parameters(self) -> dict[str, Any]: + r""" + The `self.parameters` property can be useful to run different pipelines with the same weights and + configurations without reallocating additional memory. + + Returns (`dict`): + A dictionary containing all the optional parameters needed to initialize the pipeline. + + Examples: + + ```py + >>> from diffusers import ( + ... StableDiffusionPipeline, + ... StableDiffusionImg2ImgPipeline, + ... StableDiffusionInpaintPipeline, + ... ) + + >>> text2img = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components, **text2img.parameters) + >>> inpaint = StableDiffusionInpaintPipeline(**text2img.components, **text2img.parameters) + ``` + """ + expected_modules, optional_parameters = self._get_signature_keys(self) + pipeline_parameters = { + k: self.config[k] for k in self.config.keys() if not k.startswith("_") and k in optional_parameters + } + + return pipeline_parameters + + @property + def components(self) -> dict[str, Any]: + r""" + The `self.components` property can be useful to run different pipelines with the same weights and + configurations without reallocating additional memory. + + Returns (`dict`): + A dictionary containing all the modules needed to initialize the pipeline. + + Examples: + + ```py + >>> from diffusers import ( + ... StableDiffusionPipeline, + ... StableDiffusionImg2ImgPipeline, + ... StableDiffusionInpaintPipeline, + ... ) + + >>> text2img = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components) + >>> inpaint = StableDiffusionInpaintPipeline(**text2img.components) + ``` + """ + expected_modules, optional_parameters = self._get_signature_keys(self) + components = { + k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters + } + + actual = sorted(set(components.keys())) + expected = sorted(expected_modules) + if actual != expected: + raise ValueError( + f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected" + f" {expected} to be defined, but {actual} are defined." + ) + + return components + + @staticmethod + def numpy_to_pil(images): + """ + Convert a NumPy image or a batch of images to a PIL image. + """ + return numpy_to_pil(images) + + @torch.compiler.disable + def progress_bar(self, iterable=None, total=None): + if not hasattr(self, "_progress_bar_config"): + self._progress_bar_config = {} + elif not isinstance(self._progress_bar_config, dict): + raise ValueError( + f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}." + ) + + progress_bar_config = dict(self._progress_bar_config) + if "disable" not in progress_bar_config: + progress_bar_config["disable"] = not is_torch_dist_rank_zero() + + if iterable is not None: + return tqdm(iterable, **progress_bar_config) + elif total is not None: + return tqdm(total=total, **progress_bar_config) + else: + raise ValueError("Either `total` or `iterable` has to be defined.") + + def set_progress_bar_config(self, **kwargs): + self._progress_bar_config = kwargs + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + r""" + Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). When this + option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed + up during training is not guaranteed. + + > [!WARNING] > ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient + attention takes > precedent. + + Parameters: + attention_op (`Callable`, *optional*): + Override the default `None` operator for use as `op` argument to the + [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention) + function of xFormers. + + Examples: + + ```py + >>> import torch + >>> from diffusers import DiffusionPipeline + >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp + + >>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", dtype=torch.float16) + >>> pipe = pipe.to("cuda") + >>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp) + >>> # Workaround for not accepting attention shape using VAE for Flash Attention + >>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None) + ``` + """ + self.set_use_memory_efficient_attention_xformers(True, attention_op) + + def disable_xformers_memory_efficient_attention(self): + r""" + Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). + """ + self.set_use_memory_efficient_attention_xformers(False) + + def set_use_memory_efficient_attention_xformers(self, valid: bool, attention_op: Callable | None = None) -> None: + # Recursively walk through all the children. + # Any children which exposes the set_use_memory_efficient_attention_xformers method + # gets the message + def fn_recursive_set_mem_eff(module: torch.nn.Module): + if hasattr(module, "set_use_memory_efficient_attention_xformers"): + module.set_use_memory_efficient_attention_xformers(valid, attention_op) + + for child in module.children(): + fn_recursive_set_mem_eff(child) + + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module)] + + for module in modules: + fn_recursive_set_mem_eff(module) + + def enable_attention_slicing(self, slice_size: str | int = "auto"): + r""" + Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor + in slices to compute attention in several steps. For more than one attention head, the computation is performed + sequentially over each head. This is useful to save some memory in exchange for a small speed decrease. + + > [!WARNING] > ⚠️ Don't enable attention slicing if you're already using `scaled_dot_product_attention` (SDPA) + from PyTorch > 2.0 or xFormers. These attention computations are already very memory efficient so you won't + need to enable > this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious + slow downs! + + Args: + slice_size (`str` or `int`, *optional*, defaults to `"auto"`): + When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If + `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + + Examples: + + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", + ... dtype=torch.float16, + ... use_safetensors=True, + ... ) + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> pipe.enable_attention_slicing() + >>> image = pipe(prompt).images[0] + ``` + """ + self.set_attention_slice(slice_size) + + def disable_attention_slicing(self): + r""" + Disable sliced attention computation. If `enable_attention_slicing` was previously called, attention is + computed in one step. + """ + # set slice_size = `None` to disable `attention slicing` + self.enable_attention_slicing(None) + + def set_attention_slice(self, slice_size: int | None): + module_names, _ = self._get_signature_keys(self) + modules = [getattr(self, n, None) for n in module_names] + modules = [m for m in modules if isinstance(m, torch.nn.Module) and hasattr(m, "set_attention_slice")] + + for module in modules: + module.set_attention_slice(slice_size) + + @classmethod + def from_pipe(cls, pipeline, **kwargs): + r""" + Create a new pipeline from a given pipeline. This method is useful to create a new pipeline from the existing + pipeline components without reallocating additional memory. + + Arguments: + pipeline (`DiffusionPipeline`): + The pipeline from which to create a new pipeline. + + Returns: + `DiffusionPipeline`: + A new pipeline with the same weights and configurations as `pipeline`. + + Examples: + + ```py + >>> from diffusers import StableDiffusionPipeline, StableDiffusionSAGPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") + >>> new_pipe = StableDiffusionSAGPipeline.from_pipe(pipe) + ``` + """ + + original_config = dict(pipeline.config) + torch_dtype = kwargs.pop("torch_dtype", None) + dtype = _resolve_dtype(kwargs.pop("dtype", None), torch_dtype) or torch.float32 + trust_remote_code = kwargs.pop("trust_remote_code", False) + + # derive the pipeline class to instantiate + custom_pipeline = kwargs.pop("custom_pipeline", None) + custom_revision = kwargs.pop("custom_revision", None) + + if custom_pipeline is not None: + pipeline_class = _get_custom_pipeline_class( + custom_pipeline, revision=custom_revision, trust_remote_code=trust_remote_code + ) + else: + pipeline_class = cls + + expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class) + # true_optional_modules are optional components with default value in signature so it is ok not to pass them to `__init__` + # e.g. `image_encoder` for StableDiffusionPipeline + parameters = inspect.signature(cls.__init__).parameters + true_optional_modules = set( + {k for k, v in parameters.items() if v.default != inspect._empty and k in expected_modules} + ) + + # get the class of each component based on its type hint + # e.g. {"unet": UNet2DConditionModel, "text_encoder": CLIPTextMode} + component_types = pipeline_class._get_signature_types() + + pretrained_model_name_or_path = original_config.pop("_name_or_path", None) + # allow users pass modules in `kwargs` to override the original pipeline's components + passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs} + + original_class_obj = {} + for name, component in pipeline.components.items(): + if name in expected_modules and name not in passed_class_obj: + # for model components, we will not switch over if the class does not matches the type hint in the new pipeline's signature + if ( + not isinstance(component, ModelMixin) + or type(component) in component_types[name] + or (component is None and name in cls._optional_components) + ): + original_class_obj[name] = component + else: + logger.warning( + f"component {name} is not switched over to new pipeline because type does not match the expected." + f" {name} is {type(component)} while the new pipeline expect {component_types[name]}." + f" please pass the component of the correct type to the new pipeline. `from_pipe(..., {name}={name})`" + ) + + # allow users pass optional kwargs to override the original pipelines config attribute + passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs} + original_pipe_kwargs = { + k: original_config[k] + for k in original_config.keys() + if k in optional_kwargs and k not in passed_pipe_kwargs + } + + # config attribute that were not expected by pipeline is stored as its private attribute + # (i.e. when the original pipeline was also instantiated with `from_pipe` from another pipeline that has this config) + # in this case, we will pass them as optional arguments if they can be accepted by the new pipeline + additional_pipe_kwargs = [ + k[1:] + for k in original_config.keys() + if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs + ] + for k in additional_pipe_kwargs: + original_pipe_kwargs[k] = original_config.pop(f"_{k}") + + pipeline_kwargs = { + **passed_class_obj, + **original_class_obj, + **passed_pipe_kwargs, + **original_pipe_kwargs, + **kwargs, + } + + # store unused config as private attribute in the new pipeline + unused_original_config = { + f"{'' if k.startswith('_') else '_'}{k}": v for k, v in original_config.items() if k not in pipeline_kwargs + } + + optional_components = ( + pipeline._optional_components + if hasattr(pipeline, "_optional_components") and pipeline._optional_components + else [] + ) + missing_modules = ( + set(expected_modules) - set(optional_components) - set(pipeline_kwargs.keys()) - set(true_optional_modules) + ) + + if len(missing_modules) > 0: + raise ValueError( + f"Pipeline {pipeline_class} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed" + ) + + new_pipeline = pipeline_class(**pipeline_kwargs) + if pretrained_model_name_or_path is not None: + new_pipeline.register_to_config(_name_or_path=pretrained_model_name_or_path) + new_pipeline.register_to_config(**unused_original_config) + + new_pipeline.to(dtype=dtype) + + return new_pipeline + + def _maybe_raise_error_if_group_offload_active( + self, raise_error: bool = False, module: torch.nn.Module | None = None + ) -> bool: + from ..hooks.group_offloading import _is_group_offload_enabled + + components = self.components.values() if module is None else [module] + components = [component for component in components if isinstance(component, torch.nn.Module)] + for component in components: + if _is_group_offload_enabled(component): + if raise_error: + raise ValueError( + "You are trying to apply model/sequential CPU offloading to a pipeline that contains components " + "with group offloading enabled. This is not supported. Please disable group offloading for " + "components of the pipeline to use other offloading methods." + ) + return True + return False + + def _is_pipeline_device_mapped(self): + # We support passing `device_map="cuda"`, for example. This is helpful, in case + # users want to pass `device_map="cpu"` when initializing a pipeline. This explicit declaration is desirable + # in limited VRAM environments because quantized models often initialize directly on the accelerator. + device_map = self.hf_device_map + is_device_type_map = False + if isinstance(device_map, str): + try: + torch.device(device_map) + is_device_type_map = True + except RuntimeError: + pass + + return not is_device_type_map and isinstance(device_map, dict) and len(device_map) > 1 + + +class StableDiffusionMixin: + r""" + Helper for DiffusionPipeline with vae and unet.(mainly for LDM such as stable diffusion) + """ + + def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): + r"""Enables the FreeU mechanism as in https://huggingface.co/papers/2309.11497. + + The suffixes after the scaling factors represent the stages where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values + that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if not hasattr(self, "unet"): + raise ValueError("The pipeline must have `unet` for using FreeU.") + self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) + + def disable_freeu(self): + """Disables the FreeU mechanism if enabled.""" + self.unet.disable_freeu() + + def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value) + are fused. For cross-attention modules, key and value projection matrices are fused. + + > [!WARNING] > This API is 🧪 experimental. + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + """ + self.fusing_unet = False + self.fusing_vae = False + + if unet: + self.fusing_unet = True + self.unet.fuse_qkv_projections() + self.unet.set_attn_processor(FusedAttnProcessor2_0()) + + if vae: + if not isinstance(self.vae, AutoencoderKL): + raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") + + self.fusing_vae = True + self.vae.fuse_qkv_projections() + self.vae.set_attn_processor(FusedAttnProcessor2_0()) + + def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """Disable QKV projection fusion if enabled. + + > [!WARNING] > This API is 🧪 experimental. + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + + """ + if unet: + if not self.fusing_unet: + logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") + else: + self.unet.unfuse_qkv_projections() + self.fusing_unet = False + + if vae: + if not self.fusing_vae: + logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") + else: + self.vae.unfuse_qkv_projections() + self.fusing_vae = False diff --git a/diffusers/pipelines/pixart_alpha/__init__.py b/diffusers/pipelines/pixart_alpha/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..7cb870fc8589aeff366429c182736b2ddd6ce215 --- /dev/null +++ b/diffusers/pipelines/pixart_alpha/__init__.py @@ -0,0 +1,55 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_pixart_alpha"] = ["PixArtAlphaPipeline"] + _import_structure["pipeline_pixart_sigma"] = ["PixArtSigmaPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_pixart_alpha import ( + ASPECT_RATIO_256_BIN, + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, + PixArtAlphaPipeline, + ) + from .pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN, PixArtSigmaPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py b/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py new file mode 100644 index 0000000000000000000000000000000000000000..b98248f3ca7b864445c7d1342e022df6e97b1e52 --- /dev/null +++ b/diffusers/pipelines/pixart_alpha/pipeline_pixart_alpha.py @@ -0,0 +1,984 @@ +# Copyright 2025 PixArt-Alpha Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +from typing import Callable + +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from ...image_processor import PixArtImageProcessor +from ...models import AutoencoderKL, PixArtTransformer2DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + deprecate, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import maybe_adjust_dtype_for_device, randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import PixArtAlphaPipeline + + >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-XL-2-512x512" too. + >>> pipe = PixArtAlphaPipeline.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", torch_dtype=torch.float16) + >>> # Enable memory optimizations. + >>> pipe.enable_model_cpu_offload() + + >>> prompt = "A small cactus with a happy face in the Sahara desert." + >>> image = pipe(prompt).images[0] + ``` +""" + +ASPECT_RATIO_1024_BIN = { + "0.25": [512.0, 2048.0], + "0.28": [512.0, 1856.0], + "0.32": [576.0, 1792.0], + "0.33": [576.0, 1728.0], + "0.35": [576.0, 1664.0], + "0.4": [640.0, 1600.0], + "0.42": [640.0, 1536.0], + "0.48": [704.0, 1472.0], + "0.5": [704.0, 1408.0], + "0.52": [704.0, 1344.0], + "0.57": [768.0, 1344.0], + "0.6": [768.0, 1280.0], + "0.68": [832.0, 1216.0], + "0.72": [832.0, 1152.0], + "0.78": [896.0, 1152.0], + "0.82": [896.0, 1088.0], + "0.88": [960.0, 1088.0], + "0.94": [960.0, 1024.0], + "1.0": [1024.0, 1024.0], + "1.07": [1024.0, 960.0], + "1.13": [1088.0, 960.0], + "1.21": [1088.0, 896.0], + "1.29": [1152.0, 896.0], + "1.38": [1152.0, 832.0], + "1.46": [1216.0, 832.0], + "1.67": [1280.0, 768.0], + "1.75": [1344.0, 768.0], + "2.0": [1408.0, 704.0], + "2.09": [1472.0, 704.0], + "2.4": [1536.0, 640.0], + "2.5": [1600.0, 640.0], + "3.0": [1728.0, 576.0], + "4.0": [2048.0, 512.0], +} + +ASPECT_RATIO_512_BIN = { + "0.25": [256.0, 1024.0], + "0.28": [256.0, 928.0], + "0.32": [288.0, 896.0], + "0.33": [288.0, 864.0], + "0.35": [288.0, 832.0], + "0.4": [320.0, 800.0], + "0.42": [320.0, 768.0], + "0.48": [352.0, 736.0], + "0.5": [352.0, 704.0], + "0.52": [352.0, 672.0], + "0.57": [384.0, 672.0], + "0.6": [384.0, 640.0], + "0.68": [416.0, 608.0], + "0.72": [416.0, 576.0], + "0.78": [448.0, 576.0], + "0.82": [448.0, 544.0], + "0.88": [480.0, 544.0], + "0.94": [480.0, 512.0], + "1.0": [512.0, 512.0], + "1.07": [512.0, 480.0], + "1.13": [544.0, 480.0], + "1.21": [544.0, 448.0], + "1.29": [576.0, 448.0], + "1.38": [576.0, 416.0], + "1.46": [608.0, 416.0], + "1.67": [640.0, 384.0], + "1.75": [672.0, 384.0], + "2.0": [704.0, 352.0], + "2.09": [736.0, 352.0], + "2.4": [768.0, 320.0], + "2.5": [800.0, 320.0], + "3.0": [864.0, 288.0], + "4.0": [1024.0, 256.0], +} + +ASPECT_RATIO_256_BIN = { + "0.25": [128.0, 512.0], + "0.28": [128.0, 464.0], + "0.32": [144.0, 448.0], + "0.33": [144.0, 432.0], + "0.35": [144.0, 416.0], + "0.4": [160.0, 400.0], + "0.42": [160.0, 384.0], + "0.48": [176.0, 368.0], + "0.5": [176.0, 352.0], + "0.52": [176.0, 336.0], + "0.57": [192.0, 336.0], + "0.6": [192.0, 320.0], + "0.68": [208.0, 304.0], + "0.72": [208.0, 288.0], + "0.78": [224.0, 288.0], + "0.82": [224.0, 272.0], + "0.88": [240.0, 272.0], + "0.94": [240.0, 256.0], + "1.0": [256.0, 256.0], + "1.07": [256.0, 240.0], + "1.13": [272.0, 240.0], + "1.21": [272.0, 224.0], + "1.29": [288.0, 224.0], + "1.38": [288.0, 208.0], + "1.46": [304.0, 208.0], + "1.67": [320.0, 192.0], + "1.75": [336.0, 192.0], + "2.0": [352.0, 176.0], + "2.09": [368.0, 176.0], + "2.4": [384.0, 160.0], + "2.5": [400.0, 160.0], + "3.0": [432.0, 144.0], + "4.0": [512.0, 128.0], +} + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class PixArtAlphaPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using PixArt-Alpha. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + Frozen text-encoder. PixArt-Alpha uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant. + tokenizer (`T5Tokenizer`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + transformer ([`PixArtTransformer2DModel`]): + A text conditioned `PixArtTransformer2DModel` to denoise the encoded image latents. Initially published as + [`Transformer2DModel`](https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS/blob/main/transformer/config.json#L2) + in the config, but the mismatch can be ignored. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + bad_punct_regex = re.compile( + r"[" + + "#®•©™&@·º½¾¿¡§~" + + r"\)" + + r"\(" + + r"\]" + + r"\[" + + r"\}" + + r"\{" + + r"\|" + + "\\" + + r"\/" + + r"\*" + + r"]{1,}" + ) # noqa + + _optional_components = ["tokenizer", "text_encoder"] + model_cpu_offload_seq = "text_encoder->transformer->vae" + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + vae: AutoencoderKL, + transformer: PixArtTransformer2DModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 120, + **kwargs, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Alpha, it's should be the embeddings of the "" + string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 120): Maximum sequence length to use for the prompt. + """ + + if "mask_feature" in kwargs: + deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version." + deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False) + + if device is None: + device = self._execution_device + + # See Section 3.1. of the paper. + max_length = max_sequence_length + + if prompt_embeds is None: + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because T5 can only handle sequences up to" + f" {max_length} tokens: {removed_text}" + ) + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + elif self.transformer is not None: + dtype = self.transformer.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.repeat(1, num_images_per_prompt) + prompt_attention_mask = prompt_attention_mask.view(bs_embed * num_images_per_prompt, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens = [negative_prompt] * bs_embed if isinstance(negative_prompt, str) else negative_prompt + uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption) + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + negative_prompt_attention_mask = uncond_input.attention_mask + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(1, num_images_per_prompt) + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed * num_images_per_prompt, -1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt, + callback_steps, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + height: int | None = None, + width: int | None = None, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + clean_caption: bool = True, + use_resolution_binning: bool = True, + max_sequence_length: int = 120, + **kwargs, + ) -> ImagePipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Alpha this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + max_sequence_length (`int` defaults to 120): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images + """ + if "mask_feature" in kwargs: + deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version." + deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False) + # 1. Check inputs. Raise error if not correct + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + if use_resolution_binning: + if self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_256_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + negative_prompt, + callback_steps, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + ) + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + prompt_attention_mask = prompt_attention_mask.to( + maybe_adjust_dtype_for_device(prompt_attention_mask.dtype, prompt_attention_mask.device) + ) + + # 4. Prepare timesteps + is_neuron_device = device.type == "neuron" + if XLA_AVAILABLE or is_neuron_device: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Prepare micro-conditions. + added_cond_kwargs = {"resolution": None, "aspect_ratio": None} + if self.transformer.config.sample_size == 128: + resolution = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1) + aspect_ratio = torch.tensor([float(height / width)]).repeat(batch_size * num_images_per_prompt, 1) + resolution = resolution.to(dtype=prompt_embeds.dtype, device=device) + aspect_ratio = aspect_ratio.to(dtype=prompt_embeds.dtype, device=device) + + if do_classifier_free_guidance: + resolution = torch.cat([resolution, resolution], dim=0) + aspect_ratio = torch.cat([aspect_ratio, aspect_ratio], dim=0) + + added_cond_kwargs = {"resolution": resolution, "aspect_ratio": aspect_ratio} + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + current_timestep = t + if not torch.is_tensor(current_timestep): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + if isinstance(current_timestep, float): + dtype = maybe_adjust_dtype_for_device(torch.float64, latent_model_input.device) + else: + dtype = maybe_adjust_dtype_for_device(torch.int64, latent_model_input.device) + current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device) + elif len(current_timestep.shape) == 0: + current_timestep = current_timestep[None].to(latent_model_input.device) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + current_timestep = current_timestep.expand(latent_model_input.shape[0]) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + timestep=current_timestep, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + else: + noise_pred = noise_pred + + # compute previous image: x_t -> x_t-1 + if num_inference_steps == 1: + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[1] + else: + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py b/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py new file mode 100644 index 0000000000000000000000000000000000000000..286695aa8eb98d49ef40f6dfc8349496fac7df67 --- /dev/null +++ b/diffusers/pipelines/pixart_alpha/pipeline_pixart_sigma.py @@ -0,0 +1,910 @@ +# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +from typing import Callable + +import torch +from transformers import T5EncoderModel, T5Tokenizer + +from ...image_processor import PixArtImageProcessor +from ...models import AutoencoderKL, PixArtTransformer2DModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + BACKENDS_MAPPING, + deprecate, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput +from .pipeline_pixart_alpha import ( + ASPECT_RATIO_256_BIN, + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, +) + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +ASPECT_RATIO_2048_BIN = { + "0.25": [1024.0, 4096.0], + "0.26": [1024.0, 3968.0], + "0.27": [1024.0, 3840.0], + "0.28": [1024.0, 3712.0], + "0.32": [1152.0, 3584.0], + "0.33": [1152.0, 3456.0], + "0.35": [1152.0, 3328.0], + "0.4": [1280.0, 3200.0], + "0.42": [1280.0, 3072.0], + "0.48": [1408.0, 2944.0], + "0.5": [1408.0, 2816.0], + "0.52": [1408.0, 2688.0], + "0.57": [1536.0, 2688.0], + "0.6": [1536.0, 2560.0], + "0.68": [1664.0, 2432.0], + "0.72": [1664.0, 2304.0], + "0.78": [1792.0, 2304.0], + "0.82": [1792.0, 2176.0], + "0.88": [1920.0, 2176.0], + "0.94": [1920.0, 2048.0], + "1.0": [2048.0, 2048.0], + "1.07": [2048.0, 1920.0], + "1.13": [2176.0, 1920.0], + "1.21": [2176.0, 1792.0], + "1.29": [2304.0, 1792.0], + "1.38": [2304.0, 1664.0], + "1.46": [2432.0, 1664.0], + "1.67": [2560.0, 1536.0], + "1.75": [2688.0, 1536.0], + "2.0": [2816.0, 1408.0], + "2.09": [2944.0, 1408.0], + "2.4": [3072.0, 1280.0], + "2.5": [3200.0, 1280.0], + "2.89": [3328.0, 1152.0], + "3.0": [3456.0, 1152.0], + "3.11": [3584.0, 1152.0], + "3.62": [3712.0, 1024.0], + "3.75": [3840.0, 1024.0], + "3.88": [3968.0, 1024.0], + "4.0": [4096.0, 1024.0], +} + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import PixArtSigmaPipeline + + >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-Sigma-XL-2-512-MS" too. + >>> pipe = PixArtSigmaPipeline.from_pretrained( + ... "PixArt-alpha/PixArt-Sigma-XL-2-1024-MS", torch_dtype=torch.float16 + ... ) + >>> # Enable memory optimizations. + >>> # pipe.enable_model_cpu_offload() + + >>> prompt = "A small cactus with a happy face in the Sahara desert." + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class PixArtSigmaPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation using PixArt-Sigma. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`T5EncoderModel`]): + Frozen text-encoder. PixArt-Alpha uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant. + tokenizer (`T5Tokenizer`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + transformer ([`PixArtTransformer2DModel`]): + A text conditioned `PixArtTransformer2DModel` to denoise the encoded image latents. Initially published as + [`Transformer2DModel`](https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/blob/main/transformer/config.json#L2) + in the config, but the mismatch can be ignored. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + """ + + bad_punct_regex = re.compile( + r"[" + + "#®•©™&@·º½¾¿¡§~" + + r"\)" + + r"\(" + + r"\]" + + r"\[" + + r"\}" + + r"\{" + + r"\|" + + "\\" + + r"\/" + + r"\*" + + r"]{1,}" + ) # noqa + + _optional_components = ["tokenizer", "text_encoder"] + model_cpu_offload_seq = "text_encoder->transformer->vae" + + def __init__( + self, + tokenizer: T5Tokenizer, + text_encoder: T5EncoderModel, + vae: AutoencoderKL, + transformer: PixArtTransformer2DModel, + scheduler: KarrasDiffusionSchedulers, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.PixArtAlphaPipeline.encode_prompt with 120->300 + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + **kwargs, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Alpha, it's should be the embeddings of the "" + string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + """ + + if "mask_feature" in kwargs: + deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version." + deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False) + + if device is None: + device = self._execution_device + + # See Section 3.1. of the paper. + max_length = max_sequence_length + + if prompt_embeds is None: + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because T5 can only handle sequences up to" + f" {max_length} tokens: {removed_text}" + ) + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + elif self.transformer is not None: + dtype = self.transformer.dtype + else: + dtype = None + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.repeat(1, num_images_per_prompt) + prompt_attention_mask = prompt_attention_mask.view(bs_embed * num_images_per_prompt, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens = [negative_prompt] * bs_embed if isinstance(negative_prompt, str) else negative_prompt + uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption) + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + add_special_tokens=True, + return_tensors="pt", + ) + negative_prompt_attention_mask = uncond_input.attention_mask + negative_prompt_attention_mask = negative_prompt_attention_mask.to(device) + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(1, num_images_per_prompt) + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed * num_images_per_prompt, -1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.pixart_alpha.pipeline_pixart_alpha.PixArtAlphaPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + negative_prompt, + callback_steps, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + height: int | None = None, + width: int | None = None, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + clean_caption: bool = True, + use_resolution_binning: bool = True, + max_sequence_length: int = 300, + **kwargs, + ) -> ImagePipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + max_sequence_length (`int` defaults to 300): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is + returned where the first element is a list with the generated images + """ + # 1. Check inputs. Raise error if not correct + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + if use_resolution_binning: + if self.transformer.config.sample_size == 256: + aspect_ratio_bin = ASPECT_RATIO_2048_BIN + elif self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_256_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + negative_prompt, + callback_steps, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + ) + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Prepare micro-conditions. + added_cond_kwargs = {"resolution": None, "aspect_ratio": None} + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + current_timestep = t + if not torch.is_tensor(current_timestep): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = latent_model_input.device.type == "mps" + is_npu = latent_model_input.device.type == "npu" + if isinstance(current_timestep, float): + dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + else: + dtype = torch.int32 if (is_mps or is_npu) else torch.int64 + current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device) + elif len(current_timestep.shape) == 0: + current_timestep = current_timestep[None].to(latent_model_input.device) + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + current_timestep = current_timestep.expand(latent_model_input.shape[0]) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input, + encoder_hidden_states=prompt_embeds, + encoder_attention_mask=prompt_attention_mask, + timestep=current_timestep, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + else: + noise_pred = noise_pred + + # compute previous image: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents.to(self.vae.dtype) / self.vae.config.scaling_factor, return_dict=False)[0] + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + else: + image = latents + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/prx/__init__.py b/diffusers/pipelines/prx/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a34f55b32335831db9c0d3541a74cf4a2cc4cbf8 --- /dev/null +++ b/diffusers/pipelines/prx/__init__.py @@ -0,0 +1,76 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["PRXPipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_prx"] = ["PRXPipeline"] + _import_structure["pipeline_prx_pixel"] = ["PRXPixelPipeline"] + +# Wrap T5GemmaEncoder to pass config.encoder (T5GemmaModuleConfig) instead of the +# composite T5GemmaConfig, which lacks flat attributes expected by T5GemmaEncoder.__init__. +try: + if is_transformers_available(): + import transformers + from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder as _T5GemmaEncoder + + class T5GemmaEncoder(_T5GemmaEncoder): + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs): + if "config" not in kwargs: + from transformers.models.t5gemma.configuration_t5gemma import T5GemmaConfig + + config = T5GemmaConfig.from_pretrained(pretrained_model_name_or_path) + if hasattr(config, "encoder"): + kwargs["config"] = config.encoder + return super().from_pretrained(pretrained_model_name_or_path, *args, **kwargs) + + _additional_imports["T5GemmaEncoder"] = T5GemmaEncoder + if not hasattr(transformers, "T5GemmaEncoder"): + transformers.T5GemmaEncoder = T5GemmaEncoder +except ImportError: + pass + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_output import PRXPipelineOutput + from .pipeline_prx import PRXPipeline + from .pipeline_prx_pixel import PRXPixelPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/prx/pipeline_output.py b/diffusers/pipelines/prx/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..81f92c294735136779753592cf0b81de111abfff --- /dev/null +++ b/diffusers/pipelines/prx/pipeline_output.py @@ -0,0 +1,34 @@ +# Copyright 2025 The Photoroom and the HuggingFace Teams. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class PRXPipelineOutput(BaseOutput): + """ + Output class for PRX pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/prx/pipeline_prx.py b/diffusers/pipelines/prx/pipeline_prx.py new file mode 100644 index 0000000000000000000000000000000000000000..f4ec214313e3c9df29774ca04d155372ccc85353 --- /dev/null +++ b/diffusers/pipelines/prx/pipeline_prx.py @@ -0,0 +1,835 @@ +# Copyright 2025 The Photoroom and The HuggingFace Teams. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +from typing import Callable + +import torch +from transformers import ( + AutoTokenizer, + GemmaTokenizerFast, + T5TokenizerFast, +) +from transformers.models.t5gemma.modeling_t5gemma import T5GemmaEncoder + +from diffusers.image_processor import PixArtImageProcessor +from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderDC, AutoencoderKL +from diffusers.models.transformers.transformer_prx import PRXTransformer2DModel +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.prx.pipeline_output import PRXPipelineOutput +from diffusers.schedulers import FlowMatchEulerDiscreteScheduler +from diffusers.utils import is_ftfy_available, logging, replace_example_docstring +from diffusers.utils.torch_utils import randn_tensor + + +if is_ftfy_available(): + import ftfy + +DEFAULT_RESOLUTION = 512 + +ASPECT_RATIO_256_BIN = { + "0.46": [160, 352], + "0.6": [192, 320], + "0.78": [224, 288], + "1.0": [256, 256], + "1.29": [288, 224], + "1.67": [320, 192], + "2.2": [352, 160], +} + +ASPECT_RATIO_512_BIN = { + "0.5": [352, 704], + "0.57": [384, 672], + "0.6": [384, 640], + "0.68": [416, 608], + "0.78": [448, 576], + "0.88": [480, 544], + "1.0": [512, 512], + "1.13": [544, 480], + "1.29": [576, 448], + "1.46": [608, 416], + "1.67": [640, 384], + "1.75": [672, 384], + "2.0": [704, 352], +} + +ASPECT_RATIO_1024_BIN = { + "0.49": [704, 1440], + "0.52": [736, 1408], + "0.53": [736, 1376], + "0.57": [768, 1344], + "0.59": [768, 1312], + "0.62": [800, 1280], + "0.67": [832, 1248], + "0.68": [832, 1216], + "0.78": [896, 1152], + "0.83": [928, 1120], + "0.94": [992, 1056], + "1.0": [1024, 1024], + "1.06": [1056, 992], + "1.13": [1088, 960], + "1.21": [1120, 928], + "1.29": [1152, 896], + "1.37": [1184, 864], + "1.46": [1216, 832], + "1.5": [1248, 832], + "1.71": [1312, 768], + "1.75": [1344, 768], + "1.87": [1376, 736], + "1.91": [1408, 736], + "2.05": [1440, 704], +} + +ASPECT_RATIO_BINS = { + 256: ASPECT_RATIO_256_BIN, + 512: ASPECT_RATIO_512_BIN, + 1024: ASPECT_RATIO_1024_BIN, +} + +logger = logging.get_logger(__name__) + + +class TextPreprocessor: + """Text preprocessing utility for PRXPipeline.""" + + def __init__(self): + """Initialize text preprocessor.""" + self.bad_punct_regex = re.compile( + r"[" + + "#®•©™&@·º½¾¿¡§~" + + r"\)" + + r"\(" + + r"\]" + + r"\[" + + r"\}" + + r"\{" + + r"\|" + + r"\\" + + r"\/" + + r"\*" + + r"]{1,}" + ) + + def clean_text(self, text: str) -> str: + """Clean text using comprehensive text processing logic.""" + # See Deepfloyd https://github.com/deep-floyd/IF/blob/develop/deepfloyd_if/modules/t5.py + text = str(text) + text = ul.unquote_plus(text) + text = text.strip().lower() + text = re.sub("", "person", text) + + # Remove all urls: + text = re.sub( + r"\b((?:https?|www):(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@))", + "", + text, + ) # regex for urls + + # @ + text = re.sub(r"@[\w\d]+\b", "", text) + + # 31C0—31EF CJK Strokes through 4E00—9FFF CJK Unified Ideographs + text = re.sub(r"[\u31c0-\u31ef]+", "", text) + text = re.sub(r"[\u31f0-\u31ff]+", "", text) + text = re.sub(r"[\u3200-\u32ff]+", "", text) + text = re.sub(r"[\u3300-\u33ff]+", "", text) + text = re.sub(r"[\u3400-\u4dbf]+", "", text) + text = re.sub(r"[\u4dc0-\u4dff]+", "", text) + text = re.sub(r"[\u4e00-\u9fff]+", "", text) + + # все виды тире / all types of dash --> "-" + text = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", + "-", + text, + ) + + # кавычки к одному стандарту + text = re.sub(r"[`´«»" "¨]", '"', text) + text = re.sub(r"['']", "'", text) + + # " and & + text = re.sub(r""?", "", text) + text = re.sub(r"&", "", text) + + # ip addresses: + text = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", text) + + # article ids: + text = re.sub(r"\d:\d\d\s+$", "", text) + + # \n + text = re.sub(r"\\n", " ", text) + + # "#123", "#12345..", "123456.." + text = re.sub(r"#\d{1,3}\b", "", text) + text = re.sub(r"#\d{5,}\b", "", text) + text = re.sub(r"\b\d{6,}\b", "", text) + + # filenames: + text = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", text) + + # Clean punctuation + text = re.sub(r"[\"\']{2,}", r'"', text) # """AUSVERKAUFT""" + text = re.sub(r"[\.]{2,}", r" ", text) + + text = re.sub(self.bad_punct_regex, r" ", text) # ***AUSVERKAUFT***, #AUSVERKAUFT + text = re.sub(r"\s+\.\s+", r" ", text) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, text)) > 3: + text = re.sub(regex2, " ", text) + + # Basic cleaning + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + text = text.strip() + + # Clean alphanumeric patterns + text = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", text) # jc6640 + text = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", text) # jc6640vc + text = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", text) # 6640vc231 + + # Common spam patterns + text = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", text) + text = re.sub(r"(free\s)?download(\sfree)?", "", text) + text = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", text) + text = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", text) + text = re.sub(r"\bpage\s+\d+\b", "", text) + + text = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", text) # j2d1a2a... + text = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", text) + + # Final cleanup + text = re.sub(r"\b\s+\:\s+", r": ", text) + text = re.sub(r"(\D[,\./])\b", r"\1 ", text) + text = re.sub(r"\s+", " ", text) + + text.strip() + + text = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", text) + text = re.sub(r"^[\'\_,\-\:;]", r"", text) + text = re.sub(r"[\'\_,\-\:\-\+]$", r"", text) + text = re.sub(r"^\.\S+$", "", text) + + return text.strip() + + def basic_clean(self, text: str) -> str: + """Light cleaning: fix mojibake and unescape HTML. Used when skip_text_cleaning=True.""" + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import PRXPipeline + + >>> # Load pipeline with from_pretrained + >>> pipe = PRXPipeline.from_pretrained("Photoroom/prx-512-t2i-sft") + >>> pipe.to("cuda") + + >>> prompt = "A digital painting of a rusty, vintage tram on a sandy beach" + >>> image = pipe(prompt, num_inference_steps=28, guidance_scale=5.0).images[0] + >>> image.save("prx_output.png") + ``` +""" + + +class PRXPipeline( + DiffusionPipeline, + LoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + Pipeline for text-to-image generation using PRX Transformer. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + transformer ([`PRXTransformer2DModel`]): + The PRX transformer model to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + text_encoder ([`T5GemmaEncoder`]): + Text encoder model for encoding prompts. + tokenizer ([`T5TokenizerFast` or `GemmaTokenizerFast`]): + Tokenizer for the text encoder. + vae ([`AutoencoderKL`] or [`AutoencoderDC`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + Supports both AutoencoderKL (8x compression) and AutoencoderDC (32x compression). + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + _optional_components = ["vae"] + + def __init__( + self, + transformer: PRXTransformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder: T5GemmaEncoder, + tokenizer: T5TokenizerFast | GemmaTokenizerFast | AutoTokenizer, + vae: AutoencoderKL | AutoencoderDC | None = None, + default_sample_size: int | None = DEFAULT_RESOLUTION, + ): + super().__init__() + + if PRXTransformer2DModel is None: + raise ImportError( + "PRXTransformer2DModel is not available. Please ensure the transformer_prx module is properly installed." + ) + + self.text_preprocessor = TextPreprocessor() + self.default_sample_size = default_sample_size + self._guidance_scale = 1.0 + + self.register_modules( + transformer=transformer, + scheduler=scheduler, + text_encoder=text_encoder, + tokenizer=tokenizer, + vae=vae, + ) + + self.register_to_config(default_sample_size=self.default_sample_size) + + if vae is not None: + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + else: + self.image_processor = None + + @property + def vae_scale_factor(self): + if self.vae is None: + return 8 + if hasattr(self.vae, "spatial_compression_ratio"): + return self.vae.spatial_compression_ratio + else: # Flux VAE + return 2 ** (len(self.vae.config.block_out_channels) - 1) + + @property + def do_classifier_free_guidance(self): + """Check if classifier-free guidance is enabled based on guidance scale.""" + return self._guidance_scale > 1.0 + + @property + def guidance_scale(self): + return self._guidance_scale + + def get_default_resolution(self): + """Determine the default resolution based on the loaded VAE and config. + + Returns: + int: The default sample size (height/width) to use for generation. + """ + default_from_config = getattr(self.config, "default_sample_size", None) + if default_from_config is not None: + return default_from_config + + return DEFAULT_RESOLUTION + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ): + """Prepare initial latents for the diffusion process.""" + if latents is None: + spatial_compression = self.vae_scale_factor + latent_height, latent_width = ( + height // spatial_compression, + width // spatial_compression, + ) + shape = (batch_size, num_channels_latents, latent_height, latent_width) + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + return latents + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + prompt_attention_mask: torch.BoolTensor | None = None, + negative_prompt_attention_mask: torch.BoolTensor | None = None, + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ): + """Encode text prompt using standard text encoder and tokenizer, or use precomputed embeddings.""" + if device is None: + device = self._execution_device + + if prompt_embeds is None: + if isinstance(prompt, str): + prompt = [prompt] + # Encode the prompts + prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask = ( + self._encode_prompt_standard( + prompt, + device, + do_classifier_free_guidance, + negative_prompt, + tokenizer_max_length=tokenizer_max_length, + skip_text_cleaning=skip_text_cleaning, + ) + ) + + # Duplicate embeddings for each generation per prompt + if num_images_per_prompt > 1: + # Repeat prompt embeddings + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if prompt_attention_mask is not None: + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + # Repeat negative embeddings if using CFG + if do_classifier_free_guidance and negative_prompt_embeds is not None: + bs_embed, seq_len, _ = negative_prompt_embeds.shape + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if negative_prompt_attention_mask is not None: + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + return ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds if do_classifier_free_guidance else None, + negative_prompt_attention_mask if do_classifier_free_guidance else None, + ) + + def _tokenize_prompts( + self, + prompts: list[str], + device: torch.device, + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ): + """Tokenize and clean prompts.""" + clean_fn = self.text_preprocessor.basic_clean if skip_text_cleaning else self.text_preprocessor.clean_text + cleaned = [clean_fn(text) for text in prompts] + max_length = tokenizer_max_length or self.tokenizer.model_max_length + tokens = self.tokenizer( + cleaned, + padding="max_length", + max_length=max_length, + truncation=True, + return_attention_mask=True, + return_tensors="pt", + ) + return tokens["input_ids"].to(device), tokens["attention_mask"].bool().to(device) + + def _encode_prompt_standard( + self, + prompt: list[str], + device: torch.device, + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ): + """Encode prompt using standard text encoder and tokenizer with batch processing.""" + batch_size = len(prompt) + + if do_classifier_free_guidance: + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + + prompts_to_encode = negative_prompt + prompt + else: + prompts_to_encode = prompt + + input_ids, attention_mask = self._tokenize_prompts( + prompts_to_encode, device, tokenizer_max_length=tokenizer_max_length, skip_text_cleaning=skip_text_cleaning + ) + + with torch.no_grad(): + embeddings = self.text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + )["last_hidden_state"] + + if do_classifier_free_guidance: + uncond_text_embeddings, text_embeddings = embeddings.split(batch_size, dim=0) + uncond_cross_attn_mask, cross_attn_mask = attention_mask.split(batch_size, dim=0) + else: + text_embeddings = embeddings + cross_attn_mask = attention_mask + uncond_text_embeddings = None + uncond_cross_attn_mask = None + + return text_embeddings, cross_attn_mask, uncond_text_embeddings, uncond_cross_attn_mask + + def check_inputs( + self, + prompt: str | list[str], + height: int, + width: int, + guidance_scale: float, + callback_on_step_end_tensor_inputs: list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + ): + """Check that all inputs are in correct format.""" + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + + if prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and guidance_scale > 1.0 and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided and `guidance_scale > 1.0`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + + # The latents must be divisible by the transformer's patch size after VAE compression. + dimension_multiple = self.vae_scale_factor * self.transformer.config.patch_size + if height % dimension_multiple != 0 or width % dimension_multiple != 0: + raise ValueError( + f"`height` and `width` have to be divisible by {dimension_multiple} (vae_scale_factor *" + f" transformer patch_size) but are {height} and {width}." + ) + + if guidance_scale < 1.0: + raise ValueError(f"guidance_scale has to be >= 1.0 but is {guidance_scale}") + + if callback_on_step_end_tensor_inputs is not None and not isinstance(callback_on_step_end_tensor_inputs, list): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be a list but is {callback_on_step_end_tensor_inputs}" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + timesteps: list[int] = None, + guidance_scale: float = 4.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + prompt_attention_mask: torch.BoolTensor | None = None, + negative_prompt_attention_mask: torch.BoolTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + use_resolution_binning: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str`, *optional*, defaults to `""`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 28): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided and `guidance_scale > 1`, negative embeddings will be generated from an + empty string. + prompt_attention_mask (`torch.BoolTensor`, *optional*): + Pre-generated attention mask for `prompt_embeds`. If not provided, attention mask will be generated + from `prompt` input argument. + negative_prompt_attention_mask (`torch.BoolTensor`, *optional*): + Pre-generated attention mask for `negative_prompt_embeds`. If not provided and `guidance_scale > 1`, + attention mask will be generated from an empty string. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + tokenizer_max_length (`int`, *optional*): + Override the maximum number of tokens used when tokenizing the prompt. Defaults to the tokenizer's own + ``model_max_length`` when not set. + skip_text_cleaning (`bool`, *optional*, defaults to `False`): + If `True`, uses only light prompt cleaning (fix encoding + unescape HTML) instead of the full DeepFloyd + cleaning pipeline. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.prx.PRXPipelineOutput`] instead of a plain tuple. + use_resolution_binning (`bool`, *optional*, defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + predefined aspect ratio bins. After the produced latents are decoded into images, they are resized back + to the requested resolution. Useful for generating non-square images at optimal resolutions. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self, step, timestep, callback_kwargs)`. + `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include tensors that are listed + in the `._callback_tensor_inputs` attribute. + + Examples: + + Returns: + [`~pipelines.prx.PRXPipelineOutput`] or `tuple`: [`~pipelines.prx.PRXPipelineOutput`] if `return_dict` is + True, otherwise a `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + # 0. Set height and width + default_resolution = self.get_default_resolution() + height = height or default_resolution + width = width or default_resolution + + if use_resolution_binning: + if self.default_sample_size not in ASPECT_RATIO_BINS: + raise ValueError( + f"Resolution binning is only supported for default_sample_size in {list(ASPECT_RATIO_BINS.keys())}, " + f"but got {self.default_sample_size}. Set use_resolution_binning=False to disable aspect ratio binning." + ) + aspect_ratio_bin = ASPECT_RATIO_BINS[self.default_sample_size] + + # Store original dimensions + orig_height, orig_width = height, width + # Map to closest resolution in the bin + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + # 1. Check inputs + self.check_inputs( + prompt, + height, + width, + guidance_scale, + callback_on_step_end_tensor_inputs, + prompt_embeds, + negative_prompt_embeds, + ) + + if self.vae is None and output_type not in ["latent", "pt"]: + raise ValueError( + f"VAE is required for output_type='{output_type}' but it is not available. " + "Either provide a VAE or set output_type='latent' or 'pt' to get latent outputs." + ) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Use execution device (handles offloading scenarios including group offloading) + device = self._execution_device + + self._guidance_scale = guidance_scale + + # 2. Encode input prompt + text_embeddings, cross_attn_mask, uncond_text_embeddings, uncond_cross_attn_mask = self.encode_prompt( + prompt, + device, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + tokenizer_max_length=tokenizer_max_length, + skip_text_cleaning=skip_text_cleaning, + ) + # Expose standard names for callbacks parity + prompt_embeds = text_embeddings + negative_prompt_embeds = uncond_text_embeddings + + # 3. Prepare timesteps + if timesteps is not None: + self.scheduler.set_timesteps(timesteps=timesteps, device=device) + timesteps = self.scheduler.timesteps + num_inference_steps = len(timesteps) + else: + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + self.num_timesteps = len(timesteps) + + # 4. Prepare latent variables + if self.vae is not None: + num_channels_latents = self.vae.config.latent_channels + else: + # When vae is None, get latent channels from transformer + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + text_embeddings.dtype, + device, + generator, + latents, + ) + + # 5. Prepare extra step kwargs + extra_step_kwargs = {} + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_eta: + extra_step_kwargs["eta"] = 0.0 + + # 6. Prepare cross-attention embeddings and masks + if self.do_classifier_free_guidance: + ca_embed = torch.cat([uncond_text_embeddings, text_embeddings], dim=0) + ca_mask = None + if cross_attn_mask is not None and uncond_cross_attn_mask is not None: + ca_mask = torch.cat([uncond_cross_attn_mask, cross_attn_mask], dim=0) + else: + ca_embed = text_embeddings + ca_mask = cross_attn_mask + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Duplicate latents if using classifier-free guidance + if self.do_classifier_free_guidance: + latents_in = torch.cat([latents, latents], dim=0) + # Normalize timestep for the transformer + t_cont = (t.float() / self.scheduler.config.num_train_timesteps).view(1).repeat(2).to(device) + else: + latents_in = latents + # Normalize timestep for the transformer + t_cont = (t.float() / self.scheduler.config.num_train_timesteps).view(1).to(device) + + # Forward through transformer + noise_pred = self.transformer( + hidden_states=latents_in, + timestep=t_cont, + encoder_hidden_states=ca_embed, + attention_mask=ca_mask, + return_dict=False, + )[0] + + # Apply CFG + if self.do_classifier_free_guidance: + noise_uncond, noise_text = noise_pred.chunk(2, dim=0) + noise_pred = noise_uncond + guidance_scale * (noise_text - noise_uncond) + + # Compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_on_step_end(self, i, t, callback_kwargs) + + # Call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + # 8. Post-processing + if output_type == "latent" or (output_type == "pt" and self.vae is None): + image = latents + else: + # Unscale latents for VAE (supports both AutoencoderKL and AutoencoderDC) + scaling_factor = getattr(self.vae.config, "scaling_factor", 0.18215) + shift_factor = getattr(self.vae.config, "shift_factor", 0.0) + latents = (latents / scaling_factor) + shift_factor + # Decode using VAE (AutoencoderKL or AutoencoderDC) + image = self.vae.decode(latents, return_dict=False)[0] + # Resize back to original resolution if using binning + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + # Use standard image processor for post-processing + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return PRXPipelineOutput(images=image) diff --git a/diffusers/pipelines/prx/pipeline_prx_pixel.py b/diffusers/pipelines/prx/pipeline_prx_pixel.py new file mode 100644 index 0000000000000000000000000000000000000000..22a4d8dd4b187f6c788668886e355e8c81fbff9e --- /dev/null +++ b/diffusers/pipelines/prx/pipeline_prx_pixel.py @@ -0,0 +1,666 @@ +# Copyright 2025 The Photoroom and The HuggingFace Teams. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +from typing import Callable + +import torch +from transformers import AutoTokenizer, PreTrainedModel, PreTrainedTokenizerBase + +from ...image_processor import PixArtImageProcessor +from ...models.transformers.transformer_prx import PRXTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import PRXPipelineOutput +from .pipeline_prx import TextPreprocessor + + +if is_ftfy_available(): + import ftfy + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# PRXPixel is a 1024px model. +PRX_PIXEL_DEFAULT_RESOLUTION = 1024 +# Number of text tokens used at training time (the Qwen tokenizer's own ``model_max_length`` is far larger). +PRX_PIXEL_DEFAULT_MAX_TOKENS = 256 + +# Predefined aspect-ratio bins for 1024px generation (mirrors ASPECT_RATIO_1024_BIN in pipeline_prx). +ASPECT_RATIO_1024_BIN = { + "0.49": [704, 1440], + "0.52": [736, 1408], + "0.53": [736, 1376], + "0.57": [768, 1344], + "0.59": [768, 1312], + "0.62": [800, 1280], + "0.67": [832, 1248], + "0.68": [832, 1216], + "0.78": [896, 1152], + "0.83": [928, 1120], + "0.94": [992, 1056], + "1.0": [1024, 1024], + "1.06": [1056, 992], + "1.13": [1088, 960], + "1.21": [1120, 928], + "1.29": [1152, 896], + "1.37": [1184, 864], + "1.46": [1216, 832], + "1.5": [1248, 832], + "1.71": [1312, 768], + "1.75": [1344, 768], + "1.87": [1376, 736], + "1.91": [1408, 736], + "2.05": [1440, 704], +} + +ASPECT_RATIO_BINS = { + 1024: ASPECT_RATIO_1024_BIN, +} + + +def _basic_clean(text: str) -> str: + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import PRXPixelPipeline + + >>> pipe = PRXPixelPipeline.from_pretrained("Photoroom/prxpixel-t2i", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A front-facing portrait of a lion in the golden savanna at sunset." + >>> image = pipe(prompt, num_inference_steps=28, guidance_scale=5.0).images[0] + >>> image.save("prxpixel_output.png") + ``` +""" + + +class PRXPixelPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-image generation with the PRXPixel model. + + PRXPixel is a standalone, pixel-space text-to-image pipeline. It denoises raw RGB directly with a ~7B-parameter + [`PRXTransformer2DModel`] and has no VAE (generation happens entirely in pixel space, so the denoised output *is* + the image). Prompts are encoded with a Qwen3-VL text encoder (the vision tower is discarded). Unlike + [`PRXPipeline`] the transformer is trained with x-prediction: at every step it predicts the clean image `x0`, which + is converted to a flow-matching velocity before the scheduler step. Sampling starts from `randn * noise_scale` + (`noise_scale=2.0` by default) and the default resolution is 1024px. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Examples: + ```py + >>> import torch + >>> from diffusers import PRXPixelPipeline + + >>> pipe = PRXPixelPipeline.from_pretrained("Photoroom/prxpixel-t2i", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> prompt = "A front-facing portrait of a lion in the golden savanna at sunset." + >>> image = pipe(prompt, num_inference_steps=28, guidance_scale=5.0).images[0] + >>> image.save("prxpixel_output.png") + ``` + + Args: + transformer ([`PRXTransformer2DModel`]): + The ~7B-parameter PRX denoiser. For PRXPixel this is built with `in_channels=3`, a bottleneck `img_in`, and + `resolution_embeds=True`, and it is trained to predict the clean image `x0`. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + Flow-matching scheduler used to denoise the (pixel-space) latents. + text_encoder ([`PreTrainedModel`]): + The Qwen3-VL text backbone used to encode prompts (the vision tower is discarded). Must return a + `last_hidden_state`. + tokenizer ([`PreTrainedTokenizerBase`]): + Tokenizer for `text_encoder` (typically loaded via `AutoTokenizer`). + default_sample_size (`int`, *optional*, defaults to 1024): + Default height/width used when none is provided to `__call__`. + prompt_max_tokens (`int`, *optional*, defaults to 256): + Number of text tokens the prompt is padded/truncated to before encoding. + noise_scale (`float`, *optional*, defaults to 2.0): + Scale applied to the initial Gaussian noise. PRXPixel trains with a non-unit initial-noise scale, so + sampling must start from `randn * noise_scale`. + """ + + model_cpu_offload_seq = "text_encoder->transformer" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + transformer: PRXTransformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer | PreTrainedTokenizerBase, + default_sample_size: int | None = PRX_PIXEL_DEFAULT_RESOLUTION, + prompt_max_tokens: int = PRX_PIXEL_DEFAULT_MAX_TOKENS, + noise_scale: float = 2.0, + ): + super().__init__() + + self.text_preprocessor = TextPreprocessor() + self._guidance_scale = 1.0 + + self.register_modules( + transformer=transformer, + scheduler=scheduler, + text_encoder=text_encoder, + tokenizer=tokenizer, + ) + self.register_to_config( + default_sample_size=default_sample_size, + prompt_max_tokens=prompt_max_tokens, + noise_scale=noise_scale, + ) + + # Pixel pipeline always has an image_processor (vae_scale_factor=1) + # so that output_type="pil"/"np" work without a VAE. + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + @property + def vae_scale_factor(self): + # PRXPixel operates directly in RGB pixel space: no VAE, no spatial compression. + return 1 + + @property + # Copied from diffusers.pipelines.prx.pipeline_prx.PRXPipeline.do_classifier_free_guidance + def do_classifier_free_guidance(self): + """Check if classifier-free guidance is enabled based on guidance scale.""" + return self._guidance_scale > 1.0 + + @property + # Copied from diffusers.pipelines.prx.pipeline_prx.PRXPipeline.guidance_scale + def guidance_scale(self): + return self._guidance_scale + + def _tokenize_prompts( + self, + prompts: list[str], + device: torch.device, + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ) -> tuple[torch.Tensor, torch.Tensor]: + """Tokenize and (lightly) clean prompts. + + PRXPixel always uses light cleaning (`_basic_clean`) and the training-time token budget + (`self.config.prompt_max_tokens`). The `tokenizer_max_length` and `skip_text_cleaning` arguments are accepted + for API compatibility with the copied callers but are ignored. + """ + cleaned = [_basic_clean(text) for text in prompts] + tokens = self.tokenizer( + cleaned, + padding="max_length", + max_length=self.config.prompt_max_tokens, + truncation=True, + return_attention_mask=True, + return_tensors="pt", + ) + return tokens["input_ids"].to(device), tokens["attention_mask"].bool().to(device) + + # Copied from diffusers.pipelines.prx.pipeline_prx.PRXPipeline._encode_prompt_standard + def _encode_prompt_standard( + self, + prompt: list[str], + device: torch.device, + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ): + """Encode prompt using standard text encoder and tokenizer with batch processing.""" + batch_size = len(prompt) + + if do_classifier_free_guidance: + if isinstance(negative_prompt, str): + negative_prompt = [negative_prompt] * batch_size + + prompts_to_encode = negative_prompt + prompt + else: + prompts_to_encode = prompt + + input_ids, attention_mask = self._tokenize_prompts( + prompts_to_encode, device, tokenizer_max_length=tokenizer_max_length, skip_text_cleaning=skip_text_cleaning + ) + + with torch.no_grad(): + embeddings = self.text_encoder( + input_ids=input_ids, + attention_mask=attention_mask, + output_hidden_states=True, + )["last_hidden_state"] + + if do_classifier_free_guidance: + uncond_text_embeddings, text_embeddings = embeddings.split(batch_size, dim=0) + uncond_cross_attn_mask, cross_attn_mask = attention_mask.split(batch_size, dim=0) + else: + text_embeddings = embeddings + cross_attn_mask = attention_mask + uncond_text_embeddings = None + uncond_cross_attn_mask = None + + return text_embeddings, cross_attn_mask, uncond_text_embeddings, uncond_cross_attn_mask + + # Copied from diffusers.pipelines.prx.pipeline_prx.PRXPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + prompt_attention_mask: torch.BoolTensor | None = None, + negative_prompt_attention_mask: torch.BoolTensor | None = None, + tokenizer_max_length: int | None = None, + skip_text_cleaning: bool = False, + ): + """Encode text prompt using standard text encoder and tokenizer, or use precomputed embeddings.""" + if device is None: + device = self._execution_device + + if prompt_embeds is None: + if isinstance(prompt, str): + prompt = [prompt] + # Encode the prompts + prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask = ( + self._encode_prompt_standard( + prompt, + device, + do_classifier_free_guidance, + negative_prompt, + tokenizer_max_length=tokenizer_max_length, + skip_text_cleaning=skip_text_cleaning, + ) + ) + + # Duplicate embeddings for each generation per prompt + if num_images_per_prompt > 1: + # Repeat prompt embeddings + bs_embed, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if prompt_attention_mask is not None: + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + # Repeat negative embeddings if using CFG + if do_classifier_free_guidance and negative_prompt_embeds is not None: + bs_embed, seq_len, _ = negative_prompt_embeds.shape + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if negative_prompt_attention_mask is not None: + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + + return ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds if do_classifier_free_guidance else None, + negative_prompt_attention_mask if do_classifier_free_guidance else None, + ) + + def check_inputs( + self, + prompt: str | list[str], + height: int, + width: int, + guidance_scale: float, + callback_on_step_end_tensor_inputs: list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + ): + """Check that all inputs are in correct format.""" + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + + if prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and guidance_scale > 1.0 and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided and `guidance_scale > 1.0`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + + # The latents must be divisible by the transformer's patch size after VAE compression. + dimension_multiple = self.vae_scale_factor * self.transformer.config.patch_size + if height % dimension_multiple != 0 or width % dimension_multiple != 0: + raise ValueError( + f"`height` and `width` have to be divisible by {dimension_multiple} (vae_scale_factor *" + f" transformer patch_size) but are {height} and {width}." + ) + + if guidance_scale < 1.0: + raise ValueError(f"guidance_scale has to be >= 1.0 but is {guidance_scale}") + + if callback_on_step_end_tensor_inputs is not None and not isinstance(callback_on_step_end_tensor_inputs, list): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be a list but is {callback_on_step_end_tensor_inputs}" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: torch.device, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + ): + """Prepare initial latents for the diffusion process. + + PRXPixel trains with a non-unit initial-noise scale, so the sampled noise is multiplied by + `self.config.noise_scale`. + """ + if latents is None: + spatial_compression = self.vae_scale_factor + latent_height, latent_width = ( + height // spatial_compression, + width // spatial_compression, + ) + shape = (batch_size, num_channels_latents, latent_height, latent_width) + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) * self.config.noise_scale + else: + latents = latents.to(device) + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + timesteps: list[int] = None, + guidance_scale: float = 4.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + prompt_attention_mask: torch.BoolTensor | None = None, + negative_prompt_attention_mask: torch.BoolTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + use_resolution_binning: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str`, *optional*, defaults to `""`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to `default_sample_size`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `default_sample_size`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 28): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided and `guidance_scale > 1`, negative embeddings will be generated from an + empty string. + prompt_attention_mask (`torch.BoolTensor`, *optional*): + Pre-generated attention mask for `prompt_embeds`. If not provided, attention mask will be generated + from `prompt` input argument. + negative_prompt_attention_mask (`torch.BoolTensor`, *optional*): + Pre-generated attention mask for `negative_prompt_embeds`. If not provided and `guidance_scale > 1`, + attention mask will be generated from an empty string. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.prx.PRXPipelineOutput`] instead of a plain tuple. + use_resolution_binning (`bool`, *optional*, defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + predefined aspect ratio bins. After the produced latents are decoded into images, they are resized back + to the requested resolution. Useful for generating non-square images at optimal resolutions. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self, step, timestep, callback_kwargs)`. + `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include tensors that are listed + in the `._callback_tensor_inputs` attribute. + + Examples: + + Returns: + [`~pipelines.prx.PRXPipelineOutput`] or `tuple`: [`~pipelines.prx.PRXPipelineOutput`] if `return_dict` is + True, otherwise a `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + # 0. Set height and width + default_resolution = getattr(self.config, "default_sample_size", None) or PRX_PIXEL_DEFAULT_RESOLUTION + height = height or default_resolution + width = width or default_resolution + + if use_resolution_binning: + if self.config.default_sample_size not in ASPECT_RATIO_BINS: + raise ValueError( + f"Resolution binning is only supported for default_sample_size in {list(ASPECT_RATIO_BINS.keys())}, " + f"but got {self.config.default_sample_size}. Set use_resolution_binning=False to disable aspect ratio binning." + ) + aspect_ratio_bin = ASPECT_RATIO_BINS[self.config.default_sample_size] + + # Store original dimensions + orig_height, orig_width = height, width + # Map to closest resolution in the bin + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + # 1. Check inputs + self.check_inputs( + prompt, + height, + width, + guidance_scale, + callback_on_step_end_tensor_inputs, + prompt_embeds, + negative_prompt_embeds, + ) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Use execution device (handles offloading scenarios including group offloading) + device = self._execution_device + + self._guidance_scale = guidance_scale + + # 2. Encode input prompt + text_embeddings, cross_attn_mask, uncond_text_embeddings, uncond_cross_attn_mask = self.encode_prompt( + prompt, + device, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + ) + # Expose standard names for callbacks parity + prompt_embeds = text_embeddings + negative_prompt_embeds = uncond_text_embeddings + + # 3. Prepare timesteps + if timesteps is not None: + self.scheduler.set_timesteps(timesteps=timesteps, device=device) + timesteps = self.scheduler.timesteps + num_inference_steps = len(timesteps) + else: + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + self.num_timesteps = len(timesteps) + + # 4. Prepare latent variables (pixel space: in_channels RGB tensors, no VAE) + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + text_embeddings.dtype, + device, + generator, + latents, + ) + + # 5. Prepare extra step kwargs + extra_step_kwargs = {} + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_eta: + extra_step_kwargs["eta"] = 0.0 + + # 6. Prepare cross-attention embeddings and masks + if self.do_classifier_free_guidance: + ca_embed = torch.cat([uncond_text_embeddings, text_embeddings], dim=0) + ca_mask = None + if cross_attn_mask is not None and uncond_cross_attn_mask is not None: + ca_mask = torch.cat([uncond_cross_attn_mask, cross_attn_mask], dim=0) + else: + ca_embed = text_embeddings + ca_mask = cross_attn_mask + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Duplicate latents if using classifier-free guidance + if self.do_classifier_free_guidance: + latents_in = torch.cat([latents, latents], dim=0) + # Normalize timestep for the transformer + t_cont = (t.float() / self.scheduler.config.num_train_timesteps).view(1).repeat(2).to(device) + else: + latents_in = latents + # Normalize timestep for the transformer + t_cont = (t.float() / self.scheduler.config.num_train_timesteps).view(1).to(device) + + # Forward through transformer + noise_pred = self.transformer( + hidden_states=latents_in, + timestep=t_cont, + encoder_hidden_states=ca_embed, + attention_mask=ca_mask, + return_dict=False, + )[0] + + # Apply CFG + if self.do_classifier_free_guidance: + noise_uncond, noise_text = noise_pred.chunk(2, dim=0) + noise_pred = noise_uncond + guidance_scale * (noise_text - noise_uncond) + + # PRXPixel predicts x0; convert to flow-matching velocity before the scheduler step. + t_x = torch.clamp(t.float() / self.scheduler.config.num_train_timesteps, min=0.05) + noise_pred = (latents - noise_pred) / t_x + + # Compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_on_step_end(self, i, t, callback_kwargs) + + # Call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + # 8. Post-processing (pixel space: the denoised output IS the image in [-1, 1]; no VAE decode). + if output_type in ["latent", "pt"]: + image = latents + else: + image = latents + # Resize back to original resolution if using binning + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + # Use standard image processor for post-processing + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return PRXPipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/__init__.py b/diffusers/pipelines/qwenimage/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3f43d0ebb0b9f0c2ddf7786da461ead8541cc15f --- /dev/null +++ b/diffusers/pipelines/qwenimage/__init__.py @@ -0,0 +1,65 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["QwenImagePipelineOutput", "QwenImagePriorReduxPipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["modeling_qwenimage"] = ["ReduxImageEncoder"] + _import_structure["pipeline_qwenimage"] = ["QwenImagePipeline"] + _import_structure["pipeline_qwenimage_controlnet"] = ["QwenImageControlNetPipeline"] + _import_structure["pipeline_qwenimage_controlnet_inpaint"] = ["QwenImageControlNetInpaintPipeline"] + _import_structure["pipeline_qwenimage_edit"] = ["QwenImageEditPipeline"] + _import_structure["pipeline_qwenimage_edit_inpaint"] = ["QwenImageEditInpaintPipeline"] + _import_structure["pipeline_qwenimage_edit_plus"] = ["QwenImageEditPlusPipeline"] + _import_structure["pipeline_qwenimage_img2img"] = ["QwenImageImg2ImgPipeline"] + _import_structure["pipeline_qwenimage_inpaint"] = ["QwenImageInpaintPipeline"] + _import_structure["pipeline_qwenimage_layered"] = ["QwenImageLayeredPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_qwenimage import QwenImagePipeline + from .pipeline_qwenimage_controlnet import QwenImageControlNetPipeline + from .pipeline_qwenimage_controlnet_inpaint import QwenImageControlNetInpaintPipeline + from .pipeline_qwenimage_edit import QwenImageEditPipeline + from .pipeline_qwenimage_edit_inpaint import QwenImageEditInpaintPipeline + from .pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline + from .pipeline_qwenimage_img2img import QwenImageImg2ImgPipeline + from .pipeline_qwenimage_inpaint import QwenImageInpaintPipeline + from .pipeline_qwenimage_layered import QwenImageLayeredPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/qwenimage/pipeline_output.py b/diffusers/pipelines/qwenimage/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..e4ed06856e6a60128d9286344aaa1d19ffdac6e8 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class QwenImagePipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage.py new file mode 100644 index 0000000000000000000000000000000000000000..1da0518a4f65fcdbf503abb237d3819dc7b4aa9b --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage.py @@ -0,0 +1,723 @@ +# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import QwenImagePipeline + + >>> pipe = QwenImagePipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(prompt, num_inference_steps=50).images[0] + >>> image.save("qwenimage.png") + ``` +""" + + +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImagePipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is enabled by + setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale encourages to + generate images that are closely linked to the text `prompt`, usually at the expense of lower image + quality. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [[(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)]] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet.py new file mode 100644 index 0000000000000000000000000000000000000000..f946fdf27d00177ca5318d525a93bc67b5bf4a4e --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet.py @@ -0,0 +1,966 @@ +# Copyright 2025 Qwen-Image Team, InstantX Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...models.controlnets.controlnet_qwenimage import QwenImageControlNetModel, QwenImageMultiControlNetModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers.utils import load_image + >>> from diffusers import QwenImageControlNetModel, QwenImageMultiControlNetModel, QwenImageControlNetPipeline + + >>> # QwenImageControlNetModel + >>> controlnet = QwenImageControlNetModel.from_pretrained( + ... "InstantX/Qwen-Image-ControlNet-Union", torch_dtype=torch.bfloat16 + ... ) + >>> pipe = QwenImageControlNetPipeline.from_pretrained( + ... "Qwen/Qwen-Image", controlnet=controlnet, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "Aesthetics art, traditional asian pagoda, elaborate golden accents, sky blue and white color palette, swirling cloud pattern, digital illustration, east asian architecture, ornamental rooftop, intricate detailing on building, cultural representation." + >>> negative_prompt = " " + >>> control_image = load_image( + ... "https://huggingface.co/InstantX/Qwen-Image-ControlNet-Union/resolve/main/conds/canny.png" + ... ) + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe( + ... prompt, + ... negative_prompt=negative_prompt, + ... control_image=control_image, + ... controlnet_conditioning_scale=1.0, + ... num_inference_steps=30, + ... true_cfg_scale=4.0, + ... ).images[0] + >>> image.save("qwenimage_cn_union.png") + + >>> # QwenImageMultiControlNetModel + >>> controlnet = QwenImageControlNetModel.from_pretrained( + ... "InstantX/Qwen-Image-ControlNet-Union", torch_dtype=torch.bfloat16 + ... ) + >>> controlnet = QwenImageMultiControlNetModel([controlnet]) + >>> pipe = QwenImageControlNetPipeline.from_pretrained( + ... "Qwen/Qwen-Image", controlnet=controlnet, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "Aesthetics art, traditional asian pagoda, elaborate golden accents, sky blue and white color palette, swirling cloud pattern, digital illustration, east asian architecture, ornamental rooftop, intricate detailing on building, cultural representation." + >>> negative_prompt = " " + >>> control_image = load_image( + ... "https://huggingface.co/InstantX/Qwen-Image-ControlNet-Union/resolve/main/conds/canny.png" + ... ) + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe( + ... prompt, + ... negative_prompt=negative_prompt, + ... control_image=[control_image, control_image], + ... controlnet_conditioning_scale=[0.5, 0.5], + ... num_inference_steps=30, + ... true_cfg_scale=4.0, + ... ).images[0] + >>> image.save("qwenimage_cn_union_multi.png") + ``` +""" + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImageControlNetPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + controlnet: QwenImageControlNetModel | QwenImageMultiControlNetModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + controlnet=controlnet, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + control_image: PipelineImageInput = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is enabled by + setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale encourages to + generate images that are closely linked to the text `prompt`, usually at the expense of lower image + quality. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + control_image (`PipelineImageInput`, *optional*): + The ControlNet input condition to provide guidance for the generation. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `transformer`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(control_image) if isinstance(self.controlnet, QwenImageMultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 3. Prepare control image + num_channels_latents = self.transformer.config.in_channels // 4 + if isinstance(self.controlnet, QwenImageControlNetModel): + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image.shape[-2:] + + if control_image.ndim == 4: + control_image = control_image.unsqueeze(2) + + # vae encode + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) + latents_mean = (torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1)).to( + device + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device + ) + + control_image = retrieve_latents(self.vae.encode(control_image), generator=generator) + control_image = (control_image - latents_mean) * latents_std + + control_image = control_image.permute(0, 2, 1, 3, 4) + + # pack + control_image = self._pack_latents( + control_image, + batch_size=control_image.shape[0], + num_channels_latents=num_channels_latents, + height=control_image.shape[3], + width=control_image.shape[4], + ).to(dtype=prompt_embeds.dtype, device=device) + + else: + if isinstance(self.controlnet, QwenImageMultiControlNetModel): + control_images = [] + for control_image_ in control_image: + control_image_ = self.prepare_image( + image=control_image_, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + + height, width = control_image_.shape[-2:] + + if control_image_.ndim == 4: + control_image_ = control_image_.unsqueeze(2) + + # vae encode + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1) + ).to(device) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view( + 1, self.vae.config.z_dim, 1, 1, 1 + ).to(device) + + control_image_ = retrieve_latents(self.vae.encode(control_image_), generator=generator) + control_image_ = (control_image_ - latents_mean) * latents_std + + control_image_ = control_image_.permute(0, 2, 1, 3, 4) + + # pack + control_image_ = self._pack_latents( + control_image_, + batch_size=control_image_.shape[0], + num_channels_latents=num_channels_latents, + height=control_image_.shape[3], + width=control_image_.shape[4], + ).to(dtype=prompt_embeds.dtype, device=device) + + control_images.append(control_image_) + + control_image = control_images + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(self.controlnet, QwenImageControlNetModel) else keeps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + # controlnet + controlnet_block_samples = self.controlnet( + hidden_states=latents, + controlnet_cond=control_image, + conditioning_scale=cond_scale, + timestep=timestep / 1000, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_mask=prompt_embeds_mask, + img_shapes=img_shapes, + return_dict=False, + ) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_mask=prompt_embeds_mask, + img_shapes=img_shapes, + controlnet_block_samples=controlnet_block_samples, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + controlnet_block_samples=controlnet_block_samples, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet_inpaint.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..97f510a6dbf40f690afc9cc18831e98481b9d330 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_controlnet_inpaint.py @@ -0,0 +1,937 @@ +# Copyright 2025 Qwen-Image Team, The InstantX Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...models.controlnets.controlnet_qwenimage import QwenImageControlNetModel, QwenImageMultiControlNetModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers.utils import load_image + >>> from diffusers import QwenImageControlNetModel, QwenImageControlNetInpaintPipeline + + >>> base_model_path = "Qwen/Qwen-Image" + >>> controlnet_model_path = "InstantX/Qwen-Image-ControlNet-Inpainting" + >>> controlnet = QwenImageControlNetModel.from_pretrained(controlnet_model_path, torch_dtype=torch.bfloat16) + >>> pipe = QwenImageControlNetInpaintPipeline.from_pretrained( + ... base_model_path, controlnet=controlnet, torch_dtype=torch.bfloat16 + ... ).to("cuda") + >>> image = load_image( + ... "https://huggingface.co/InstantX/Qwen-Image-ControlNet-Inpainting/resolve/main/assets/images/image1.png" + ... ) + >>> mask_image = load_image( + ... "https://huggingface.co/InstantX/Qwen-Image-ControlNet-Inpainting/resolve/main/assets/masks/mask1.png" + ... ) + >>> prompt = "一辆绿色的出租车行驶在路上" + >>> result = pipe( + ... prompt=prompt, + ... control_image=image, + ... control_mask=mask_image, + ... controlnet_conditioning_scale=1.0, + ... width=mask_image.size[0], + ... height=mask_image.size[1], + ... true_cfg_scale=4.0, + ... ).images[0] + >>> image.save("qwenimage_controlnet_inpaint.png") + ``` +""" + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImageControlNetInpaintPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + controlnet: QwenImageControlNetModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + controlnet=controlnet, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + do_resize=True, + do_convert_grayscale=True, + do_normalize=False, + do_binarize=True, + ) + + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.prepare_latents + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + def prepare_image_with_mask( + self, + image, + mask, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + image = image.to(device=device, dtype=dtype) # (bsz, 3, height_ori, width_ori) + + # Prepare mask + if isinstance(mask, torch.Tensor): + pass + else: + mask = self.mask_processor.preprocess(mask, height=height, width=width) + mask = mask.repeat_interleave(repeat_by, dim=0) + mask = mask.to(device=device, dtype=dtype) # (bsz, 1, height_ori, width_ori) + + if image.ndim == 4: + image = image.unsqueeze(2) + + if mask.ndim == 4: + mask = mask.unsqueeze(2) + + # Get masked image + masked_image = image.clone() + masked_image[(mask > 0.5).repeat(1, 3, 1, 1, 1)] = -1 # (bsz, 3, 1, height_ori, width_ori) + + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) + latents_mean = (torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1)).to(device) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device + ) + + # Encode to latents + image_latents = self.vae.encode(masked_image.to(self.vae.dtype)).latent_dist.sample() + image_latents = (image_latents - latents_mean) * latents_std + image_latents = image_latents.to(dtype) # torch.Size([1, 16, 1, height_ori//8, width_ori//8]) + + mask = torch.nn.functional.interpolate( + mask, size=(image_latents.shape[-3], image_latents.shape[-2], image_latents.shape[-1]) + ) + mask = 1 - mask # torch.Size([1, 1, 1, height_ori//8, width_ori//8]) + + control_image = torch.cat( + [image_latents, mask], dim=1 + ) # torch.Size([1, 16+1, 1, height_ori//8, width_ori//8]) + + control_image = control_image.permute(0, 2, 1, 3, 4) # torch.Size([1, 1, 16+1, height_ori//8, width_ori//8]) + + # pack + control_image = self._pack_latents( + control_image, + batch_size=control_image.shape[0], + num_channels_latents=control_image.shape[2], + height=control_image.shape[3], + width=control_image.shape[4], + ) + + if do_classifier_free_guidance and not guess_mode: + control_image = torch.cat([control_image] * 2) + + return control_image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 1.0, + control_guidance_start: float | list[float] = 0.0, + control_guidance_end: float | list[float] = 1.0, + control_image: PipelineImageInput = None, + control_mask: PipelineImageInput = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 3.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + control_guidance_start (`float` or `list[float]`, *optional*, defaults to 0.0): + The percentage of total steps at which the ControlNet starts applying. + control_guidance_end (`float` or `list[float]`, *optional*, defaults to 1.0): + The percentage of total steps at which the ControlNet stops applying. + control_image (`PipelineImageInput`, *optional*): + The ControlNet input condition to provide guidance for the generation. + control_mask (`PipelineImageInput`, *optional*): + The inpainting mask for the ControlNet input condition. White pixels are repainted while black pixels + are preserved. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `transformer`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(control_image) if isinstance(self.controlnet, QwenImageMultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = ( + mult * [control_guidance_start], + mult * [control_guidance_end], + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 3. Prepare control image + num_channels_latents = self.transformer.config.in_channels // 4 + if isinstance(self.controlnet, QwenImageControlNetModel): + control_image = self.prepare_image_with_mask( + image=control_image, + mask=control_mask, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(self.controlnet, QwenImageControlNetModel) else keeps) + + # handle guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + # controlnet + controlnet_block_samples = self.controlnet( + hidden_states=latents, + controlnet_cond=control_image.to(dtype=latents.dtype, device=device), + conditioning_scale=cond_scale, + timestep=timestep / 1000, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_mask=prompt_embeds_mask, + img_shapes=img_shapes, + return_dict=False, + ) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_mask=prompt_embeds_mask, + img_shapes=img_shapes, + controlnet_block_samples=controlnet_block_samples, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + controlnet_block_samples=controlnet_block_samples, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py new file mode 100644 index 0000000000000000000000000000000000000000..85abb815cf23c9393a743ff82e769da5d6d36a64 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit.py @@ -0,0 +1,852 @@ +# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import QwenImageEditPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageEditPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png" + ... ).convert("RGB") + >>> prompt = ( + ... "Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors" + ... ) + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(image, prompt, num_inference_steps=50).images[0] + >>> image.save("qwenimage_edit.png") + ``` +""" + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def calculate_dimensions(target_area, ratio): + width = math.sqrt(target_area * ratio) + height = width / ratio + + width = round(width / 32) * 32 + height = round(height / 32) * 32 + + return width, height, None + + +class QwenImageEditPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The Qwen-Image-Edit pipeline for image editing. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + processor: Qwen2VLProcessor, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + processor=processor, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 1024 + + self.prompt_template_encode = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 64 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + image: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + + model_inputs = self.processor( + text=txt, + images=image, + padding=True, + return_tensors="pt", + ).to(device) + + outputs = self.text_encoder( + input_ids=model_inputs["input_ids"], + attention_mask=model_inputs["attention_mask"], + pixel_values=model_inputs.get("pixel_values"), + image_grid_thw=model_inputs.get("image_grid_thw"), + output_hidden_states=True, + ) + + hidden_states = outputs.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs["attention_mask"]) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + image: torch.Tensor | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + image (`torch.Tensor`, *optional*): + image to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, image, device) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.latent_channels, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, self.latent_channels, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + image_latents = (image_latents - latents_mean) / latents_std + + return image_latents + + def prepare_latents( + self, + image, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + image_latents = None + if image is not None: + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latent_height, image_latent_width = image_latents.shape[3:] + image_latents = self._pack_latents( + image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free + Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is + enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale + encourages to generate images that are closely linked to the text `prompt`, usually at the expense of + lower image quality. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + image_size = image[0].size if isinstance(image, list) else image.size + calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1]) + height = height or calculated_height + width = width or calculated_width + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # 3. Preprocess image + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + image = self.image_processor.resize(image, calculated_height, calculated_width) + prompt_image = image + image = self.image_processor.preprocess(image, calculated_height, calculated_width) + image = image.unsqueeze(2) + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + image=prompt_image, + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + image=prompt_image, + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents = self.prepare_latents( + image, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [ + [ + (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2), + (1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2), + ] + ] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_inpaint.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..57d1fdaaf99f84402bdc697ceba25249d0057284 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_inpaint.py @@ -0,0 +1,1083 @@ +# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import QwenImageEditInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageEditInpaintPipeline.from_pretrained("Qwen/Qwen-Image-Edit", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe( + ... prompt=prompt, negative_prompt=" ", image=source, mask_image=mask, strength=1.0, num_inference_steps=50 + ... ).images[0] + >>> image.save("qwenimage_inpainting.png") + ``` +""" + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.calculate_dimensions +def calculate_dimensions(target_area, ratio): + width = math.sqrt(target_area * ratio) + height = width / ratio + + width = round(width / 32) * 32 + height = round(height / 32) * 32 + + return width, height, None + + +class QwenImageEditInpaintPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The Qwen-Image-Edit pipeline for image editing. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + processor: Qwen2VLProcessor, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + processor=processor, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.vl_processor = processor + self.tokenizer_max_length = 1024 + + self.prompt_template_encode = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 64 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + image: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + + model_inputs = self.processor( + text=txt, + images=image, + padding=True, + return_tensors="pt", + ).to(device) + + outputs = self.text_encoder( + input_ids=model_inputs["input_ids"], + attention_mask=model_inputs["attention_mask"], + pixel_values=model_inputs.get("pixel_values"), + image_grid_thw=model_inputs.get("image_grid_thw"), + output_hidden_states=True, + ) + + hidden_states = outputs.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs["attention_mask"]) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + image: torch.Tensor | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + image (`torch.Tensor`, *optional*): + image to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, image, device) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_inpaint.QwenImageInpaintPipeline.check_inputs + def check_inputs( + self, + prompt, + image, + mask_image, + strength, + height, + width, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_img2img.QwenImageImg2ImgPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + image_latents.device, image_latents.dtype + ) + + image_latents = (image_latents - latents_mean) * latents_std + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_inpaint.QwenImageInpaintPipeline.prepare_latents + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + # If image is [B,C,H,W] -> add T=1. If it's already [B,C,T,H,W], leave it. + if image.dim() == 4: + image = image.unsqueeze(2) + elif image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {image.dim()}.") + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) # [B,z,1,H',W'] + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latents = image_latents.transpose(1, 2) # [B,1,z,H',W'] + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width) + image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents, noise, image_latents + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_inpaint.QwenImageInpaintPipeline.prepare_mask_latents + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate(mask, size=(height, width)) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if masked_image.dim() == 4: + masked_image = masked_image.unsqueeze(2) + elif masked_image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {masked_image.dim()}.") + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == self.latent_channels: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(image=masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask = self._pack_latents( + mask.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + true_cfg_scale: float = 4.0, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free + Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is + enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale + encourages to generate images that are closely linked to the text `prompt`, usually at the expense of + lower image quality. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, `list[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will ge generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + image_size = image[0].size if isinstance(image, list) else image.size + calculated_width, calculated_height, _ = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1]) + + # height and width are the same as the calculated height and width + height = calculated_height + width = calculated_width + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + mask_image, + strength, + height, + width, + output_type=output_type, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + padding_mask_crop=padding_mask_crop, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # 3. Preprocess image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + image = self.image_processor.resize(image, calculated_height, calculated_width) + original_image = image + prompt_image = image + image = self.image_processor.preprocess( + image, + height=calculated_height, + width=calculated_width, + crops_coords=crops_coords, + resize_mode=resize_mode, + ) + image = image.to(dtype=torch.float32) + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + image=prompt_image, + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + image=prompt_image, + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, noise, image_latents = self.prepare_latents( + image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + img_shapes = [ + [ + (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2), + (1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2), + ] + ] * batch_size + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + # for 64 channel transformer only. + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [ + self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image + ] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_plus.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_plus.py new file mode 100644 index 0000000000000000000000000000000000000000..84d1b60152b11b335020cc4649b4f5e81a0e65fb --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_edit_plus.py @@ -0,0 +1,896 @@ +# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import QwenImageEditPlusPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageEditPlusPipeline.from_pretrained("Qwen/Qwen-Image-Edit-2509", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png" + ... ).convert("RGB") + >>> prompt = ( + ... "Make Pikachu hold a sign that says 'Qwen Edit is awesome', yarn art style, detailed, vibrant colors" + ... ) + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> image = pipe(image, prompt, num_inference_steps=50).images[0] + >>> image.save("qwenimage_edit_plus.png") + ``` +""" + +CONDITION_IMAGE_SIZE = 384 * 384 +VAE_IMAGE_SIZE = 1024 * 1024 + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def calculate_dimensions(target_area, ratio): + width = math.sqrt(target_area * ratio) + height = width / ratio + + width = round(width / 32) * 32 + height = round(height / 32) * 32 + + return width, height + + +class QwenImageEditPlusPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The Qwen-Image-Edit pipeline for image editing. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + processor: Qwen2VLProcessor, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + processor=processor, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.tokenizer_max_length = 1024 + + self.prompt_template_encode = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 64 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + image: torch.Tensor | None = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + img_prompt_template = "Picture {}: <|vision_start|><|image_pad|><|vision_end|>" + if isinstance(image, list): + base_img_prompt = "" + for i, img in enumerate(image): + base_img_prompt += img_prompt_template.format(i + 1) + elif image is not None: + base_img_prompt = img_prompt_template.format(1) + else: + base_img_prompt = "" + + template = self.prompt_template_encode + + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(base_img_prompt + e) for e in prompt] + + model_inputs = self.processor( + text=txt, + images=image, + padding=True, + return_tensors="pt", + ).to(device) + + outputs = self.text_encoder( + input_ids=model_inputs["input_ids"], + attention_mask=model_inputs["attention_mask"], + pixel_values=model_inputs.get("pixel_values"), + image_grid_thw=model_inputs.get("image_grid_thw"), + output_hidden_states=True, + ) + + hidden_states = outputs.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, model_inputs["attention_mask"]) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + image: torch.Tensor | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + image (`torch.Tensor`, *optional*): + image to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, image, device) + + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline.check_inputs + def check_inputs( + self, + prompt, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.latent_channels, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, self.latent_channels, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + image_latents = (image_latents - latents_mean) / latents_std + + return image_latents + + def prepare_latents( + self, + images, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + image_latents = None + if images is not None: + if not isinstance(images, list): + images = [images] + all_image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latent_height, image_latent_width = image_latents.shape[3:] + image_latents = self._pack_latents( + image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width + ) + all_image_latents.append(image_latents) + image_latents = torch.cat(all_image_latents, dim=1) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free + Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is + enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale + encourages to generate images that are closely linked to the text `prompt`, usually at the expense of + lower image quality. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + image_size = image[-1].size if isinstance(image, list) else image.size + calculated_width, calculated_height = calculate_dimensions(1024 * 1024, image_size[0] / image_size[1]) + height = height or calculated_height + width = width or calculated_width + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # QwenImageEditPlusPipeline does not currently support batch_size > 1 + if batch_size > 1: + raise ValueError( + f"QwenImageEditPlusPipeline currently only supports batch_size=1, but received batch_size={batch_size}. " + "Please process prompts one at a time." + ) + + device = self._execution_device + # 3. Preprocess image + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + if not isinstance(image, list): + image = [image] + condition_image_sizes = [] + condition_images = [] + vae_image_sizes = [] + vae_images = [] + for img in image: + image_width, image_height = img.size + condition_width, condition_height = calculate_dimensions( + CONDITION_IMAGE_SIZE, image_width / image_height + ) + vae_width, vae_height = calculate_dimensions(VAE_IMAGE_SIZE, image_width / image_height) + condition_image_sizes.append((condition_width, condition_height)) + vae_image_sizes.append((vae_width, vae_height)) + condition_images.append(self.image_processor.resize(img, condition_height, condition_width)) + vae_images.append(self.image_processor.preprocess(img, vae_height, vae_width).unsqueeze(2)) + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + image=condition_images, + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + image=condition_images, + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents = self.prepare_latents( + vae_images, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [ + [ + (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2), + *[ + (1, vae_height // self.vae_scale_factor // 2, vae_width // self.vae_scale_factor // 2) + for vae_width, vae_height in vae_image_sizes + ], + ] + ] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = latents.shape[1] + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..9b9af83737e55056dfc87fd09b83765c77067fc3 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_img2img.py @@ -0,0 +1,826 @@ +import inspect +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import QwenImageImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageImg2ImgPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16) + >>> pipe = pipe.to("cuda") + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + >>> prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney" + >>> images = pipe(prompt=prompt, negative_prompt=" ", image=init_image, strength=0.95).images[0] + >>> images.save("qwenimage_img2img.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImageImg2ImgPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + image_latents.device, image_latents.dtype + ) + + image_latents = (image_latents - latents_mean) * latents_std + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + strength, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + # If image is [B,C,H,W] -> add T=1. If it's already [B,C,T,H,W], leave it. + if image.dim() == 4: + image = image.unsqueeze(2) + elif image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {image.dim()}.") + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) # [B,z,1,H',W'] + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latents = image_latents.transpose(1, 2) # [B,1,z,H',W'] + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is enabled by + setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale encourages to + generate images that are closely linked to the text `prompt`, usually at the expense of lower image + quality. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Preprocess image + init_image = self.image_processor.preprocess(image, height=height, width=width) + init_image = init_image.to(dtype=torch.float32) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [[(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)]] * batch_size + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..3d5f0040932a82f591ec8bb7322e7c6bbe8db679 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_inpaint.py @@ -0,0 +1,1012 @@ +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import QwenImageInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageInpaintPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, negative_prompt=" ", image=source, mask_image=mask, strength=0.85).images[0] + >>> image.save("qwenimage_inpainting.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class QwenImageInpaintPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The QwenImage pipeline for text-to-image generation. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + vae_latent_channels=self.latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = 1024 + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._get_qwen_prompt_embeds + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt" + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_img2img.QwenImageImg2ImgPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + image_latents.device, image_latents.dtype + ) + + image_latents = (image_latents - latents_mean) * latents_std + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def check_inputs( + self, + prompt, + image, + mask_image, + strength, + height, + width, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._unpack_latents + def _unpack_latents(latents, height, width, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 3, 1, 4, 2, 5) + + latents = latents.reshape(batch_size, channels // (2 * 2), 1, height, width) + + return latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, 1, num_channels_latents, height, width) + + # If image is [B,C,H,W] -> add T=1. If it's already [B,C,T,H,W], leave it. + if image.dim() == 4: + image = image.unsqueeze(2) + elif image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {image.dim()}.") + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) # [B,z,1,H',W'] + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latents = image_latents.transpose(1, 2) # [B,1,z,H',W'] + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + noise = self._pack_latents(noise, batch_size, num_channels_latents, height, width) + image_latents = self._pack_latents(image_latents, batch_size, num_channels_latents, height, width) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width) + + return latents, noise, image_latents + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate(mask, size=(height, width)) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if masked_image.dim() == 4: + masked_image = masked_image.unsqueeze(2) + elif masked_image.dim() != 5: + raise ValueError(f"Expected image dims 4 or 5, got {masked_image.dim()}.") + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == self.latent_channels: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(image=masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1, 1) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + masked_image_latents = self._pack_latents( + masked_image_latents, + batch_size, + num_channels_latents, + height, + width, + ) + mask = self._pack_latents( + mask.repeat(1, num_channels_latents, 1, 1), + batch_size, + num_channels_latents, + height, + width, + ) + + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + true_cfg_scale (`float`, *optional*, defaults to 1.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is enabled by + setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale encourages to + generate images that are closely linked to the text `prompt`, usually at the expense of lower image + quality. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, `list[torch.Tensor]`): + `Tensor` representing an image batch to mask `image` generated by VAE. If not provided, the mask + latents tensor will be generated by `mask_image`. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + image, + mask_image, + strength, + height, + width, + output_type=output_type, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + padding_mask_crop=padding_mask_crop, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + # 2. Preprocess image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + + latents, noise, image_latents = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_channels_latents, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + ) + + img_shapes = [[(1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2)]] * batch_size + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latents, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + return_dict=False, + )[0] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + # for 64 channel transformer only. + init_latents_proper = image_latents + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + + latents = latents / latents_std + latents_mean + image = self.vae.decode(latents, return_dict=False)[0][:, :, 0] + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [ + self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image + ] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return QwenImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/qwenimage/pipeline_qwenimage_layered.py b/diffusers/pipelines/qwenimage/pipeline_qwenimage_layered.py new file mode 100644 index 0000000000000000000000000000000000000000..7e06a7d36ffdca11017b13e7c3148f17cd104504 --- /dev/null +++ b/diffusers/pipelines/qwenimage/pipeline_qwenimage_layered.py @@ -0,0 +1,913 @@ +# Copyright 2025 Qwen-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import math +from typing import Any, Callable + +import numpy as np +import torch +from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2Tokenizer, Qwen2VLProcessor + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import QwenImageLoraLoaderMixin +from ...models import AutoencoderKLQwenImage, QwenImageTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import QwenImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from PIL import Image + >>> from diffusers import QwenImageLayeredPipeline + >>> from diffusers.utils import load_image + + >>> pipe = QwenImageLayeredPipeline.from_pretrained("Qwen/Qwen-Image-Layered", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/yarn-art-pikachu.png" + ... ).convert("RGBA") + >>> prompt = "" + >>> # Depending on the variant being used, the pipeline call will slightly vary. + >>> # Refer to the pipeline documentation for more details. + >>> images = pipe( + ... image, + ... prompt, + ... num_inference_steps=50, + ... true_cfg_scale=4.0, + ... layers=4, + ... resolution=640, + ... cfg_normalize=False, + ... use_en_prompt=True, + ... ).images[0] + >>> for i, image in enumerate(images): + ... image.save(f"{i}.out.png") + ``` +""" + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit_plus.calculate_dimensions +def calculate_dimensions(target_area, ratio): + width = math.sqrt(target_area * ratio) + height = width / ratio + + width = round(width / 32) * 32 + height = round(height / 32) * 32 + + return width, height + + +class QwenImageLayeredPipeline(DiffusionPipeline, QwenImageLoraLoaderMixin): + r""" + The Qwen-Image-Layered pipeline for image decomposing. + + Args: + transformer ([`QwenImageTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`Qwen2.5-VL-7B-Instruct`]): + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), specifically the + [Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) variant. + tokenizer (`Qwen2Tokenizer`): + Tokenizer of class https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2Tokenizer + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKLQwenImage, + text_encoder: Qwen2_5_VLForConditionalGeneration, + tokenizer: Qwen2Tokenizer, + processor: Qwen2VLProcessor, + transformer: QwenImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + processor=processor, + transformer=transformer, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.latent_channels = self.vae.config.z_dim if getattr(self, "vae", None) else 16 + # QwenImage latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.vl_processor = processor + self.tokenizer_max_length = 1024 + + self.prompt_template_encode = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n" + self.prompt_template_encode_start_idx = 34 + self.image_caption_prompt_cn = """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# 图像标注器\n你是一个专业的图像标注器。请基于输入图像，撰写图注:\n1. +使用自然、描述性的语言撰写图注，不要使用结构化形式或富文本形式。\n2. 通过加入以下内容，丰富图注细节：\n - 对象的属性：如数量、颜色、形状、大小、位置、材质、状态、动作等\n - +对象间的视觉关系：如空间关系、功能关系、动作关系、从属关系、比较关系、因果关系等\n - 环境细节：例如天气、光照、颜色、纹理、气氛等\n - 文字内容：识别图像中清晰可见的文字，不做翻译和解释，用引号在图注中强调\n3. +保持真实性与准确性：\n - 不要使用笼统的描述\n - +描述图像中所有可见的信息，但不要加入没有在图像中出现的内容\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n""" + self.image_caption_prompt_en = """<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n# Image Annotator\nYou are a professional +image annotator. Please write an image caption based on the input image:\n1. Write the caption using natural, +descriptive language without structured formats or rich text.\n2. Enrich caption details by including: \n - Object +attributes, such as quantity, color, shape, size, material, state, position, actions, and so on\n - Vision Relations +between objects, such as spatial relations, functional relations, possessive relations, attachment relations, action +relations, comparative relations, causal relations, and so on\n - Environmental details, such as weather, lighting, +colors, textures, atmosphere, and so on\n - Identify the text clearly visible in the image, without translation or +explanation, and highlight it in the caption with quotation marks\n3. Maintain authenticity and accuracy:\n - Avoid +generalizations\n - Describe all visible information in the image, while do not add information not explicitly shown in +the image\n<|vision_start|><|image_pad|><|vision_end|><|im_end|>\n<|im_start|>assistant\n""" + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline._extract_masked_hidden + def _extract_masked_hidden(self, hidden_states: torch.Tensor, mask: torch.Tensor): + bool_mask = mask.bool() + valid_lengths = bool_mask.sum(dim=1) + selected = hidden_states[bool_mask] + split_result = torch.split(selected, valid_lengths.tolist(), dim=0) + + return split_result + + def _get_qwen_prompt_embeds( + self, + prompt: str | list[str] = None, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + + template = self.prompt_template_encode + drop_idx = self.prompt_template_encode_start_idx + txt = [template.format(e) for e in prompt] + txt_tokens = self.tokenizer( + txt, + padding=True, + return_tensors="pt", + ).to(device) + encoder_hidden_states = self.text_encoder( + input_ids=txt_tokens.input_ids, + attention_mask=txt_tokens.attention_mask, + output_hidden_states=True, + ) + hidden_states = encoder_hidden_states.hidden_states[-1] + split_hidden_states = self._extract_masked_hidden(hidden_states, txt_tokens.attention_mask) + split_hidden_states = [e[drop_idx:] for e in split_hidden_states] + attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states] + max_seq_len = max([e.size(0) for e in split_hidden_states]) + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states] + ) + encoder_attention_mask = torch.stack( + [torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list] + ) + + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + return prompt_embeds, encoder_attention_mask + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage.QwenImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + max_sequence_length: int = 1024, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds, prompt_embeds_mask = self._get_qwen_prompt_embeds(prompt, device) + + prompt_embeds = prompt_embeds[:, :max_sequence_length] + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if prompt_embeds_mask is not None: + prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length] + prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1) + prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len) + + if prompt_embeds_mask.all(): + prompt_embeds_mask = None + + return prompt_embeds, prompt_embeds_mask + + def get_image_caption(self, prompt_image, use_en_prompt=True, device=None): + if use_en_prompt: + prompt = self.image_caption_prompt_en + else: + prompt = self.image_caption_prompt_cn + model_inputs = self.vl_processor( + text=prompt, + images=prompt_image, + padding=True, + return_tensors="pt", + ).to(device) + generated_ids = self.text_encoder.generate(**model_inputs, max_new_tokens=512) + generated_ids_trimmed = [ + out_ids[len(in_ids) :] for in_ids, out_ids in zip(model_inputs.input_ids, generated_ids) + ] + output_text = self.vl_processor.batch_decode( + generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False + )[0] + return output_text.strip() + + def check_inputs( + self, + height, + width, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_embeds_mask=None, + negative_prompt_embeds_mask=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_embeds_mask is None: + logger.warning( + "`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all" + " prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as" + " `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was" + " used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None: + logger.warning( + "`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model will treat all" + " negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you should provide the padding mask as" + " `negative_prompt_embeds_mask`. Make sure to generate `negative_prompt_embeds_mask` from the same text encoder that was" + " used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 1024: + raise ValueError(f"`max_sequence_length` cannot be greater than 1024 but is {max_sequence_length}") + + @staticmethod + def _pack_latents(latents, batch_size, num_channels_latents, height, width, layers): + latents = latents.view(batch_size, layers, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 1, 3, 5, 2, 4, 6) + latents = latents.reshape(batch_size, layers * (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, height, width, layers, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (vae_scale_factor * 2)) + width = 2 * (int(width) // (vae_scale_factor * 2)) + + latents = latents.view(batch_size, layers + 1, height // 2, width // 2, channels // 4, 2, 2) + latents = latents.permute(0, 1, 4, 2, 5, 3, 6) + + latents = latents.reshape(batch_size, layers + 1, channels // (2 * 2), height, width) + latents = latents.permute(0, 2, 1, 3, 4) # (b, c, f, h, w) + + return latents + + # Copied from diffusers.pipelines.qwenimage.pipeline_qwenimage_edit.QwenImageEditPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i], sample_mode="argmax") + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax") + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.latent_channels, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std) + .view(1, self.latent_channels, 1, 1, 1) + .to(image_latents.device, image_latents.dtype) + ) + image_latents = (image_latents - latents_mean) / latents_std + + return image_latents + + def prepare_latents( + self, + image, + batch_size, + num_channels_latents, + height, + width, + layers, + dtype, + device, + generator, + latents=None, + ): + # VAE applies 8x compression on images but we must also account for packing which requires + # latent height and width to be divisible by 2. + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = ( + batch_size, + layers + 1, + num_channels_latents, + height, + width, + ) ### the generated first image is combined image + + image_latents = None + if image is not None: + image = image.to(device=device, dtype=dtype) + if image.shape[1] != self.latent_channels: + image_latents = self._encode_vae_image(image=image, generator=generator) + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + image_latent_height, image_latent_width = image_latents.shape[3:] + image_latents = image_latents.permute(0, 2, 1, 3, 4) # (b, c, f, h, w) -> (b, f, c, h, w) + image_latents = self._pack_latents( + image_latents, batch_size, num_channels_latents, image_latent_height, image_latent_width, 1 + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width, layers + 1) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents, image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + true_cfg_scale: float = 4.0, + layers: int | None = 4, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + resolution: int = 640, + cfg_normalize: bool = False, + use_en_prompt: bool = False, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is + not greater than `1`). + true_cfg_scale (`float`, *optional*, defaults to 1.0): + true_cfg_scale (`float`, *optional*, defaults to 1.0): Guidance scale as defined in [Classifier-Free + Diffusion Guidance](https://huggingface.co/papers/2207.12598). `true_cfg_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Classifier-free guidance is + enabled by setting `true_cfg_scale > 1` and a provided `negative_prompt`. Higher guidance scale + encourages to generate images that are closely linked to the text `prompt`, usually at the expense of + lower image quality. + layers (`int`, *optional*, defaults to 4): + Number of latent layers to generate for the layered output. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to None): + A guidance scale value for guidance distilled models. Unlike the traditional classifier-free guidance + where the guidance scale is applied during inference through noise prediction rescaling, guidance + distilled models take the guidance scale directly as an input parameter during forward pass. Guidance + scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images + that are closely linked to the text `prompt`, usually at the expense of lower image quality. This + parameter in the pipeline is there to support future guidance-distilled models when they come up. It is + ignored when not using guidance distilled models. To enable traditional classifier-free guidance, + please pass `true_cfg_scale > 1.0` and `negative_prompt` (even an empty negative prompt like " " should + enable classifier-free guidance computations). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `prompt_embeds`. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_mask (`torch.Tensor`, *optional*): + Attention mask for `negative_prompt_embeds`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + resolution (`int`, *optional*, defaults to 640): + using different bucket in (640, 1024) to determin the condition and output resolution + cfg_normalize (`bool`, *optional*, defaults to `False`) + whether enable cfg normalization. + use_en_prompt (`bool`, *optional*, defaults to `False`) + automatic caption language if user does not provide caption + + Examples: + + Returns: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] or `tuple`: + [`~pipelines.qwenimage.QwenImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is a list with the generated images. + """ + image_size = image[0].size if isinstance(image, list) else image.size + assert resolution in [640, 1024], f"resolution must be either 640 or 1024, but got {resolution}" + calculated_width, calculated_height = calculate_dimensions( + resolution * resolution, image_size[0] / image_size[1] + ) + height = calculated_height + width = calculated_width + + multiple_of = self.vae_scale_factor * 2 + width = width // multiple_of * multiple_of + height = height // multiple_of * multiple_of + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + height, + width, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + negative_prompt_embeds_mask=negative_prompt_embeds_mask, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + # 2. Preprocess image + if image is not None and not (isinstance(image, torch.Tensor) and image.size(1) == self.latent_channels): + image = self.image_processor.resize(image, calculated_height, calculated_width) + prompt_image = image + image = self.image_processor.preprocess(image, calculated_height, calculated_width) + image = image.unsqueeze(2) + image = image.to(dtype=self.text_encoder.dtype) + + if prompt is None or prompt == "" or prompt == " ": + prompt = self.get_image_caption(prompt_image, use_en_prompt=use_en_prompt, device=device) + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + has_neg_prompt = negative_prompt is not None or negative_prompt_embeds is not None + + if true_cfg_scale > 1 and not has_neg_prompt: + logger.warning( + f"true_cfg_scale is passed as {true_cfg_scale}, but classifier-free guidance is not enabled since no negative_prompt is provided." + ) + elif true_cfg_scale <= 1 and has_neg_prompt: + logger.warning( + " negative_prompt is passed but classifier-free guidance is not enabled since true_cfg_scale <= 1" + ) + + do_true_cfg = true_cfg_scale > 1 and has_neg_prompt + prompt_embeds, prompt_embeds_mask = self.encode_prompt( + prompt=prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_mask=prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + if do_true_cfg: + negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt( + prompt=negative_prompt, + prompt_embeds=negative_prompt_embeds, + prompt_embeds_mask=negative_prompt_embeds_mask, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels // 4 + latents, image_latents = self.prepare_latents( + image, + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + layers, + prompt_embeds.dtype, + device, + generator, + latents, + ) + img_shapes = [ + [ + *[ + (1, height // self.vae_scale_factor // 2, width // self.vae_scale_factor // 2) + for _ in range(layers + 1) + ], + (1, calculated_height // self.vae_scale_factor // 2, calculated_width // self.vae_scale_factor // 2), + ] + ] * batch_size + + # 5. Prepare timesteps + sigmas = np.linspace(1.0, 0, num_inference_steps + 1)[:-1] if sigmas is None else sigmas + image_seq_len = latents.shape[1] + base_seqlen = 256 * 256 / 16 / 16 + mu = (image_latents.shape[1] / base_seqlen) ** 0.5 + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # handle guidance + if self.transformer.config.guidance_embeds and guidance_scale is None: + raise ValueError("guidance_scale is required for guidance-distilled model.") + elif self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + elif not self.transformer.config.guidance_embeds and guidance_scale is not None: + logger.warning( + f"guidance_scale is passed as {guidance_scale}, but ignored since the model is not guidance-distilled." + ) + guidance = None + elif not self.transformer.config.guidance_embeds and guidance_scale is None: + guidance = None + + if self.attention_kwargs is None: + self._attention_kwargs = {} + + is_rgb = torch.tensor([0] * batch_size).to(device=device, dtype=torch.long) + # 6. Denoising loop + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + latent_model_input = latents + if image_latents is not None: + latent_model_input = torch.cat([latents, image_latents], dim=1) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=prompt_embeds_mask, + encoder_hidden_states=prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + additional_t_cond=is_rgb, + return_dict=False, + )[0] + noise_pred = noise_pred[:, : latents.size(1)] + + if do_true_cfg: + with self.transformer.cache_context("uncond"): + neg_noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + encoder_hidden_states_mask=negative_prompt_embeds_mask, + encoder_hidden_states=negative_prompt_embeds, + img_shapes=img_shapes, + attention_kwargs=self.attention_kwargs, + additional_t_cond=is_rgb, + return_dict=False, + )[0] + neg_noise_pred = neg_noise_pred[:, : latents.size(1)] + comb_pred = neg_noise_pred + true_cfg_scale * (noise_pred - neg_noise_pred) + + if cfg_normalize: + cond_norm = torch.norm(noise_pred, dim=-1, keepdim=True) + noise_norm = torch.norm(comb_pred, dim=-1, keepdim=True) + noise_pred = comb_pred * (cond_norm / noise_norm) + else: + noise_pred = comb_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + if output_type == "latent": + image = latents + else: + latents = self._unpack_latents(latents, height, width, layers, self.vae_scale_factor) + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + + b, c, f, h, w = latents.shape + + latents = latents[:, :, 1:] # remove the first frame as it is the orgin input + + latents = latents.permute(0, 2, 1, 3, 4).reshape(-1, c, 1, h, w) + + image = self.vae.decode(latents, return_dict=False)[0] # (b f) c 1 h w + + image = image.squeeze(2) + + image = self.image_processor.postprocess(image, output_type=output_type) + images = [] + for bidx in range(b): + images.append(image[bidx * f : (bidx + 1) * f]) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (images,) + + return QwenImagePipelineOutput(images=images) diff --git a/diffusers/pipelines/sana/__init__.py b/diffusers/pipelines/sana/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..91684f35f153516de83cd65ebed28164970b7548 --- /dev/null +++ b/diffusers/pipelines/sana/__init__.py @@ -0,0 +1,53 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_sana"] = ["SanaPipeline"] + _import_structure["pipeline_sana_controlnet"] = ["SanaControlNetPipeline"] + _import_structure["pipeline_sana_sprint"] = ["SanaSprintPipeline"] + _import_structure["pipeline_sana_sprint_img2img"] = ["SanaSprintImg2ImgPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_sana import SanaPipeline + from .pipeline_sana_controlnet import SanaControlNetPipeline + from .pipeline_sana_sprint import SanaSprintPipeline + from .pipeline_sana_sprint_img2img import SanaSprintImg2ImgPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/sana/pipeline_output.py b/diffusers/pipelines/sana/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..b9d095906e06ba975233a2a1577d1dac651c813d --- /dev/null +++ b/diffusers/pipelines/sana/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class SanaPipelineOutput(BaseOutput): + """ + Output class for Sana pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/sana/pipeline_sana.py b/diffusers/pipelines/sana/pipeline_sana.py new file mode 100644 index 0000000000000000000000000000000000000000..553e45a628d954c89d752e19df9467e2ca2dbed6 --- /dev/null +++ b/diffusers/pipelines/sana/pipeline_sana.py @@ -0,0 +1,986 @@ +# Copyright 2025 SANA Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable + +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PixArtImageProcessor +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, SanaTransformer2DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..pixart_alpha.pipeline_pixart_alpha import ( + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, +) +from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN +from .pipeline_output import SanaPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +ASPECT_RATIO_4096_BIN = { + "0.25": [2048.0, 8192.0], + "0.26": [2048.0, 7936.0], + "0.27": [2048.0, 7680.0], + "0.28": [2048.0, 7424.0], + "0.32": [2304.0, 7168.0], + "0.33": [2304.0, 6912.0], + "0.35": [2304.0, 6656.0], + "0.4": [2560.0, 6400.0], + "0.42": [2560.0, 6144.0], + "0.48": [2816.0, 5888.0], + "0.5": [2816.0, 5632.0], + "0.52": [2816.0, 5376.0], + "0.57": [3072.0, 5376.0], + "0.6": [3072.0, 5120.0], + "0.68": [3328.0, 4864.0], + "0.72": [3328.0, 4608.0], + "0.78": [3584.0, 4608.0], + "0.82": [3584.0, 4352.0], + "0.88": [3840.0, 4352.0], + "0.94": [3840.0, 4096.0], + "1.0": [4096.0, 4096.0], + "1.07": [4096.0, 3840.0], + "1.13": [4352.0, 3840.0], + "1.21": [4352.0, 3584.0], + "1.29": [4608.0, 3584.0], + "1.38": [4608.0, 3328.0], + "1.46": [4864.0, 3328.0], + "1.67": [5120.0, 3072.0], + "1.75": [5376.0, 3072.0], + "2.0": [5632.0, 2816.0], + "2.09": [5888.0, 2816.0], + "2.4": [6144.0, 2560.0], + "2.5": [6400.0, 2560.0], + "2.89": [6656.0, 2304.0], + "3.0": [6912.0, 2304.0], + "3.11": [7168.0, 2304.0], + "3.62": [7424.0, 2048.0], + "3.75": [7680.0, 2048.0], + "3.88": [7936.0, 2048.0], + "4.0": [8192.0, 2048.0], +} + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaPipeline + + >>> pipe = SanaPipeline.from_pretrained( + ... "Efficient-Large-Model/Sana_1600M_1024px_BF16_diffusers", torch_dtype=torch.float32 + ... ) + >>> pipe.to("cuda") + >>> pipe.text_encoder.to(torch.bfloat16) + >>> pipe.transformer = pipe.transformer.to(torch.bfloat16) + + >>> image = pipe(prompt='a cyberpunk cat with a neon sign that says "Sana"')[0] + >>> image[0].save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629). + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC, + transformer: SanaTransformer2DModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + if hasattr(self, "vae") and self.vae is not None + else 32 + ) + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=False, + ) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + height: int = 1024, + width: int = 1024, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> SanaPipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + attention_kwargs: + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_4096_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_2048_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 16: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + timestep = timestep * self.transformer.config.timestep_scale + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + + # compute previous image: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + try: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return SanaPipelineOutput(images=image) diff --git a/diffusers/pipelines/sana/pipeline_sana_controlnet.py b/diffusers/pipelines/sana/pipeline_sana_controlnet.py new file mode 100644 index 0000000000000000000000000000000000000000..de1910f68192e7e5c6a712db86b38cf48f8cbc0e --- /dev/null +++ b/diffusers/pipelines/sana/pipeline_sana_controlnet.py @@ -0,0 +1,1081 @@ +# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable + +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, PixArtImageProcessor +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, SanaControlNetModel, SanaTransformer2DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..pixart_alpha.pipeline_pixart_alpha import ( + ASPECT_RATIO_512_BIN, + ASPECT_RATIO_1024_BIN, +) +from ..pixart_alpha.pipeline_pixart_sigma import ASPECT_RATIO_2048_BIN +from .pipeline_output import SanaPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +ASPECT_RATIO_4096_BIN = { + "0.25": [2048.0, 8192.0], + "0.26": [2048.0, 7936.0], + "0.27": [2048.0, 7680.0], + "0.28": [2048.0, 7424.0], + "0.32": [2304.0, 7168.0], + "0.33": [2304.0, 6912.0], + "0.35": [2304.0, 6656.0], + "0.4": [2560.0, 6400.0], + "0.42": [2560.0, 6144.0], + "0.48": [2816.0, 5888.0], + "0.5": [2816.0, 5632.0], + "0.52": [2816.0, 5376.0], + "0.57": [3072.0, 5376.0], + "0.6": [3072.0, 5120.0], + "0.68": [3328.0, 4864.0], + "0.72": [3328.0, 4608.0], + "0.78": [3584.0, 4608.0], + "0.82": [3584.0, 4352.0], + "0.88": [3840.0, 4352.0], + "0.94": [3840.0, 4096.0], + "1.0": [4096.0, 4096.0], + "1.07": [4096.0, 3840.0], + "1.13": [4352.0, 3840.0], + "1.21": [4352.0, 3584.0], + "1.29": [4608.0, 3584.0], + "1.38": [4608.0, 3328.0], + "1.46": [4864.0, 3328.0], + "1.67": [5120.0, 3072.0], + "1.75": [5376.0, 3072.0], + "2.0": [5632.0, 2816.0], + "2.09": [5888.0, 2816.0], + "2.4": [6144.0, 2560.0], + "2.5": [6400.0, 2560.0], + "2.89": [6656.0, 2304.0], + "3.0": [6912.0, 2304.0], + "3.11": [7168.0, 2304.0], + "3.62": [7424.0, 2048.0], + "3.75": [7680.0, 2048.0], + "3.88": [7936.0, 2048.0], + "4.0": [8192.0, 2048.0], +} + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaControlNetPipeline + >>> from diffusers.utils import load_image + + >>> pipe = SanaControlNetPipeline.from_pretrained( + ... "ishan24/Sana_600M_1024px_ControlNetPlus_diffusers", + ... variant="fp16", + ... torch_dtype={"default": torch.bfloat16, "controlnet": torch.float16, "transformer": torch.float16}, + ... device_map="balanced", + ... ) + >>> cond_image = load_image( + ... "https://huggingface.co/ishan24/Sana_600M_1024px_ControlNet_diffusers/resolve/main/hed_example.png" + ... ) + >>> prompt = 'a cat with a neon sign that says "Sana"' + >>> image = pipe( + ... prompt, + ... control_image=cond_image, + ... ).images[0] + >>> image.save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaControlNetPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629). + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->controlnet->transformer->vae" + _callback_tensor_inputs = ["latents", "control_image", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC, + transformer: SanaTransformer2DModel, + controlnet: SanaControlNetModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + vae=vae, + transformer=transformer, + controlnet=controlnet, + scheduler=scheduler, + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + if hasattr(self, "vae") and self.vae is not None + else 32 + ) + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=False, + ) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 20, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 4.5, + control_image: PipelineImageInput = None, + controlnet_conditioning_scale: float | list[float] = 1.0, + num_images_per_prompt: int | None = 1, + height: int = 1024, + width: int = 1024, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> SanaPipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + control_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, `list[np.ndarray]`,: + `list[list[torch.Tensor]]`, `list[list[np.ndarray]]` or `list[list[PIL.Image.Image]]`): + The ControlNet input condition to provide guidance to the `unet` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or + width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, + images must be passed as a list such that each element of the list can be correctly batched for input + to a single ControlNet. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set + the corresponding scale as a list. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + attention_kwargs: + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 128: + aspect_ratio_bin = ASPECT_RATIO_4096_BIN + elif self.transformer.config.sample_size == 64: + aspect_ratio_bin = ASPECT_RATIO_2048_BIN + elif self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + elif self.transformer.config.sample_size == 16: + aspect_ratio_bin = ASPECT_RATIO_512_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare control image + if isinstance(self.controlnet, SanaControlNetModel): + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + do_classifier_free_guidance=self.do_classifier_free_guidance, + guess_mode=False, + ) + height, width = control_image.shape[-2:] + + control_image = self.vae.encode(control_image).latent + control_image = control_image * self.vae.config.scaling_factor + else: + raise ValueError("`controlnet` must be of type `SanaControlNetModel`.") + + # 5. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 6. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + controlnet_dtype = self.controlnet.dtype + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + # controlnet(s) inference + controlnet_block_samples = self.controlnet( + latent_model_input.to(dtype=controlnet_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=controlnet_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + controlnet_cond=control_image, + conditioning_scale=controlnet_conditioning_scale, + )[0] + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + controlnet_block_samples=tuple(t.to(dtype=transformer_dtype) for t in controlnet_block_samples), + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + + # compute previous image: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + try: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return SanaPipelineOutput(images=image) diff --git a/diffusers/pipelines/sana/pipeline_sana_sprint.py b/diffusers/pipelines/sana/pipeline_sana_sprint.py new file mode 100644 index 0000000000000000000000000000000000000000..812441d8e46262998a6b54914269b25ff705fe72 --- /dev/null +++ b/diffusers/pipelines/sana/pipeline_sana_sprint.py @@ -0,0 +1,868 @@ +# Copyright 2025 SANA-Sprint Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable + +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PixArtImageProcessor +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, SanaTransformer2DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..pixart_alpha.pipeline_pixart_alpha import ASPECT_RATIO_1024_BIN +from .pipeline_output import SanaPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaSprintPipeline + + >>> pipe = SanaSprintPipeline.from_pretrained( + ... "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe(prompt="a tiny astronaut hatching from an egg on the moon")[0] + >>> image[0].save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaSprintPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for text-to-image generation using [SANA-Sprint](https://huggingface.co/papers/2503.09641). + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC, + transformer: SanaTransformer2DModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + if hasattr(self, "vae") and self.vae is not None + else 32 + ) + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + num_inference_steps, + timesteps, + max_timesteps, + intermediate_timesteps, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if timesteps is not None and len(timesteps) != num_inference_steps + 1: + raise ValueError("If providing custom timesteps, `timesteps` must be of length `num_inference_steps + 1`.") + + if timesteps is not None and max_timesteps is not None: + raise ValueError("If providing custom timesteps, `max_timesteps` should not be provided.") + + if timesteps is None and max_timesteps is None: + raise ValueError("Should provide either `timesteps` or `max_timesteps`.") + + if intermediate_timesteps is not None and num_inference_steps != 2: + raise ValueError("Intermediate timesteps for SCM is not supported when num_inference_steps != 2.") + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + num_inference_steps: int = 2, + timesteps: list[int] = None, + max_timesteps: float = 1.57080, + intermediate_timesteps: float = 1.3, + guidance_scale: float = 4.5, + num_images_per_prompt: int | None = 1, + height: int = 1024, + width: int = 1024, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> SanaPipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + max_timesteps (`float`, *optional*, defaults to 1.57080): + The maximum timestep value used in the SCM scheduler. + intermediate_timesteps (`float`, *optional*, defaults to 1.3): + The intermediate timestep value used in SCM scheduler (only used when num_inference_steps=2). + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 4.5): + Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages + a model to generate images more aligned with `prompt` at the expense of lower image quality. + + Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to + the [paper](https://huggingface.co/papers/2210.03142) to learn more. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + attention_kwargs: + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt=prompt, + height=height, + width=width, + num_inference_steps=num_inference_steps, + timesteps=timesteps, + max_timesteps=max_timesteps, + intermediate_timesteps=intermediate_timesteps, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + ) = self.encode_prompt( + prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + timesteps, + sigmas=None, + max_timesteps=max_timesteps, + intermediate_timesteps=intermediate_timesteps, + ) + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(0) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + latents = latents * self.scheduler.config.sigma_data + + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]).to(prompt_embeds.dtype) + guidance = guidance * self.transformer.config.guidance_embeds_scale + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + timesteps = timesteps[:-1] + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + latents_model_input = latents / self.scheduler.config.sigma_data + + scm_timestep = torch.sin(timestep) / (torch.cos(timestep) + torch.sin(timestep)) + + scm_timestep_expanded = scm_timestep.view(-1, 1, 1, 1) + latent_model_input = latents_model_input * torch.sqrt( + scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2 + ) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + guidance=guidance, + timestep=scm_timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + )[0] + + noise_pred = ( + (1 - 2 * scm_timestep_expanded) * latent_model_input + + (1 - 2 * scm_timestep_expanded + 2 * scm_timestep_expanded**2) * noise_pred + ) / torch.sqrt(scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2) + noise_pred = noise_pred.float() * self.scheduler.config.sigma_data + + # compute previous image: x_t -> x_t-1 + latents, denoised = self.scheduler.step( + noise_pred, timestep, latents, **extra_step_kwargs, return_dict=False + ) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + latents = denoised / self.scheduler.config.sigma_data + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + try: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return SanaPipelineOutput(images=image) diff --git a/diffusers/pipelines/sana/pipeline_sana_sprint_img2img.py b/diffusers/pipelines/sana/pipeline_sana_sprint_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..e149e0c597b2abd863fea3bd7315efe393cc00f7 --- /dev/null +++ b/diffusers/pipelines/sana/pipeline_sana_sprint_img2img.py @@ -0,0 +1,958 @@ +# Copyright 2025 PixArt-Sigma Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable + +import torch +import torch.nn.functional as F +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, PixArtImageProcessor +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, SanaTransformer2DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ..pipeline_utils import DiffusionPipeline +from ..pixart_alpha.pipeline_pixart_alpha import ASPECT_RATIO_1024_BIN +from .pipeline_output import SanaPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaSprintImg2ImgPipeline + >>> from diffusers.utils.loading_utils import load_image + + >>> pipe = SanaSprintImg2ImgPipeline.from_pretrained( + ... "Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/penguin.png" + ... ) + + + >>> image = pipe(prompt="a cute pink bear", image=image, strength=0.5, height=832, width=480).images[0] + >>> image[0].save("output.png") + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaSprintImg2ImgPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for text-to-image generation using [SANA-Sprint](https://huggingface.co/papers/2503.09641). + """ + + # fmt: off + bad_punct_regex = re.compile( + r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC, + transformer: SanaTransformer2DModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.encoder_block_out_channels) - 1) + if hasattr(self, "vae") and self.vae is not None + else 32 + ) + self.image_processor = PixArtImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.sana.pipeline_sana_sprint.SanaSprintPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + + num_images_per_prompt (`int`, *optional*, defaults to 1): + number of images that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1) + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + prompt, + strength, + height, + width, + num_inference_steps, + timesteps, + max_timesteps, + intermediate_timesteps, + callback_on_step_end_tensor_inputs=None, + prompt_embeds=None, + prompt_attention_mask=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if timesteps is not None and len(timesteps) != num_inference_steps + 1: + raise ValueError("If providing custom timesteps, `timesteps` must be of length `num_inference_steps + 1`.") + + if timesteps is not None and max_timesteps is not None: + raise ValueError("If providing custom timesteps, `max_timesteps` should not be provided.") + + if timesteps is None and max_timesteps is None: + raise ValueError("Should provide either `timesteps` or `max_timesteps`.") + + if intermediate_timesteps is not None and num_inference_steps != 2: + raise ValueError("Intermediate timesteps for SCM is not supported when num_inference_steps != 2.") + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + def prepare_image( + self, + image: PipelineImageInput, + width: int, + height: int, + device: torch.device, + dtype: torch.dtype, + ): + if isinstance(image, torch.Tensor): + if image.ndim == 3: + image = image.unsqueeze(0) + # Resize if current dimensions do not match target dimensions. + if image.shape[2] != height or image.shape[3] != width: + image = F.interpolate(image, size=(height, width), mode="bilinear", align_corners=False) + + image = self.image_processor.preprocess(image, height=height, width=width) + + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image = image.to(device=device, dtype=dtype) + + return image + + def prepare_latents( + self, image, timestep, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + + if image.shape[1] != num_channels_latents: + image = self.vae.encode(image).latent + image_latents = image * self.vae.config.scaling_factor * self.scheduler.config.sigma_data + else: + image_latents = image + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # adapt from https://github.com/huggingface/diffusers/blob/c36f8487df35895421c15f351c7d360bd680[…]/examples/research_projects/sana/train_sana_sprint_diffusers.py + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) * self.scheduler.config.sigma_data + latents = torch.cos(timestep) * image_latents + torch.sin(timestep) * noise + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + num_inference_steps: int = 2, + timesteps: list[int] = None, + max_timesteps: float = 1.57080, + intermediate_timesteps: float = 1.3, + guidance_scale: float = 4.5, + image: PipelineImageInput = None, + strength: float = 0.6, + num_images_per_prompt: int | None = 1, + height: int = 1024, + width: int = 1024, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for image generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, and spatial relationships to create vivid and concrete scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat curled up in a round shape, sleeping peacefully on a warm sunny windowsill, surrounded by pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps, a diverse crowd of people in colorful clothing, and a double-decker bus passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> SanaPipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + max_timesteps (`float`, *optional*, defaults to 1.57080): + The maximum timestep value used in the SCM scheduler. + intermediate_timesteps (`float`, *optional*, defaults to 1.3): + The intermediate timestep value used in SCM scheduler (only used when num_inference_steps=2). + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. Can also + accept image latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.6): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to self.unet.config.sample_size): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size): + The width in pixels of the generated image. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple. + attention_kwargs: + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to + the requested resolution. Useful for generating non-square images. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana.pipeline_output.SanaPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 32: + aspect_ratio_bin = ASPECT_RATIO_1024_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.image_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt=prompt, + strength=strength, + height=height, + width=width, + num_inference_steps=num_inference_steps, + timesteps=timesteps, + max_timesteps=max_timesteps, + intermediate_timesteps=intermediate_timesteps, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + prompt_embeds=prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 2. Preprocess image + init_image = self.prepare_image(image, width, height, device, self.vae.dtype) + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + ) = self.encode_prompt( + prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + + # 5. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + timesteps, + sigmas=None, + max_timesteps=max_timesteps, + intermediate_timesteps=intermediate_timesteps, + ) + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(0) + + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1] + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + init_image, + latent_timestep, + batch_size * num_images_per_prompt, + latent_channels, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # I think this is redundant given the scaling in prepare_latents + # latents = latents * self.scheduler.config.sigma_data + + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]).to(prompt_embeds.dtype) + guidance = guidance * self.transformer.config.guidance_embeds_scale + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + timesteps = timesteps[:-1] + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + latents_model_input = latents / self.scheduler.config.sigma_data + + scm_timestep = torch.sin(timestep) / (torch.cos(timestep) + torch.sin(timestep)) + + scm_timestep_expanded = scm_timestep.view(-1, 1, 1, 1) + latent_model_input = latents_model_input * torch.sqrt( + scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2 + ) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + guidance=guidance, + timestep=scm_timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + )[0] + + noise_pred = ( + (1 - 2 * scm_timestep_expanded) * latent_model_input + + (1 - 2 * scm_timestep_expanded + 2 * scm_timestep_expanded**2) * noise_pred + ) / torch.sqrt(scm_timestep_expanded**2 + (1 - scm_timestep_expanded) ** 2) + noise_pred = noise_pred.float() * self.scheduler.config.sigma_data + + # compute previous image: x_t -> x_t-1 + latents, denoised = self.scheduler.step( + noise_pred, timestep, latents, **extra_step_kwargs, return_dict=False + ) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + latents = denoised / self.scheduler.config.sigma_data + if output_type == "latent": + image = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + try: + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + if use_resolution_binning: + image = self.image_processor.resize_and_crop_tensor(image, orig_width, orig_height) + + if not output_type == "latent": + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return SanaPipelineOutput(images=image) diff --git a/diffusers/pipelines/sana_video/__init__.py b/diffusers/pipelines/sana_video/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..73e224bf749d358e46a2cd4e762c7875b8eeb761 --- /dev/null +++ b/diffusers/pipelines/sana_video/__init__.py @@ -0,0 +1,49 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_sana_video"] = ["SanaVideoPipeline"] + _import_structure["pipeline_sana_video_i2v"] = ["SanaImageToVideoPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_sana_video import SanaVideoPipeline + from .pipeline_sana_video_i2v import SanaImageToVideoPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/sana_video/pipeline_output.py b/diffusers/pipelines/sana_video/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..a625ad3e1d3449b48308bb7ce4a51fcfcc36b23b --- /dev/null +++ b/diffusers/pipelines/sana_video/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from ...utils import BaseOutput + + +@dataclass +class SanaVideoPipelineOutput(BaseOutput): + r""" + Output class for Sana-Video pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/sana_video/pipeline_sana_video.py b/diffusers/pipelines/sana_video/pipeline_sana_video.py new file mode 100644 index 0000000000000000000000000000000000000000..7ae85639e358ad1c720b97dbaa8906a241a2d8b4 --- /dev/null +++ b/diffusers/pipelines/sana_video/pipeline_sana_video.py @@ -0,0 +1,1035 @@ +# Copyright 2025 SANA-Video Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable + +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, AutoencoderKLLTX2Video, AutoencoderKLWan, SanaVideoTransformer3DModel +from ...schedulers import DPMSolverMultistepScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SanaVideoPipelineOutput + + +ASPECT_RATIO_480_BIN = { + "0.5": [448.0, 896.0], + "0.57": [480.0, 832.0], + "0.68": [528.0, 768.0], + "0.78": [560.0, 720.0], + "1.0": [624.0, 624.0], + "1.13": [672.0, 592.0], + "1.29": [720.0, 560.0], + "1.46": [768.0, 528.0], + "1.67": [816.0, 496.0], + "1.75": [832.0, 480.0], + "2.0": [896.0, 448.0], +} + + +ASPECT_RATIO_720_BIN = { + "0.5": [672.0, 1344.0], + "0.57": [704.0, 1280.0], + "0.68": [800.0, 1152.0], + "0.78": [832.0, 1088.0], + "1.0": [960.0, 960.0], + "1.13": [1024.0, 896.0], + "1.29": [1088.0, 832.0], + "1.46": [1152.0, 800.0], + "1.67": [1248.0, 736.0], + "1.75": [1280.0, 704.0], + "2.0": [1344.0, 672.0], +} + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaVideoPipeline + >>> from diffusers.utils import export_to_video + + >>> pipe = SanaVideoPipeline.from_pretrained("Efficient-Large-Model/SANA-Video_2B_480p_diffusers") + >>> pipe.transformer.to(torch.bfloat16) + >>> pipe.text_encoder.to(torch.bfloat16) + >>> pipe.vae.to(torch.float32) + >>> pipe.to("cuda") + >>> motion_score = 30 + + >>> prompt = "Evening, backlight, side lighting, soft light, high contrast, mid-shot, centered composition, clean solo shot, warm color. A young Caucasian man stands in a forest, golden light glimmers on his hair as sunlight filters through the leaves. He wears a light shirt, wind gently blowing his hair and collar, light dances across his face with his movements. The background is blurred, with dappled light and soft tree shadows in the distance. The camera focuses on his lifted gaze, clear and emotional." + >>> negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience." + >>> motion_prompt = f" motion score: {motion_score}." + >>> prompt = prompt + motion_prompt + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=480, + ... width=832, + ... frames=81, + ... guidance_scale=6, + ... num_inference_steps=50, + ... generator=torch.Generator(device="cuda").manual_seed(42), + ... ).frames[0] + + >>> export_to_video(output, "sana-video-output.mp4", fps=16) + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class SanaVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for text-to-video generation using [Sana](https://huggingface.co/papers/2509.24695). This model inherits + from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods implemented for all + pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`GemmaTokenizer`] or [`GemmaTokenizerFast`]): + The tokenizer used to tokenize the prompt. + text_encoder ([`Gemma2PreTrainedModel`]): + Text encoder model to encode the input prompts. + vae ([`AutoencoderKLWan`, `AutoencoderDC`, or `AutoencoderKLLTX2Video`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + transformer ([`SanaVideoTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`DPMSolverMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC | AutoencoderKLLTX2Video | AutoencoderKLWan, + transformer: SanaVideoTransformer3DModel, + scheduler: DPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + if getattr(self, "vae", None): + if isinstance(self.vae, AutoencoderKLLTX2Video): + self.vae_scale_factor_temporal = self.vae.config.temporal_compression_ratio + self.vae_scale_factor_spatial = self.vae.config.spatial_compression_ratio + elif isinstance(self.vae, (AutoencoderDC, AutoencoderKLWan)): + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial + else: + self.vae_scale_factor_temporal = 4 + self.vae_scale_factor_spatial = 8 + else: + self.vae_scale_factor_temporal = 4 + self.vae_scale_factor_spatial = 8 + + self.vae_scale_factor = self.vae_scale_factor_spatial + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the video generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_videos_per_prompt (`int`, *optional*, defaults to 1): + number of videos that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=False, + ) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_videos_per_prompt, 1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_on_step_end_tensor_inputs=None, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + height: int = 480, + width: int = 832, + frames: int = 81, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for video generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, motion, and temporal relationships to create vivid and dynamic scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat slowly settling into a curled position, peacefully falling asleep on a warm sunny windowsill, with gentle sunlight filtering through surrounding pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps gradually lighting up, a diverse crowd of people in colorful clothing walking past, and a double-decker bus smoothly passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> SanaVideoPipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the video generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate videos that are closely linked to + the text `prompt`, usually at the expense of lower video quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + height (`int`, *optional*, defaults to 480): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to 832): + The width in pixels of the generated video. + frames (`int`, *optional*, defaults to 81): + The number of frames in the generated video. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated video. Choose between mp4 or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`SanaVideoPipelineOutput`] instead of a plain tuple. + attention_kwargs: + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_480_BIN` or `ASPECT_RATIO_720_BIN`. After the produced latents are decoded into videos, + they are resized back to the requested resolution. Useful for generating non-square videos. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana_video.pipeline_output.SanaVideoPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana_video.pipeline_output.SanaVideoPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated videos + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 30: + aspect_ratio_bin = ASPECT_RATIO_480_BIN + elif self.transformer.config.sample_size == 22: + aspect_ratio_bin = ASPECT_RATIO_720_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.video_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + height, + width, + callback_on_step_end_tensor_inputs, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + latent_channels, + height, + width, + frames, + torch.float32, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + + # compute previous image: x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + video = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + if isinstance(self.vae, AutoencoderKLLTX2Video): + latents_mean = self.vae.latents_mean + latents_std = self.vae.latents_std + z_dim = self.vae.config.latent_channels + elif isinstance(self.vae, AutoencoderKLWan): + latents_mean = torch.tensor(self.vae.config.latents_mean) + latents_std = torch.tensor(self.vae.config.latents_std) + z_dim = self.vae.config.z_dim + else: + latents_mean = torch.zeros(latents.shape[1], device=latents.device, dtype=latents.dtype) + latents_std = torch.ones(latents.shape[1], device=latents.device, dtype=latents.dtype) + z_dim = latents.shape[1] + + latents_mean = latents_mean.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = 1.0 / latents_std.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents / latents_std + latents_mean + try: + video = self.vae.decode(latents, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + + if use_resolution_binning: + video = self.video_processor.resize_and_crop_tensor(video, orig_width, orig_height) + + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SanaVideoPipelineOutput(frames=video) diff --git a/diffusers/pipelines/sana_video/pipeline_sana_video_i2v.py b/diffusers/pipelines/sana_video/pipeline_sana_video_i2v.py new file mode 100644 index 0000000000000000000000000000000000000000..81df1d0759da923bbfb9ace3f8956507079a7c74 --- /dev/null +++ b/diffusers/pipelines/sana_video/pipeline_sana_video_i2v.py @@ -0,0 +1,1088 @@ +# Copyright 2025 SANA-Video Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import re +import urllib.parse as ul +import warnings +from typing import Any, Callable + +import PIL +import torch +from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFast + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import SanaLoraLoaderMixin +from ...models import AutoencoderDC, AutoencoderKLLTX2Video, AutoencoderKLWan, SanaVideoTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + BACKENDS_MAPPING, + USE_PEFT_BACKEND, + is_bs4_available, + is_ftfy_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import get_device, is_torch_version, randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SanaVideoPipelineOutput +from .pipeline_sana_video import ASPECT_RATIO_480_BIN, ASPECT_RATIO_720_BIN + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_bs4_available(): + from bs4 import BeautifulSoup + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import SanaImageToVideoPipeline + >>> from diffusers.utils import export_to_video, load_image + + >>> pipe = SanaImageToVideoPipeline.from_pretrained("Efficient-Large-Model/SANA-Video_2B_480p_diffusers") + >>> pipe.transformer.to(torch.bfloat16) + >>> pipe.text_encoder.to(torch.bfloat16) + >>> pipe.vae.to(torch.float32) + >>> pipe.to("cuda") + >>> motion_score = 30 + + >>> prompt = "A woman stands against a stunning sunset backdrop, her long, wavy brown hair gently blowing in the breeze. She wears a sleeveless, light-colored blouse with a deep V-neckline, which accentuates her graceful posture. The warm hues of the setting sun cast a golden glow across her face and hair, creating a serene and ethereal atmosphere. The background features a blurred landscape with soft, rolling hills and scattered clouds, adding depth to the scene. The camera remains steady, capturing the tranquil moment from a medium close-up angle." + >>> negative_prompt = "A chaotic sequence with misshapen, deformed limbs in heavy motion blur, sudden disappearance, jump cuts, jerky movements, rapid shot changes, frames out of sync, inconsistent character shapes, temporal artifacts, jitter, and ghosting effects, creating a disorienting visual experience." + >>> motion_prompt = f" motion score: {motion_score}." + >>> prompt = prompt + motion_prompt + >>> image = load_image("https://raw.githubusercontent.com/NVlabs/Sana/refs/heads/main/asset/samples/i2v-1.png") + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=480, + ... width=832, + ... frames=81, + ... guidance_scale=6, + ... num_inference_steps=50, + ... generator=torch.Generator(device="cuda").manual_seed(42), + ... ).frames[0] + + >>> export_to_video(output, "sana-ti2v-output.mp4", fps=16) + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class SanaImageToVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin): + r""" + Pipeline for image/text-to-video generation using [Sana](https://huggingface.co/papers/2509.24695). This model + inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods implemented for all + pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`GemmaTokenizer`] or [`GemmaTokenizerFast`]): + The tokenizer used to tokenize the prompt. + text_encoder ([`Gemma2PreTrainedModel`]): + Text encoder model to encode the input prompts. + vae ([`AutoencoderKLWan`, `AutoencoderDC`, or `AutoencoderKLLTX2Video`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + transformer ([`SanaVideoTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded video latents. + """ + + # fmt: off + bad_punct_regex = re.compile(r"[" + "#®•©™&@·º½¾¿¡§~" + r"\)" + r"\(" + r"\]" + r"\[" + r"\}" + r"\{" + r"\|" + "\\" + r"\/" + r"\*" + r"]{1,}") + # fmt: on + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: GemmaTokenizer | GemmaTokenizerFast, + text_encoder: Gemma2PreTrainedModel, + vae: AutoencoderDC | AutoencoderKLLTX2Video | AutoencoderKLWan, + transformer: SanaVideoTransformer3DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler + ) + + if getattr(self, "vae", None): + if isinstance(self.vae, AutoencoderKLLTX2Video): + self.vae_scale_factor_temporal = self.vae.config.temporal_compression_ratio + self.vae_scale_factor_spatial = self.vae.config.spatial_compression_ratio + elif isinstance(self.vae, (AutoencoderDC, AutoencoderKLWan)): + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial + else: + self.vae_scale_factor_temporal = 4 + self.vae_scale_factor_spatial = 8 + else: + self.vae_scale_factor_temporal = 4 + self.vae_scale_factor_spatial = 8 + + self.vae_scale_factor = self.vae_scale_factor_spatial + + self.transformer_spatial_patch_size = ( + self.transformer.config.patch_size[1] if getattr(self, "transformer", None) is not None else 1 + ) + self.transformer_temporal_patch_size = ( + self.transformer.config.patch_size[0] if getattr(self, "transformer") is not None else 1 + ) + + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline._get_gemma_prompt_embeds + def _get_gemma_prompt_embeds( + self, + prompt: str | list[str], + device: torch.device, + dtype: torch.dtype, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + prompt = [prompt] if isinstance(prompt, str) else prompt + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + prompt = self._text_preprocessing(prompt, clean_caption=clean_caption) + + # prepare complex human instruction + if not complex_human_instruction: + max_length_all = max_sequence_length + else: + chi_prompt = "\n".join(complex_human_instruction) + prompt = [chi_prompt + p for p in prompt] + num_chi_prompt_tokens = len(self.tokenizer.encode(chi_prompt)) + max_length_all = num_chi_prompt_tokens + max_sequence_length - 2 + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_length_all, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + prompt_attention_mask = text_inputs.attention_mask + prompt_attention_mask = prompt_attention_mask.to(device) + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask) + prompt_embeds = prompt_embeds[0].to(dtype=dtype, device=device) + + return prompt_embeds, prompt_attention_mask + + # Copied from diffusers.pipelines.sana_video.pipeline_sana_video.SanaVideoPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + do_classifier_free_guidance: bool = True, + negative_prompt: str = "", + num_videos_per_prompt: int = 1, + device: torch.device | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + clean_caption: bool = False, + max_sequence_length: int = 300, + complex_human_instruction: list[str] | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt not to guide the video generation. If not defined, one has to pass `negative_prompt_embeds` + instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For + PixArt-Alpha, this should be "". + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + whether to use classifier free guidance or not + num_videos_per_prompt (`int`, *optional*, defaults to 1): + number of videos that should be generated per prompt + device: (`torch.device`, *optional*): + torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For Sana, it's should be the embeddings of the "" string. + clean_caption (`bool`, defaults to `False`): + If `True`, the function will preprocess and clean the provided caption before encoding. + max_sequence_length (`int`, defaults to 300): Maximum sequence length to use for the prompt. + complex_human_instruction (`list[str]`, defaults to `complex_human_instruction`): + If `complex_human_instruction` is not empty, the function will use the complex Human instruction for + the prompt. + """ + + if device is None: + device = self._execution_device + + if self.text_encoder is not None: + dtype = self.text_encoder.dtype + else: + dtype = None + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if getattr(self, "tokenizer", None) is not None: + self.tokenizer.padding_side = "right" + + # See Section 3.1. of the paper. + max_length = max_sequence_length + select_index = [0] + list(range(-max_length + 1, 0)) + + if prompt_embeds is None: + prompt_embeds, prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + ) + + prompt_embeds = prompt_embeds[:, select_index] + prompt_attention_mask = prompt_attention_mask[:, select_index] + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1) + prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1) + prompt_attention_mask = prompt_attention_mask.repeat(num_videos_per_prompt, 1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = [negative_prompt] * batch_size if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_embeds, negative_prompt_attention_mask = self._get_gemma_prompt_embeds( + prompt=negative_prompt, + device=device, + dtype=dtype, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=False, + ) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1) + negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_videos_per_prompt, 1) + else: + negative_prompt_embeds = None + negative_prompt_attention_mask = None + + if self.text_encoder is not None: + if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + image, + height, + width, + callback_on_step_end_tensor_inputs=None, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + prompt_attention_mask=None, + negative_prompt_attention_mask=None, + ): + if height % 32 != 0 or width % 32 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 32 but are {height} and {width}.") + + if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and prompt_attention_mask is None: + raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.") + + if negative_prompt_embeds is not None and negative_prompt_attention_mask is None: + raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.") + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if prompt_attention_mask.shape != negative_prompt_attention_mask.shape: + raise ValueError( + "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but" + f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`" + f" {negative_prompt_attention_mask.shape}." + ) + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing + def _text_preprocessing(self, text, clean_caption=False): + if clean_caption and not is_bs4_available(): + logger.warning(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if clean_caption and not is_ftfy_available(): + logger.warning(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`")) + logger.warning("Setting `clean_caption` to False...") + clean_caption = False + + if not isinstance(text, (tuple, list)): + text = [text] + + def process(text: str): + if clean_caption: + text = self._clean_caption(text) + text = self._clean_caption(text) + else: + text = text.lower().strip() + return text + + return [process(t) for t in text] + + # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption + def _clean_caption(self, caption): + caption = str(caption) + caption = ul.unquote_plus(caption) + caption = caption.strip().lower() + caption = re.sub("", "person", caption) + # urls: + caption = re.sub( + r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + caption = re.sub( + r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa + "", + caption, + ) # regex for urls + # html: + caption = BeautifulSoup(caption, features="html.parser").text + + # @ + caption = re.sub(r"@[\w\d]+\b", "", caption) + + # 31C0—31EF CJK Strokes + # 31F0—31FF Katakana Phonetic Extensions + # 3200—32FF Enclosed CJK Letters and Months + # 3300—33FF CJK Compatibility + # 3400—4DBF CJK Unified Ideographs Extension A + # 4DC0—4DFF Yijing Hexagram Symbols + # 4E00—9FFF CJK Unified Ideographs + caption = re.sub(r"[\u31c0-\u31ef]+", "", caption) + caption = re.sub(r"[\u31f0-\u31ff]+", "", caption) + caption = re.sub(r"[\u3200-\u32ff]+", "", caption) + caption = re.sub(r"[\u3300-\u33ff]+", "", caption) + caption = re.sub(r"[\u3400-\u4dbf]+", "", caption) + caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption) + caption = re.sub(r"[\u4e00-\u9fff]+", "", caption) + ####################################################### + + # все виды тире / all types of dash --> "-" + caption = re.sub( + r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa + "-", + caption, + ) + + # кавычки к одному стандарту + caption = re.sub(r"[`´«»“”¨]", '"', caption) + caption = re.sub(r"[‘’]", "'", caption) + + # " + caption = re.sub(r""?", "", caption) + # & + caption = re.sub(r"&", "", caption) + + # ip addresses: + caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption) + + # article ids: + caption = re.sub(r"\d:\d\d\s+$", "", caption) + + # \n + caption = re.sub(r"\\n", " ", caption) + + # "#123" + caption = re.sub(r"#\d{1,3}\b", "", caption) + # "#12345.." + caption = re.sub(r"#\d{5,}\b", "", caption) + # "123456.." + caption = re.sub(r"\b\d{6,}\b", "", caption) + # filenames: + caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption) + + # + caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT""" + caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT""" + + caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT + caption = re.sub(r"\s+\.\s+", r" ", caption) # " . " + + # this-is-my-cute-cat / this_is_my_cute_cat + regex2 = re.compile(r"(?:\-|\_)") + if len(re.findall(regex2, caption)) > 3: + caption = re.sub(regex2, " ", caption) + + caption = ftfy.fix_text(caption) + caption = html.unescape(html.unescape(caption)) + + caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640 + caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc + caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231 + + caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption) + caption = re.sub(r"(free\s)?download(\sfree)?", "", caption) + caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption) + caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption) + caption = re.sub(r"\bpage\s+\d+\b", "", caption) + + caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a... + + caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption) + + caption = re.sub(r"\b\s+\:\s+", r": ", caption) + caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption) + caption = re.sub(r"\s+", " ", caption) + + caption.strip() + + caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption) + caption = re.sub(r"^[\'\_,\-\:;]", r"", caption) + caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption) + caption = re.sub(r"^\.\S+$", "", caption) + + return caption.strip() + + def prepare_latents( + self, + image: PipelineImageInput, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + image = image.unsqueeze(2) # [B, C, 1, H, W] + image = image.to(device=device, dtype=self.vae.dtype) + + if isinstance(generator, list): + image_latents = [retrieve_latents(self.vae.encode(image), sample_mode="argmax") for _ in generator] + image_latents = torch.cat(image_latents) + else: + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + image_latents = image_latents.repeat(batch_size, 1, 1, 1, 1) + + if isinstance(self.vae, AutoencoderKLLTX2Video): + _latents_mean = self.vae.latents_mean + _latents_std = self.vae.latents_std + elif isinstance(self.vae, AutoencoderKLWan): + _latents_mean = torch.tensor(self.vae.config.latents_mean) + _latents_std = torch.tensor(self.vae.config.latents_std) + else: + _latents_mean = torch.zeros(image_latents.shape[1], device=image_latents.device, dtype=image_latents.dtype) + _latents_std = torch.ones(image_latents.shape[1], device=image_latents.device, dtype=image_latents.dtype) + + latents_mean = _latents_mean.view(1, -1, 1, 1, 1).to(image_latents.device, image_latents.dtype) + latents_std = 1.0 / _latents_std.view(1, -1, 1, 1, 1).to(image_latents.device, image_latents.dtype) + image_latents = (image_latents - latents_mean) * latents_std + + latents[:, :, 0:1] = image_latents.to(dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + negative_prompt: str = "", + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + height: int = 480, + width: int = 832, + frames: int = 81, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_attention_mask: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_attention_mask: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + clean_caption: bool = False, + use_resolution_binning: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 300, + complex_human_instruction: list[str] = [ + "Given a user prompt, generate an 'Enhanced prompt' that provides detailed visual descriptions suitable for video generation. Evaluate the level of detail in the user prompt:", + "- If the prompt is simple, focus on adding specifics about colors, shapes, sizes, textures, motion, and temporal relationships to create vivid and dynamic scenes.", + "- If the prompt is already detailed, refine and enhance the existing details slightly without overcomplicating.", + "Here are examples of how to transform or refine prompts:", + "- User Prompt: A cat sleeping -> Enhanced: A small, fluffy white cat slowly settling into a curled position, peacefully falling asleep on a warm sunny windowsill, with gentle sunlight filtering through surrounding pots of blooming red flowers.", + "- User Prompt: A busy city street -> Enhanced: A bustling city street scene at dusk, featuring glowing street lamps gradually lighting up, a diverse crowd of people in colorful clothing walking past, and a double-decker bus smoothly passing by towering glass skyscrapers.", + "Please generate only the enhanced description for the prompt below and avoid including any additional commentary or evaluations:", + "User Prompt: ", + ], + ) -> SanaVideoPipelineOutput | tuple: + """ + Function invoked when calling the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the video generation on. The first frame of the generated video will be + conditioned on this image. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the video generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 4.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate videos that are closely linked to + the text `prompt`, usually at the expense of lower video quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + height (`int`, *optional*, defaults to 480): + The height in pixels of the generated video. + width (`int`, *optional*, defaults to 832): + The width in pixels of the generated video. + frames (`int`, *optional*, defaults to 81): + The number of frames in the generated video. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. For PixArt-Sigma this negative prompt should be "". If not + provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. + negative_prompt_attention_mask (`torch.Tensor`, *optional*): + Pre-generated attention mask for negative text embeddings. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated video. Choose between mp4 or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`SanaVideoPipelineOutput`] instead of a plain tuple. + attention_kwargs: + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clean_caption (`bool`, *optional*, defaults to `True`): + Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to + be installed. If the dependencies are not installed, the embeddings will be created from the raw + prompt. + use_resolution_binning (`bool` defaults to `True`): + If set to `True`, the requested height and width are first mapped to the closest resolutions using + `ASPECT_RATIO_480_BIN` or `ASPECT_RATIO_720_BIN`. After the produced latents are decoded into videos, + they are resized back to the requested resolution. Useful for generating non-square videos. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to `300`): + Maximum sequence length to use with the `prompt`. + complex_human_instruction (`list[str]`, *optional*): + Instructions for complex human attention: + https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55. + + Examples: + + Returns: + [`~pipelines.sana_video.pipeline_output.SanaVideoPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.sana_video.pipeline_output.SanaVideoPipelineOutput`] is + returned, otherwise a `tuple` is returned where the first element is a list with the generated videos + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + if use_resolution_binning: + if self.transformer.config.sample_size == 30: + aspect_ratio_bin = ASPECT_RATIO_480_BIN + elif self.transformer.config.sample_size == 22: + aspect_ratio_bin = ASPECT_RATIO_720_BIN + else: + raise ValueError("Invalid sample size") + orig_height, orig_width = height, width + height, width = self.video_processor.classify_height_width_bin(height, width, ratios=aspect_ratio_bin) + + self.check_inputs( + prompt, + image, + height, + width, + callback_on_step_end_tensor_inputs, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + prompt_attention_mask, + negative_prompt_attention_mask, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._interrupt = False + + # 2. Default height and width to transformer + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None + + # 3. Encode input prompt + ( + prompt_embeds, + prompt_attention_mask, + negative_prompt_embeds, + negative_prompt_attention_mask, + ) = self.encode_prompt( + prompt, + self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + device=device, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + prompt_attention_mask=prompt_attention_mask, + negative_prompt_attention_mask=negative_prompt_attention_mask, + clean_caption=clean_caption, + max_sequence_length=max_sequence_length, + complex_human_instruction=complex_human_instruction, + lora_scale=lora_scale, + ) + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, device, timesteps, sigmas + ) + + # 5. Prepare latents. + latent_channels = self.transformer.config.in_channels + image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.float32) + + latents = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + latent_channels, + height, + width, + frames, + torch.float32, + device, + generator, + latents, + ) + + conditioning_mask = latents.new_zeros( + batch_size, + 1, + latents.shape[2] // self.transformer_temporal_patch_size, + latents.shape[3] // self.transformer_spatial_patch_size, + latents.shape[4] // self.transformer_spatial_patch_size, + ) + conditioning_mask[:, :, 0] = 1.0 + if self.do_classifier_free_guidance: + conditioning_mask = torch.cat([conditioning_mask, conditioning_mask]) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + transformer_dtype = self.transformer.dtype + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(conditioning_mask.shape) + timestep = timestep * (1 - conditioning_mask) + + # predict noise model_output + noise_pred = self.transformer( + latent_model_input.to(dtype=transformer_dtype), + encoder_hidden_states=prompt_embeds.to(dtype=transformer_dtype), + encoder_attention_mask=prompt_attention_mask, + timestep=timestep, + return_dict=False, + attention_kwargs=self.attention_kwargs, + )[0] + noise_pred = noise_pred.float() + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + timestep, _ = timestep.chunk(2) + + # learned sigma + if self.transformer.config.out_channels // 2 == latent_channels: + noise_pred = noise_pred.chunk(2, dim=1)[0] + + noise_pred = noise_pred[:, :, 1:] + noise_latents = latents[:, :, 1:] + pred_latents = self.scheduler.step( + noise_pred, t, noise_latents, **extra_step_kwargs, return_dict=False + )[0] + + latents = torch.cat([latents[:, :, :1], pred_latents], dim=2) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + video = latents + else: + latents = latents.to(self.vae.dtype) + torch_accelerator_module = getattr(torch, get_device(), torch.cuda) + oom_error = ( + torch.OutOfMemoryError + if is_torch_version(">=", "2.5.0") + else torch_accelerator_module.OutOfMemoryError + ) + if isinstance(self.vae, AutoencoderKLLTX2Video): + latents_mean = self.vae.latents_mean + latents_std = self.vae.latents_std + z_dim = self.vae.config.latent_channels + elif isinstance(self.vae, AutoencoderKLWan): + latents_mean = torch.tensor(self.vae.config.latents_mean) + latents_std = torch.tensor(self.vae.config.latents_std) + z_dim = self.vae.config.z_dim + else: + latents_mean = torch.zeros(latents.shape[1], device=latents.device, dtype=latents.dtype) + latents_std = torch.ones(latents.shape[1], device=latents.device, dtype=latents.dtype) + z_dim = latents.shape[1] + + latents_mean = latents_mean.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype) + latents_std = 1.0 / latents_std.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype) + latents = latents / latents_std + latents_mean + try: + video = self.vae.decode(latents, return_dict=False)[0] + except oom_error as e: + warnings.warn( + f"{e}. \n" + f"Try to use VAE tiling for large images. For example: \n" + f"pipe.vae.enable_tiling(tile_sample_min_width=512, tile_sample_min_height=512)" + ) + + if use_resolution_binning: + video = self.video_processor.resize_and_crop_tensor(video, orig_width, orig_height) + + video = self.video_processor.postprocess_video(video, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SanaVideoPipelineOutput(frames=video) diff --git a/diffusers/pipelines/shap_e/__init__.py b/diffusers/pipelines/shap_e/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..4ed563c4a51f6e627c06711b60fe3a0709ff22f7 --- /dev/null +++ b/diffusers/pipelines/shap_e/__init__.py @@ -0,0 +1,71 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["camera"] = ["create_pan_cameras"] + _import_structure["pipeline_shap_e"] = ["ShapEPipeline"] + _import_structure["pipeline_shap_e_img2img"] = ["ShapEImg2ImgPipeline"] + _import_structure["renderer"] = [ + "BoundingBoxVolume", + "ImportanceRaySampler", + "MLPNeRFModelOutput", + "MLPNeRSTFModel", + "ShapEParamsProjModel", + "ShapERenderer", + "StratifiedRaySampler", + "VoidNeRFModel", + ] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .camera import create_pan_cameras + from .pipeline_shap_e import ShapEPipeline + from .pipeline_shap_e_img2img import ShapEImg2ImgPipeline + from .renderer import ( + BoundingBoxVolume, + ImportanceRaySampler, + MLPNeRFModelOutput, + MLPNeRSTFModel, + ShapEParamsProjModel, + ShapERenderer, + StratifiedRaySampler, + VoidNeRFModel, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/shap_e/camera.py b/diffusers/pipelines/shap_e/camera.py new file mode 100644 index 0000000000000000000000000000000000000000..81807b6ff35a9e3abf400fd7cddfc710244f367b --- /dev/null +++ b/diffusers/pipelines/shap_e/camera.py @@ -0,0 +1,146 @@ +# Copyright 2025 Open AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import torch + + +@dataclass +class DifferentiableProjectiveCamera: + """ + Implements a batch, differentiable, standard pinhole camera + """ + + origin: torch.Tensor # [batch_size x 3] + x: torch.Tensor # [batch_size x 3] + y: torch.Tensor # [batch_size x 3] + z: torch.Tensor # [batch_size x 3] + width: int + height: int + x_fov: float + y_fov: float + shape: tuple[int] + + def __post_init__(self): + assert self.x.shape[0] == self.y.shape[0] == self.z.shape[0] == self.origin.shape[0] + assert self.x.shape[1] == self.y.shape[1] == self.z.shape[1] == self.origin.shape[1] == 3 + assert len(self.x.shape) == len(self.y.shape) == len(self.z.shape) == len(self.origin.shape) == 2 + + def resolution(self): + return torch.from_numpy(np.array([self.width, self.height], dtype=np.float32)) + + def fov(self): + return torch.from_numpy(np.array([self.x_fov, self.y_fov], dtype=np.float32)) + + def get_image_coords(self) -> torch.Tensor: + """ + :return: coords of shape (width * height, 2) + """ + pixel_indices = torch.arange(self.height * self.width) + coords = torch.stack( + [ + pixel_indices % self.width, + torch.div(pixel_indices, self.width, rounding_mode="trunc"), + ], + axis=1, + ) + return coords + + @property + def camera_rays(self): + batch_size, *inner_shape = self.shape + inner_batch_size = int(np.prod(inner_shape)) + + coords = self.get_image_coords() + coords = torch.broadcast_to(coords.unsqueeze(0), [batch_size * inner_batch_size, *coords.shape]) + rays = self.get_camera_rays(coords) + + rays = rays.view(batch_size, inner_batch_size * self.height * self.width, 2, 3) + + return rays + + def get_camera_rays(self, coords: torch.Tensor) -> torch.Tensor: + batch_size, *shape, n_coords = coords.shape + assert n_coords == 2 + assert batch_size == self.origin.shape[0] + + flat = coords.view(batch_size, -1, 2) + + res = self.resolution() + fov = self.fov() + + fracs = (flat.float() / (res - 1)) * 2 - 1 + fracs = fracs * torch.tan(fov / 2) + + fracs = fracs.view(batch_size, -1, 2) + directions = ( + self.z.view(batch_size, 1, 3) + + self.x.view(batch_size, 1, 3) * fracs[:, :, :1] + + self.y.view(batch_size, 1, 3) * fracs[:, :, 1:] + ) + directions = directions / directions.norm(dim=-1, keepdim=True) + rays = torch.stack( + [ + torch.broadcast_to(self.origin.view(batch_size, 1, 3), [batch_size, directions.shape[1], 3]), + directions, + ], + dim=2, + ) + return rays.view(batch_size, *shape, 2, 3) + + def resize_image(self, width: int, height: int) -> "DifferentiableProjectiveCamera": + """ + Creates a new camera for the resized view assuming the aspect ratio does not change. + """ + assert width * self.height == height * self.width, "The aspect ratio should not change." + return DifferentiableProjectiveCamera( + origin=self.origin, + x=self.x, + y=self.y, + z=self.z, + width=width, + height=height, + x_fov=self.x_fov, + y_fov=self.y_fov, + ) + + +def create_pan_cameras(size: int) -> DifferentiableProjectiveCamera: + origins = [] + xs = [] + ys = [] + zs = [] + for theta in np.linspace(0, 2 * np.pi, num=20): + z = np.array([np.sin(theta), np.cos(theta), -0.5]) + z /= np.sqrt(np.sum(z**2)) + origin = -z * 4 + x = np.array([np.cos(theta), -np.sin(theta), 0.0]) + y = np.cross(z, x) + origins.append(origin) + xs.append(x) + ys.append(y) + zs.append(z) + return DifferentiableProjectiveCamera( + origin=torch.from_numpy(np.stack(origins, axis=0)).float(), + x=torch.from_numpy(np.stack(xs, axis=0)).float(), + y=torch.from_numpy(np.stack(ys, axis=0)).float(), + z=torch.from_numpy(np.stack(zs, axis=0)).float(), + width=size, + height=size, + x_fov=0.7, + y_fov=0.7, + shape=(1, len(xs)), + ) diff --git a/diffusers/pipelines/shap_e/pipeline_shap_e.py b/diffusers/pipelines/shap_e/pipeline_shap_e.py new file mode 100644 index 0000000000000000000000000000000000000000..eea83aff9e1098dc083a0a4a656e77bb503b2f57 --- /dev/null +++ b/diffusers/pipelines/shap_e/pipeline_shap_e.py @@ -0,0 +1,345 @@ +# Copyright 2025 Open AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPTextModelWithProjection, CLIPTokenizer + +from ...models import PriorTransformer +from ...schedulers import HeunDiscreteScheduler +from ...utils import ( + BaseOutput, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .renderer import ShapERenderer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import DiffusionPipeline + >>> from diffusers.utils import export_to_gif + + >>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + >>> repo = "openai/shap-e" + >>> pipe = DiffusionPipeline.from_pretrained(repo, torch_dtype=torch.float16) + >>> pipe = pipe.to(device) + + >>> guidance_scale = 15.0 + >>> prompt = "a shark" + + >>> images = pipe( + ... prompt, + ... guidance_scale=guidance_scale, + ... num_inference_steps=64, + ... frame_size=256, + ... ).images + + >>> gif_path = export_to_gif(images[0], "shark_3d.gif") + ``` +""" + + +@dataclass +class ShapEPipelineOutput(BaseOutput): + """ + Output class for [`ShapEPipeline`] and [`ShapEImg2ImgPipeline`]. + + Args: + images (`torch.Tensor`) + A list of images for 3D rendering. + """ + + images: list[list[PIL.Image.Image]] | list[list[np.ndarray]] + + +class ShapEPipeline(DiffusionPipeline): + """ + Pipeline for generating latent representation of a 3D asset and rendering with the NeRF method. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + text_encoder ([`~transformers.CLIPTextModelWithProjection`]): + Frozen text-encoder. + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + scheduler ([`HeunDiscreteScheduler`]): + A scheduler to be used in combination with the `prior` model to generate image embedding. + shap_e_renderer ([`ShapERenderer`]): + Shap-E renderer projects the generated latents into parameters of a MLP to create 3D objects with the NeRF + rendering method. + """ + + model_cpu_offload_seq = "text_encoder->prior" + _exclude_from_cpu_offload = ["shap_e_renderer"] + + def __init__( + self, + prior: PriorTransformer, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + scheduler: HeunDiscreteScheduler, + shap_e_renderer: ShapERenderer, + ): + super().__init__() + + self.register_modules( + prior=prior, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + shap_e_renderer=shap_e_renderer, + ) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + ): + len(prompt) if isinstance(prompt, list) else 1 + + # YiYi Notes: set pad_token_id to be 0, not sure why I can't set in the config file + self.tokenizer.pad_token_id = 0 + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_encoder_output = self.text_encoder(text_input_ids.to(device)) + prompt_embeds = text_encoder_output.text_embeds + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + # in Shap-E it normalize the prompt_embeds and then later rescale it + prompt_embeds = prompt_embeds / torch.linalg.norm(prompt_embeds, dim=-1, keepdim=True) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # Rescale the features to have unit variance + prompt_embeds = math.sqrt(prompt_embeds.shape[1]) * prompt_embeds + + return prompt_embeds + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str, + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + guidance_scale: float = 4.0, + frame_size: int = 64, + output_type: str | None = "pil", # pil, np, latent, mesh + return_dict: bool = True, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + frame_size (`int`, *optional*, default to 64): + The width and height of each image frame of the generated 3D output. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`), `"latent"` (`torch.Tensor`), or mesh ([`MeshDecoderOutput`]). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] instead of a plain + tuple. + + Examples: + + Returns: + [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + + if isinstance(prompt, str): + batch_size = 1 + elif isinstance(prompt, list): + batch_size = len(prompt) + else: + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = guidance_scale > 1.0 + prompt_embeds = self._encode_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance) + + # prior + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + num_embeddings = self.prior.config.num_embeddings + embedding_dim = self.prior.config.embedding_dim + + latents = self.prepare_latents( + (batch_size, num_embeddings * embedding_dim), + prompt_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + + # YiYi notes: for testing only to match ldm, we can directly create a latents with desired shape: batch_size, num_embeddings, embedding_dim + latents = latents.reshape(latents.shape[0], num_embeddings, embedding_dim) + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + noise_pred = self.prior( + scaled_model_input, + timestep=t, + proj_embedding=prompt_embeds, + ).predicted_image_embedding + + # remove the variance + noise_pred, _ = noise_pred.split( + scaled_model_input.shape[2], dim=2 + ) # batch_size, num_embeddings, embedding_dim + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred - noise_pred_uncond) + + latents = self.scheduler.step( + noise_pred, + timestep=t, + sample=latents, + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + # Offload all models + self.maybe_free_model_hooks() + + if output_type not in ["np", "pil", "latent", "mesh"]: + raise ValueError( + f"Only the output types `pil`, `np`, `latent` and `mesh` are supported not output_type={output_type}" + ) + + if output_type == "latent": + return ShapEPipelineOutput(images=latents) + + images = [] + if output_type == "mesh": + for i, latent in enumerate(latents): + mesh = self.shap_e_renderer.decode_to_mesh( + latent[None, :], + device, + ) + images.append(mesh) + + else: + # np, pil + for i, latent in enumerate(latents): + image = self.shap_e_renderer.decode_to_image( + latent[None, :], + device, + size=frame_size, + ) + images.append(image) + + images = torch.stack(images) + + images = images.cpu().numpy() + + if output_type == "pil": + images = [self.numpy_to_pil(image) for image in images] + + if not return_dict: + return (images,) + + return ShapEPipelineOutput(images=images) diff --git a/diffusers/pipelines/shap_e/pipeline_shap_e_img2img.py b/diffusers/pipelines/shap_e/pipeline_shap_e_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..f59fd298c68447e4b2453c608d5eaf47e8354ce9 --- /dev/null +++ b/diffusers/pipelines/shap_e/pipeline_shap_e_img2img.py @@ -0,0 +1,332 @@ +# Copyright 2025 Open AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPVisionModel + +from ...models import PriorTransformer +from ...schedulers import HeunDiscreteScheduler +from ...utils import ( + BaseOutput, + is_torch_xla_available, + logging, + replace_example_docstring, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .renderer import ShapERenderer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from PIL import Image + >>> import torch + >>> from diffusers import DiffusionPipeline + >>> from diffusers.utils import export_to_gif, load_image + + >>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu") + + >>> repo = "openai/shap-e-img2img" + >>> pipe = DiffusionPipeline.from_pretrained(repo, torch_dtype=torch.float16) + >>> pipe = pipe.to(device) + + >>> guidance_scale = 3.0 + >>> image_url = "https://hf.co/datasets/diffusers/docs-images/resolve/main/shap-e/corgi.png" + >>> image = load_image(image_url).convert("RGB") + + >>> images = pipe( + ... image, + ... guidance_scale=guidance_scale, + ... num_inference_steps=64, + ... frame_size=256, + ... ).images + + >>> gif_path = export_to_gif(images[0], "corgi_3d.gif") + ``` +""" + + +@dataclass +class ShapEPipelineOutput(BaseOutput): + """ + Output class for [`ShapEPipeline`] and [`ShapEImg2ImgPipeline`]. + + Args: + images (`torch.Tensor`) + A list of images for 3D rendering. + """ + + images: PIL.Image.Image | np.ndarray + + +class ShapEImg2ImgPipeline(DiffusionPipeline): + """ + Pipeline for generating latent representation of a 3D asset and rendering with the NeRF method from an image. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + image_encoder ([`~transformers.CLIPVisionModel`]): + Frozen image-encoder. + image_processor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to process images. + scheduler ([`HeunDiscreteScheduler`]): + A scheduler to be used in combination with the `prior` model to generate image embedding. + shap_e_renderer ([`ShapERenderer`]): + Shap-E renderer projects the generated latents into parameters of a MLP to create 3D objects with the NeRF + rendering method. + """ + + model_cpu_offload_seq = "image_encoder->prior" + _exclude_from_cpu_offload = ["shap_e_renderer"] + + def __init__( + self, + prior: PriorTransformer, + image_encoder: CLIPVisionModel, + image_processor: CLIPImageProcessor, + scheduler: HeunDiscreteScheduler, + shap_e_renderer: ShapERenderer, + ): + super().__init__() + + self.register_modules( + prior=prior, + image_encoder=image_encoder, + image_processor=image_processor, + scheduler=scheduler, + shap_e_renderer=shap_e_renderer, + ) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def _encode_image( + self, + image, + device, + num_images_per_prompt, + do_classifier_free_guidance, + ): + if isinstance(image, list) and isinstance(image[0], torch.Tensor): + image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0) + + if not isinstance(image, torch.Tensor): + image = self.image_processor(image, return_tensors="pt").pixel_values[0].unsqueeze(0) + + image = image.to(dtype=self.image_encoder.dtype, device=device) + + image_embeds = self.image_encoder(image)["last_hidden_state"] + image_embeds = image_embeds[:, 1:, :].contiguous() # batch_size, dim, 256 + + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.zeros_like(image_embeds) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeds = torch.cat([negative_image_embeds, image_embeds]) + + return image_embeds + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PIL.Image.Image | list[PIL.Image.Image], + num_images_per_prompt: int = 1, + num_inference_steps: int = 25, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + guidance_scale: float = 4.0, + frame_size: int = 64, + output_type: str | None = "pil", # pil, np, latent, mesh + return_dict: bool = True, + ): + """ + The call function to the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be used as the starting point. Can also accept image + latents as image, but if passing latents directly it is not encoded again. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + frame_size (`int`, *optional*, default to 64): + The width and height of each image frame of the generated 3D output. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`), `"latent"` (`torch.Tensor`), or mesh ([`MeshDecoderOutput`]). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] instead of a plain + tuple. + + Examples: + + Returns: + [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, torch.Tensor): + batch_size = image.shape[0] + elif isinstance(image, list) and isinstance(image[0], (torch.Tensor, PIL.Image.Image)): + batch_size = len(image) + else: + raise ValueError( + f"`image` has to be of type `PIL.Image.Image`, `torch.Tensor`, `list[PIL.Image.Image]` or `list[torch.Tensor]` but is {type(image)}" + ) + + device = self._execution_device + + batch_size = batch_size * num_images_per_prompt + + do_classifier_free_guidance = guidance_scale > 1.0 + image_embeds = self._encode_image(image, device, num_images_per_prompt, do_classifier_free_guidance) + + # prior + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + num_embeddings = self.prior.config.num_embeddings + embedding_dim = self.prior.config.embedding_dim + if latents is None: + latents = self.prepare_latents( + (batch_size, num_embeddings * embedding_dim), + image_embeds.dtype, + device, + generator, + latents, + self.scheduler, + ) + + # YiYi notes: for testing only to match ldm, we can directly create a latents with desired shape: batch_size, num_embeddings, embedding_dim + latents = latents.reshape(latents.shape[0], num_embeddings, embedding_dim) + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + noise_pred = self.prior( + scaled_model_input, + timestep=t, + proj_embedding=image_embeds, + ).predicted_image_embedding + + # remove the variance + noise_pred, _ = noise_pred.split( + scaled_model_input.shape[2], dim=2 + ) # batch_size, num_embeddings, embedding_dim + + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred - noise_pred_uncond) + + latents = self.scheduler.step( + noise_pred, + timestep=t, + sample=latents, + ).prev_sample + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type not in ["np", "pil", "latent", "mesh"]: + raise ValueError( + f"Only the output types `pil`, `np`, `latent` and `mesh` are supported not output_type={output_type}" + ) + + # Offload all models + self.maybe_free_model_hooks() + + if output_type == "latent": + return ShapEPipelineOutput(images=latents) + + images = [] + if output_type == "mesh": + for i, latent in enumerate(latents): + mesh = self.shap_e_renderer.decode_to_mesh( + latent[None, :], + device, + ) + images.append(mesh) + + else: + # np, pil + for i, latent in enumerate(latents): + image = self.shap_e_renderer.decode_to_image( + latent[None, :], + device, + size=frame_size, + ) + images.append(image) + + images = torch.stack(images) + + images = images.cpu().numpy() + + if output_type == "pil": + images = [self.numpy_to_pil(image) for image in images] + + if not return_dict: + return (images,) + + return ShapEPipelineOutput(images=images) diff --git a/diffusers/pipelines/shap_e/renderer.py b/diffusers/pipelines/shap_e/renderer.py new file mode 100644 index 0000000000000000000000000000000000000000..0c2058c887fc187060e94a45534129c8efcfda04 --- /dev/null +++ b/diffusers/pipelines/shap_e/renderer.py @@ -0,0 +1,1049 @@ +# Copyright 2025 Open AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models import ModelMixin +from ...utils import BaseOutput +from .camera import create_pan_cameras + + +def sample_pmf(pmf: torch.Tensor, n_samples: int) -> torch.Tensor: + r""" + Sample from the given discrete probability distribution with replacement. + + The i-th bin is assumed to have mass pmf[i]. + + Args: + pmf: [batch_size, *shape, n_samples, 1] where (pmf.sum(dim=-2) == 1).all() + n_samples: number of samples + + Return: + indices sampled with replacement + """ + + *shape, support_size, last_dim = pmf.shape + assert last_dim == 1 + + cdf = torch.cumsum(pmf.view(-1, support_size), dim=1) + inds = torch.searchsorted(cdf, torch.rand(cdf.shape[0], n_samples, device=cdf.device)) + + return inds.view(*shape, n_samples, 1).clamp(0, support_size - 1) + + +def posenc_nerf(x: torch.Tensor, min_deg: int = 0, max_deg: int = 15) -> torch.Tensor: + """ + Concatenate x and its positional encodings, following NeRF. + + Reference: https://huggingface.co/papers/2210.04628 + """ + if min_deg == max_deg: + return x + + scales = 2.0 ** torch.arange(min_deg, max_deg, dtype=x.dtype, device=x.device) + *shape, dim = x.shape + xb = (x.reshape(-1, 1, dim) * scales.view(1, -1, 1)).reshape(*shape, -1) + assert xb.shape[-1] == dim * (max_deg - min_deg) + emb = torch.cat([xb, xb + math.pi / 2.0], axis=-1).sin() + return torch.cat([x, emb], dim=-1) + + +def encode_position(position): + return posenc_nerf(position, min_deg=0, max_deg=15) + + +def encode_direction(position, direction=None): + if direction is None: + return torch.zeros_like(posenc_nerf(position, min_deg=0, max_deg=8)) + else: + return posenc_nerf(direction, min_deg=0, max_deg=8) + + +def _sanitize_name(x: str) -> str: + return x.replace(".", "__") + + +def integrate_samples(volume_range, ts, density, channels): + r""" + Function integrating the model output. + + Args: + volume_range: Specifies the integral range [t0, t1] + ts: timesteps + density: torch.Tensor [batch_size, *shape, n_samples, 1] + channels: torch.Tensor [batch_size, *shape, n_samples, n_channels] + returns: + channels: integrated rgb output weights: torch.Tensor [batch_size, *shape, n_samples, 1] (density + *transmittance)[i] weight for each rgb output at [..., i, :]. transmittance: transmittance of this volume + ) + """ + + # 1. Calculate the weights + _, _, dt = volume_range.partition(ts) + ddensity = density * dt + + mass = torch.cumsum(ddensity, dim=-2) + transmittance = torch.exp(-mass[..., -1, :]) + + alphas = 1.0 - torch.exp(-ddensity) + Ts = torch.exp(torch.cat([torch.zeros_like(mass[..., :1, :]), -mass[..., :-1, :]], dim=-2)) + # This is the probability of light hitting and reflecting off of + # something at depth [..., i, :]. + weights = alphas * Ts + + # 2. Integrate channels + channels = torch.sum(channels * weights, dim=-2) + + return channels, weights, transmittance + + +def volume_query_points(volume, grid_size): + indices = torch.arange(grid_size**3, device=volume.bbox_min.device) + zs = indices % grid_size + ys = torch.div(indices, grid_size, rounding_mode="trunc") % grid_size + xs = torch.div(indices, grid_size**2, rounding_mode="trunc") % grid_size + combined = torch.stack([xs, ys, zs], dim=1) + return (combined.float() / (grid_size - 1)) * (volume.bbox_max - volume.bbox_min) + volume.bbox_min + + +def _convert_srgb_to_linear(u: torch.Tensor): + return torch.where(u <= 0.04045, u / 12.92, ((u + 0.055) / 1.055) ** 2.4) + + +def _create_flat_edge_indices( + flat_cube_indices: torch.Tensor, + grid_size: tuple[int, int, int], +): + num_xs = (grid_size[0] - 1) * grid_size[1] * grid_size[2] + y_offset = num_xs + num_ys = grid_size[0] * (grid_size[1] - 1) * grid_size[2] + z_offset = num_xs + num_ys + return torch.stack( + [ + # Edges spanning x-axis. + flat_cube_indices[:, 0] * grid_size[1] * grid_size[2] + + flat_cube_indices[:, 1] * grid_size[2] + + flat_cube_indices[:, 2], + flat_cube_indices[:, 0] * grid_size[1] * grid_size[2] + + (flat_cube_indices[:, 1] + 1) * grid_size[2] + + flat_cube_indices[:, 2], + flat_cube_indices[:, 0] * grid_size[1] * grid_size[2] + + flat_cube_indices[:, 1] * grid_size[2] + + flat_cube_indices[:, 2] + + 1, + flat_cube_indices[:, 0] * grid_size[1] * grid_size[2] + + (flat_cube_indices[:, 1] + 1) * grid_size[2] + + flat_cube_indices[:, 2] + + 1, + # Edges spanning y-axis. + ( + y_offset + + flat_cube_indices[:, 0] * (grid_size[1] - 1) * grid_size[2] + + flat_cube_indices[:, 1] * grid_size[2] + + flat_cube_indices[:, 2] + ), + ( + y_offset + + (flat_cube_indices[:, 0] + 1) * (grid_size[1] - 1) * grid_size[2] + + flat_cube_indices[:, 1] * grid_size[2] + + flat_cube_indices[:, 2] + ), + ( + y_offset + + flat_cube_indices[:, 0] * (grid_size[1] - 1) * grid_size[2] + + flat_cube_indices[:, 1] * grid_size[2] + + flat_cube_indices[:, 2] + + 1 + ), + ( + y_offset + + (flat_cube_indices[:, 0] + 1) * (grid_size[1] - 1) * grid_size[2] + + flat_cube_indices[:, 1] * grid_size[2] + + flat_cube_indices[:, 2] + + 1 + ), + # Edges spanning z-axis. + ( + z_offset + + flat_cube_indices[:, 0] * grid_size[1] * (grid_size[2] - 1) + + flat_cube_indices[:, 1] * (grid_size[2] - 1) + + flat_cube_indices[:, 2] + ), + ( + z_offset + + (flat_cube_indices[:, 0] + 1) * grid_size[1] * (grid_size[2] - 1) + + flat_cube_indices[:, 1] * (grid_size[2] - 1) + + flat_cube_indices[:, 2] + ), + ( + z_offset + + flat_cube_indices[:, 0] * grid_size[1] * (grid_size[2] - 1) + + (flat_cube_indices[:, 1] + 1) * (grid_size[2] - 1) + + flat_cube_indices[:, 2] + ), + ( + z_offset + + (flat_cube_indices[:, 0] + 1) * grid_size[1] * (grid_size[2] - 1) + + (flat_cube_indices[:, 1] + 1) * (grid_size[2] - 1) + + flat_cube_indices[:, 2] + ), + ], + dim=-1, + ) + + +class VoidNeRFModel(nn.Module): + """ + Implements the default empty space model where all queries are rendered as background. + """ + + def __init__(self, background, channel_scale=255.0): + super().__init__() + background = nn.Parameter(torch.from_numpy(np.array(background)).to(dtype=torch.float32) / channel_scale) + + self.register_buffer("background", background) + + def forward(self, position): + background = self.background[None].to(position.device) + + shape = position.shape[:-1] + ones = [1] * (len(shape) - 1) + n_channels = background.shape[-1] + background = torch.broadcast_to(background.view(background.shape[0], *ones, n_channels), [*shape, n_channels]) + + return background + + +@dataclass +class VolumeRange: + t0: torch.Tensor + t1: torch.Tensor + intersected: torch.Tensor + + def __post_init__(self): + assert self.t0.shape == self.t1.shape == self.intersected.shape + + def partition(self, ts): + """ + Partitions t0 and t1 into n_samples intervals. + + Args: + ts: [batch_size, *shape, n_samples, 1] + + Return: + + lower: [batch_size, *shape, n_samples, 1] upper: [batch_size, *shape, n_samples, 1] delta: [batch_size, + *shape, n_samples, 1] + + where + ts \\in [lower, upper] deltas = upper - lower + """ + + mids = (ts[..., 1:, :] + ts[..., :-1, :]) * 0.5 + lower = torch.cat([self.t0[..., None, :], mids], dim=-2) + upper = torch.cat([mids, self.t1[..., None, :]], dim=-2) + delta = upper - lower + assert lower.shape == upper.shape == delta.shape == ts.shape + return lower, upper, delta + + +class BoundingBoxVolume(nn.Module): + """ + Axis-aligned bounding box defined by the two opposite corners. + """ + + def __init__( + self, + *, + bbox_min, + bbox_max, + min_dist: float = 0.0, + min_t_range: float = 1e-3, + ): + """ + Args: + bbox_min: the left/bottommost corner of the bounding box + bbox_max: the other corner of the bounding box + min_dist: all rays should start at least this distance away from the origin. + """ + super().__init__() + + self.min_dist = min_dist + self.min_t_range = min_t_range + + self.bbox_min = torch.tensor(bbox_min) + self.bbox_max = torch.tensor(bbox_max) + self.bbox = torch.stack([self.bbox_min, self.bbox_max]) + assert self.bbox.shape == (2, 3) + assert min_dist >= 0.0 + assert min_t_range > 0.0 + + def intersect( + self, + origin: torch.Tensor, + direction: torch.Tensor, + t0_lower: torch.Tensor | None = None, + epsilon=1e-6, + ): + """ + Args: + origin: [batch_size, *shape, 3] + direction: [batch_size, *shape, 3] + t0_lower: Optional [batch_size, *shape, 1] lower bound of t0 when intersecting this volume. + params: Optional meta parameters in case Volume is parametric + epsilon: to stabilize calculations + + Return: + A tuple of (t0, t1, intersected) where each has a shape [batch_size, *shape, 1]. If a ray intersects with + the volume, `o + td` is in the volume for all t in [t0, t1]. If the volume is bounded, t1 is guaranteed to + be on the boundary of the volume. + """ + + batch_size, *shape, _ = origin.shape + ones = [1] * len(shape) + bbox = self.bbox.view(1, *ones, 2, 3).to(origin.device) + + def _safe_divide(a, b, epsilon=1e-6): + return a / torch.where(b < 0, b - epsilon, b + epsilon) + + ts = _safe_divide(bbox - origin[..., None, :], direction[..., None, :], epsilon=epsilon) + + # Cases to think about: + # + # 1. t1 <= t0: the ray does not pass through the AABB. + # 2. t0 < t1 <= 0: the ray intersects but the BB is behind the origin. + # 3. t0 <= 0 <= t1: the ray starts from inside the BB + # 4. 0 <= t0 < t1: the ray is not inside and intersects with the BB twice. + # + # 1 and 4 are clearly handled from t0 < t1 below. + # Making t0 at least min_dist (>= 0) takes care of 2 and 3. + t0 = ts.min(dim=-2).values.max(dim=-1, keepdim=True).values.clamp(self.min_dist) + t1 = ts.max(dim=-2).values.min(dim=-1, keepdim=True).values + assert t0.shape == t1.shape == (batch_size, *shape, 1) + if t0_lower is not None: + assert t0.shape == t0_lower.shape + t0 = torch.maximum(t0, t0_lower) + + intersected = t0 + self.min_t_range < t1 + t0 = torch.where(intersected, t0, torch.zeros_like(t0)) + t1 = torch.where(intersected, t1, torch.ones_like(t1)) + + return VolumeRange(t0=t0, t1=t1, intersected=intersected) + + +class StratifiedRaySampler(nn.Module): + """ + Instead of fixed intervals, a sample is drawn uniformly at random from each interval. + """ + + def __init__(self, depth_mode: str = "linear"): + """ + :param depth_mode: linear samples ts linearly in depth. harmonic ensures + closer points are sampled more densely. + """ + self.depth_mode = depth_mode + assert self.depth_mode in ("linear", "geometric", "harmonic") + + def sample( + self, + t0: torch.Tensor, + t1: torch.Tensor, + n_samples: int, + epsilon: float = 1e-3, + ) -> torch.Tensor: + """ + Args: + t0: start time has shape [batch_size, *shape, 1] + t1: finish time has shape [batch_size, *shape, 1] + n_samples: number of ts to sample + Return: + sampled ts of shape [batch_size, *shape, n_samples, 1] + """ + ones = [1] * (len(t0.shape) - 1) + ts = torch.linspace(0, 1, n_samples).view(*ones, n_samples).to(t0.dtype).to(t0.device) + + if self.depth_mode == "linear": + ts = t0 * (1.0 - ts) + t1 * ts + elif self.depth_mode == "geometric": + ts = (t0.clamp(epsilon).log() * (1.0 - ts) + t1.clamp(epsilon).log() * ts).exp() + elif self.depth_mode == "harmonic": + # The original NeRF recommends this interpolation scheme for + # spherical scenes, but there could be some weird edge cases when + # the observer crosses from the inner to outer volume. + ts = 1.0 / (1.0 / t0.clamp(epsilon) * (1.0 - ts) + 1.0 / t1.clamp(epsilon) * ts) + + mids = 0.5 * (ts[..., 1:] + ts[..., :-1]) + upper = torch.cat([mids, t1], dim=-1) + lower = torch.cat([t0, mids], dim=-1) + # yiyi notes: add a random seed here for testing, don't forget to remove + torch.manual_seed(0) + t_rand = torch.rand_like(ts) + + ts = lower + (upper - lower) * t_rand + return ts.unsqueeze(-1) + + +class ImportanceRaySampler(nn.Module): + """ + Given the initial estimate of densities, this samples more from regions/bins expected to have objects. + """ + + def __init__( + self, + volume_range: VolumeRange, + ts: torch.Tensor, + weights: torch.Tensor, + blur_pool: bool = False, + alpha: float = 1e-5, + ): + """ + Args: + volume_range: the range in which a ray intersects the given volume. + ts: earlier samples from the coarse rendering step + weights: discretized version of density * transmittance + blur_pool: if true, use 2-tap max + 2-tap blur filter from mip-NeRF. + alpha: small value to add to weights. + """ + self.volume_range = volume_range + self.ts = ts.clone().detach() + self.weights = weights.clone().detach() + self.blur_pool = blur_pool + self.alpha = alpha + + @torch.no_grad() + def sample(self, t0: torch.Tensor, t1: torch.Tensor, n_samples: int) -> torch.Tensor: + """ + Args: + t0: start time has shape [batch_size, *shape, 1] + t1: finish time has shape [batch_size, *shape, 1] + n_samples: number of ts to sample + Return: + sampled ts of shape [batch_size, *shape, n_samples, 1] + """ + lower, upper, _ = self.volume_range.partition(self.ts) + + batch_size, *shape, n_coarse_samples, _ = self.ts.shape + + weights = self.weights + if self.blur_pool: + padded = torch.cat([weights[..., :1, :], weights, weights[..., -1:, :]], dim=-2) + maxes = torch.maximum(padded[..., :-1, :], padded[..., 1:, :]) + weights = 0.5 * (maxes[..., :-1, :] + maxes[..., 1:, :]) + weights = weights + self.alpha + pmf = weights / weights.sum(dim=-2, keepdim=True) + inds = sample_pmf(pmf, n_samples) + assert inds.shape == (batch_size, *shape, n_samples, 1) + assert (inds >= 0).all() and (inds < n_coarse_samples).all() + + t_rand = torch.rand(inds.shape, device=inds.device) + lower_ = torch.gather(lower, -2, inds) + upper_ = torch.gather(upper, -2, inds) + + ts = lower_ + (upper_ - lower_) * t_rand + ts = torch.sort(ts, dim=-2).values + return ts + + +@dataclass +class MeshDecoderOutput(BaseOutput): + """ + A 3D triangle mesh with optional data at the vertices and faces. + + Args: + verts (`torch.Tensor` of shape `(N, 3)`): + array of vertext coordinates + faces (`torch.Tensor` of shape `(N, 3)`): + array of triangles, pointing to indices in verts. + vertext_channels (Dict): + vertext coordinates for each color channel + """ + + verts: torch.Tensor + faces: torch.Tensor + vertex_channels: dict[str, torch.Tensor] + + +class MeshDecoder(nn.Module): + """ + Construct meshes from Signed distance functions (SDFs) using marching cubes method + """ + + def __init__(self): + super().__init__() + cases = torch.zeros(256, 5, 3, dtype=torch.long) + masks = torch.zeros(256, 5, dtype=torch.bool) + + self.register_buffer("cases", cases) + self.register_buffer("masks", masks) + + def forward(self, field: torch.Tensor, min_point: torch.Tensor, size: torch.Tensor): + """ + For a signed distance field, produce a mesh using marching cubes. + + :param field: a 3D tensor of field values, where negative values correspond + to the outside of the shape. The dimensions correspond to the x, y, and z directions, respectively. + :param min_point: a tensor of shape [3] containing the point corresponding + to (0, 0, 0) in the field. + :param size: a tensor of shape [3] containing the per-axis distance from the + (0, 0, 0) field corner and the (-1, -1, -1) field corner. + """ + assert len(field.shape) == 3, "input must be a 3D scalar field" + dev = field.device + + cases = self.cases.to(dev) + masks = self.masks.to(dev) + + min_point = min_point.to(dev) + size = size.to(dev) + + grid_size = field.shape + grid_size_tensor = torch.tensor(grid_size).to(size) + + # Create bitmasks between 0 and 255 (inclusive) indicating the state + # of the eight corners of each cube. + bitmasks = (field > 0).to(torch.uint8) + bitmasks = bitmasks[:-1, :, :] | (bitmasks[1:, :, :] << 1) + bitmasks = bitmasks[:, :-1, :] | (bitmasks[:, 1:, :] << 2) + bitmasks = bitmasks[:, :, :-1] | (bitmasks[:, :, 1:] << 4) + + # Compute corner coordinates across the entire grid. + corner_coords = torch.empty(*grid_size, 3, device=dev, dtype=field.dtype) + corner_coords[range(grid_size[0]), :, :, 0] = torch.arange(grid_size[0], device=dev, dtype=field.dtype)[ + :, None, None + ] + corner_coords[:, range(grid_size[1]), :, 1] = torch.arange(grid_size[1], device=dev, dtype=field.dtype)[ + :, None + ] + corner_coords[:, :, range(grid_size[2]), 2] = torch.arange(grid_size[2], device=dev, dtype=field.dtype) + + # Compute all vertices across all edges in the grid, even though we will + # throw some out later. We have (X-1)*Y*Z + X*(Y-1)*Z + X*Y*(Z-1) vertices. + # These are all midpoints, and don't account for interpolation (which is + # done later based on the used edge midpoints). + edge_midpoints = torch.cat( + [ + ((corner_coords[:-1] + corner_coords[1:]) / 2).reshape(-1, 3), + ((corner_coords[:, :-1] + corner_coords[:, 1:]) / 2).reshape(-1, 3), + ((corner_coords[:, :, :-1] + corner_coords[:, :, 1:]) / 2).reshape(-1, 3), + ], + dim=0, + ) + + # Create a flat array of [X, Y, Z] indices for each cube. + cube_indices = torch.zeros( + grid_size[0] - 1, grid_size[1] - 1, grid_size[2] - 1, 3, device=dev, dtype=torch.long + ) + cube_indices[range(grid_size[0] - 1), :, :, 0] = torch.arange(grid_size[0] - 1, device=dev)[:, None, None] + cube_indices[:, range(grid_size[1] - 1), :, 1] = torch.arange(grid_size[1] - 1, device=dev)[:, None] + cube_indices[:, :, range(grid_size[2] - 1), 2] = torch.arange(grid_size[2] - 1, device=dev) + flat_cube_indices = cube_indices.reshape(-1, 3) + + # Create a flat array mapping each cube to 12 global edge indices. + edge_indices = _create_flat_edge_indices(flat_cube_indices, grid_size) + + # Apply the LUT to figure out the triangles. + flat_bitmasks = bitmasks.reshape(-1).long() # must cast to long for indexing to believe this not a mask + local_tris = cases[flat_bitmasks] + local_masks = masks[flat_bitmasks] + # Compute the global edge indices for the triangles. + global_tris = torch.gather(edge_indices, 1, local_tris.reshape(local_tris.shape[0], -1)).reshape( + local_tris.shape + ) + # Select the used triangles for each cube. + selected_tris = global_tris.reshape(-1, 3)[local_masks.reshape(-1)] + + # Now we have a bunch of indices into the full list of possible vertices, + # but we want to reduce this list to only the used vertices. + used_vertex_indices = torch.unique(selected_tris.view(-1)) + used_edge_midpoints = edge_midpoints[used_vertex_indices] + old_index_to_new_index = torch.zeros(len(edge_midpoints), device=dev, dtype=torch.long) + old_index_to_new_index[used_vertex_indices] = torch.arange( + len(used_vertex_indices), device=dev, dtype=torch.long + ) + + # Rewrite the triangles to use the new indices + faces = torch.gather(old_index_to_new_index, 0, selected_tris.view(-1)).reshape(selected_tris.shape) + + # Compute the actual interpolated coordinates corresponding to edge midpoints. + v1 = torch.floor(used_edge_midpoints).to(torch.long) + v2 = torch.ceil(used_edge_midpoints).to(torch.long) + s1 = field[v1[:, 0], v1[:, 1], v1[:, 2]] + s2 = field[v2[:, 0], v2[:, 1], v2[:, 2]] + p1 = (v1.float() / (grid_size_tensor - 1)) * size + min_point + p2 = (v2.float() / (grid_size_tensor - 1)) * size + min_point + # The signs of s1 and s2 should be different. We want to find + # t such that t*s2 + (1-t)*s1 = 0. + t = (s1 / (s1 - s2))[:, None] + verts = t * p2 + (1 - t) * p1 + + return MeshDecoderOutput(verts=verts, faces=faces, vertex_channels=None) + + +@dataclass +class MLPNeRFModelOutput(BaseOutput): + density: torch.Tensor + signed_distance: torch.Tensor + channels: torch.Tensor + ts: torch.Tensor + + +class MLPNeRSTFModel(ModelMixin, ConfigMixin): + @register_to_config + def __init__( + self, + d_hidden: int = 256, + n_output: int = 12, + n_hidden_layers: int = 6, + act_fn: str = "swish", + insert_direction_at: int = 4, + ): + super().__init__() + + # Instantiate the MLP + + # Find out the dimension of encoded position and direction + dummy = torch.eye(1, 3) + d_posenc_pos = encode_position(position=dummy).shape[-1] + d_posenc_dir = encode_direction(position=dummy).shape[-1] + + mlp_widths = [d_hidden] * n_hidden_layers + input_widths = [d_posenc_pos] + mlp_widths + output_widths = mlp_widths + [n_output] + + if insert_direction_at is not None: + input_widths[insert_direction_at] += d_posenc_dir + + self.mlp = nn.ModuleList([nn.Linear(d_in, d_out) for d_in, d_out in zip(input_widths, output_widths)]) + + if act_fn == "swish": + # self.activation = swish + # yiyi testing: + self.activation = lambda x: F.silu(x) + else: + raise ValueError(f"Unsupported activation function {act_fn}") + + self.sdf_activation = torch.tanh + self.density_activation = torch.nn.functional.relu + self.channel_activation = torch.sigmoid + + def map_indices_to_keys(self, output): + h_map = { + "sdf": (0, 1), + "density_coarse": (1, 2), + "density_fine": (2, 3), + "stf": (3, 6), + "nerf_coarse": (6, 9), + "nerf_fine": (9, 12), + } + + mapped_output = {k: output[..., start:end] for k, (start, end) in h_map.items()} + + return mapped_output + + def forward(self, *, position, direction, ts, nerf_level="coarse", rendering_mode="nerf"): + h = encode_position(position) + + h_preact = h + h_directionless = None + for i, layer in enumerate(self.mlp): + if i == self.config.insert_direction_at: # 4 in the config + h_directionless = h_preact + h_direction = encode_direction(position, direction=direction) + h = torch.cat([h, h_direction], dim=-1) + + h = layer(h) + + h_preact = h + + if i < len(self.mlp) - 1: + h = self.activation(h) + + h_final = h + if h_directionless is None: + h_directionless = h_preact + + activation = self.map_indices_to_keys(h_final) + + if nerf_level == "coarse": + h_density = activation["density_coarse"] + else: + h_density = activation["density_fine"] + + if rendering_mode == "nerf": + if nerf_level == "coarse": + h_channels = activation["nerf_coarse"] + else: + h_channels = activation["nerf_fine"] + + elif rendering_mode == "stf": + h_channels = activation["stf"] + + density = self.density_activation(h_density) + signed_distance = self.sdf_activation(activation["sdf"]) + channels = self.channel_activation(h_channels) + + # yiyi notes: I think signed_distance is not used + return MLPNeRFModelOutput(density=density, signed_distance=signed_distance, channels=channels, ts=ts) + + +class ChannelsProj(nn.Module): + def __init__( + self, + *, + vectors: int, + channels: int, + d_latent: int, + ): + super().__init__() + self.proj = nn.Linear(d_latent, vectors * channels) + self.norm = nn.LayerNorm(channels) + self.d_latent = d_latent + self.vectors = vectors + self.channels = channels + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x_bvd = x + w_vcd = self.proj.weight.view(self.vectors, self.channels, self.d_latent) + b_vc = self.proj.bias.view(1, self.vectors, self.channels) + h = torch.einsum("bvd,vcd->bvc", x_bvd, w_vcd) + h = self.norm(h) + + h = h + b_vc + return h + + +class ShapEParamsProjModel(ModelMixin, ConfigMixin): + """ + project the latent representation of a 3D asset to obtain weights of a multi-layer perceptron (MLP). + + For more details, see the original paper: + """ + + @register_to_config + def __init__( + self, + *, + param_names: tuple[str] = ( + "nerstf.mlp.0.weight", + "nerstf.mlp.1.weight", + "nerstf.mlp.2.weight", + "nerstf.mlp.3.weight", + ), + param_shapes: tuple[tuple[int]] = ( + (256, 93), + (256, 256), + (256, 256), + (256, 256), + ), + d_latent: int = 1024, + ): + super().__init__() + + # check inputs + if len(param_names) != len(param_shapes): + raise ValueError("Must provide same number of `param_names` as `param_shapes`") + self.projections = nn.ModuleDict({}) + for k, (vectors, channels) in zip(param_names, param_shapes): + self.projections[_sanitize_name(k)] = ChannelsProj( + vectors=vectors, + channels=channels, + d_latent=d_latent, + ) + + def forward(self, x: torch.Tensor): + out = {} + start = 0 + for k, shape in zip(self.config.param_names, self.config.param_shapes): + vectors, _ = shape + end = start + vectors + x_bvd = x[:, start:end] + out[k] = self.projections[_sanitize_name(k)](x_bvd).reshape(len(x), *shape) + start = end + return out + + +class ShapERenderer(ModelMixin, ConfigMixin): + @register_to_config + def __init__( + self, + *, + param_names: tuple[str] = ( + "nerstf.mlp.0.weight", + "nerstf.mlp.1.weight", + "nerstf.mlp.2.weight", + "nerstf.mlp.3.weight", + ), + param_shapes: tuple[tuple[int]] = ( + (256, 93), + (256, 256), + (256, 256), + (256, 256), + ), + d_latent: int = 1024, + d_hidden: int = 256, + n_output: int = 12, + n_hidden_layers: int = 6, + act_fn: str = "swish", + insert_direction_at: int = 4, + background: tuple[float] = ( + 255.0, + 255.0, + 255.0, + ), + ): + super().__init__() + + self.params_proj = ShapEParamsProjModel( + param_names=param_names, + param_shapes=param_shapes, + d_latent=d_latent, + ) + self.mlp = MLPNeRSTFModel(d_hidden, n_output, n_hidden_layers, act_fn, insert_direction_at) + self.void = VoidNeRFModel(background=background, channel_scale=255.0) + self.volume = BoundingBoxVolume(bbox_max=[1.0, 1.0, 1.0], bbox_min=[-1.0, -1.0, -1.0]) + self.mesh_decoder = MeshDecoder() + + @torch.no_grad() + def render_rays(self, rays, sampler, n_samples, prev_model_out=None, render_with_direction=False): + """ + Perform volumetric rendering over a partition of possible t's in the union of rendering volumes (written below + with some abuse of notations) + + C(r) := sum( + transmittance(t[i]) * integrate( + lambda t: density(t) * channels(t) * transmittance(t), [t[i], t[i + 1]], + ) for i in range(len(parts)) + ) + transmittance(t[-1]) * void_model(t[-1]).channels + + where + + 1) transmittance(s) := exp(-integrate(density, [t[0], s])) calculates the probability of light passing through + the volume specified by [t[0], s]. (transmittance of 1 means light can pass freely) 2) density and channels are + obtained by evaluating the appropriate part.model at time t. 3) [t[i], t[i + 1]] is defined as the range of t + where the ray intersects (parts[i].volume \\ union(part.volume for part in parts[:i])) at the surface of the + shell (if bounded). If the ray does not intersect, the integral over this segment is evaluated as 0 and + transmittance(t[i + 1]) := transmittance(t[i]). 4) The last term is integration to infinity (e.g. [t[-1], + math.inf]) that is evaluated by the void_model (i.e. we consider this space to be empty). + + Args: + rays: [batch_size x ... x 2 x 3] origin and direction. sampler: disjoint volume integrals. n_samples: + number of ts to sample. prev_model_outputs: model outputs from the previous rendering step, including + + :return: A tuple of + - `channels` + - A importance samplers for additional fine-grained rendering + - raw model output + """ + origin, direction = rays[..., 0, :], rays[..., 1, :] + + # Integrate over [t[i], t[i + 1]] + + # 1 Intersect the rays with the current volume and sample ts to integrate along. + vrange = self.volume.intersect(origin, direction, t0_lower=None) + ts = sampler.sample(vrange.t0, vrange.t1, n_samples) + ts = ts.to(rays.dtype) + + if prev_model_out is not None: + # Append the previous ts now before fprop because previous + # rendering used a different model and we can't reuse the output. + ts = torch.sort(torch.cat([ts, prev_model_out.ts], dim=-2), dim=-2).values + + batch_size, *_shape, _t0_dim = vrange.t0.shape + _, *ts_shape, _ts_dim = ts.shape + + # 2. Get the points along the ray and query the model + directions = torch.broadcast_to(direction.unsqueeze(-2), [batch_size, *ts_shape, 3]) + positions = origin.unsqueeze(-2) + ts * directions + + directions = directions.to(self.mlp.dtype) + positions = positions.to(self.mlp.dtype) + + optional_directions = directions if render_with_direction else None + + model_out = self.mlp( + position=positions, + direction=optional_directions, + ts=ts, + nerf_level="coarse" if prev_model_out is None else "fine", + ) + + # 3. Integrate the model results + channels, weights, transmittance = integrate_samples( + vrange, model_out.ts, model_out.density, model_out.channels + ) + + # 4. Clean up results that do not intersect with the volume. + transmittance = torch.where(vrange.intersected, transmittance, torch.ones_like(transmittance)) + channels = torch.where(vrange.intersected, channels, torch.zeros_like(channels)) + # 5. integration to infinity (e.g. [t[-1], math.inf]) that is evaluated by the void_model (i.e. we consider this space to be empty). + channels = channels + transmittance * self.void(origin) + + weighted_sampler = ImportanceRaySampler(vrange, ts=model_out.ts, weights=weights) + + return channels, weighted_sampler, model_out + + @torch.no_grad() + def decode_to_image( + self, + latents, + device, + size: int = 64, + ray_batch_size: int = 4096, + n_coarse_samples=64, + n_fine_samples=128, + ): + # project the parameters from the generated latents + projected_params = self.params_proj(latents) + + # update the mlp layers of the renderer + for name, param in self.mlp.state_dict().items(): + if f"nerstf.{name}" in projected_params.keys(): + param.copy_(projected_params[f"nerstf.{name}"].squeeze(0)) + + # create cameras object + camera = create_pan_cameras(size) + rays = camera.camera_rays + rays = rays.to(device) + n_batches = rays.shape[1] // ray_batch_size + + coarse_sampler = StratifiedRaySampler() + + images = [] + + for idx in range(n_batches): + rays_batch = rays[:, idx * ray_batch_size : (idx + 1) * ray_batch_size] + + # render rays with coarse, stratified samples. + _, fine_sampler, coarse_model_out = self.render_rays(rays_batch, coarse_sampler, n_coarse_samples) + # Then, render with additional importance-weighted ray samples. + channels, _, _ = self.render_rays( + rays_batch, fine_sampler, n_fine_samples, prev_model_out=coarse_model_out + ) + + images.append(channels) + + images = torch.cat(images, dim=1) + images = images.view(*camera.shape, camera.height, camera.width, -1).squeeze(0) + + return images + + @torch.no_grad() + def decode_to_mesh( + self, + latents, + device, + grid_size: int = 128, + query_batch_size: int = 4096, + texture_channels: tuple = ("R", "G", "B"), + ): + # 1. project the parameters from the generated latents + projected_params = self.params_proj(latents) + + # 2. update the mlp layers of the renderer + for name, param in self.mlp.state_dict().items(): + if f"nerstf.{name}" in projected_params.keys(): + param.copy_(projected_params[f"nerstf.{name}"].squeeze(0)) + + # 3. decoding with STF rendering + # 3.1 query the SDF values at vertices along a regular 128**3 grid + + query_points = volume_query_points(self.volume, grid_size) + query_positions = query_points[None].repeat(1, 1, 1).to(device=device, dtype=self.mlp.dtype) + + fields = [] + + for idx in range(0, query_positions.shape[1], query_batch_size): + query_batch = query_positions[:, idx : idx + query_batch_size] + + model_out = self.mlp( + position=query_batch, direction=None, ts=None, nerf_level="fine", rendering_mode="stf" + ) + fields.append(model_out.signed_distance) + + # predicted SDF values + fields = torch.cat(fields, dim=1) + fields = fields.float() + + assert len(fields.shape) == 3 and fields.shape[-1] == 1, ( + f"expected [meta_batch x inner_batch] SDF results, but got {fields.shape}" + ) + + fields = fields.reshape(1, *([grid_size] * 3)) + + # create grid 128 x 128 x 128 + # - force a negative border around the SDFs to close off all the models. + full_grid = torch.zeros( + 1, + grid_size + 2, + grid_size + 2, + grid_size + 2, + device=fields.device, + dtype=fields.dtype, + ) + full_grid.fill_(-1.0) + full_grid[:, 1:-1, 1:-1, 1:-1] = fields + fields = full_grid + + # apply a differentiable implementation of Marching Cubes to construct meshs + raw_meshes = [] + mesh_mask = [] + + for field in fields: + raw_mesh = self.mesh_decoder(field, self.volume.bbox_min, self.volume.bbox_max - self.volume.bbox_min) + mesh_mask.append(True) + raw_meshes.append(raw_mesh) + + mesh_mask = torch.tensor(mesh_mask, device=fields.device) + max_vertices = max(len(m.verts) for m in raw_meshes) + + # 3.2. query the texture color head at each vertex of the resulting mesh. + texture_query_positions = torch.stack( + [m.verts[torch.arange(0, max_vertices) % len(m.verts)] for m in raw_meshes], + dim=0, + ) + texture_query_positions = texture_query_positions.to(device=device, dtype=self.mlp.dtype) + + textures = [] + + for idx in range(0, texture_query_positions.shape[1], query_batch_size): + query_batch = texture_query_positions[:, idx : idx + query_batch_size] + + texture_model_out = self.mlp( + position=query_batch, direction=None, ts=None, nerf_level="fine", rendering_mode="stf" + ) + textures.append(texture_model_out.channels) + + # predict texture color + textures = torch.cat(textures, dim=1) + + textures = _convert_srgb_to_linear(textures) + textures = textures.float() + + # 3.3 augment the mesh with texture data + assert len(textures.shape) == 3 and textures.shape[-1] == len(texture_channels), ( + f"expected [meta_batch x inner_batch x texture_channels] field results, but got {textures.shape}" + ) + + for m, texture in zip(raw_meshes, textures): + texture = texture[: len(m.verts)] + m.vertex_channels = dict(zip(texture_channels, texture.unbind(-1))) + + return raw_meshes[0] diff --git a/diffusers/pipelines/skyreels_v2/__init__.py b/diffusers/pipelines/skyreels_v2/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..84d2a2dd35007a7ee1213fd772ad86c96345bf7d --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/__init__.py @@ -0,0 +1,59 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_skyreels_v2"] = ["SkyReelsV2Pipeline"] + _import_structure["pipeline_skyreels_v2_diffusion_forcing"] = ["SkyReelsV2DiffusionForcingPipeline"] + _import_structure["pipeline_skyreels_v2_diffusion_forcing_i2v"] = [ + "SkyReelsV2DiffusionForcingImageToVideoPipeline" + ] + _import_structure["pipeline_skyreels_v2_diffusion_forcing_v2v"] = [ + "SkyReelsV2DiffusionForcingVideoToVideoPipeline" + ] + _import_structure["pipeline_skyreels_v2_i2v"] = ["SkyReelsV2ImageToVideoPipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_skyreels_v2 import SkyReelsV2Pipeline + from .pipeline_skyreels_v2_diffusion_forcing import SkyReelsV2DiffusionForcingPipeline + from .pipeline_skyreels_v2_diffusion_forcing_i2v import SkyReelsV2DiffusionForcingImageToVideoPipeline + from .pipeline_skyreels_v2_diffusion_forcing_v2v import SkyReelsV2DiffusionForcingVideoToVideoPipeline + from .pipeline_skyreels_v2_i2v import SkyReelsV2ImageToVideoPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/skyreels_v2/pipeline_output.py b/diffusers/pipelines/skyreels_v2/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..dac2316362ec20c1874b514637ef1681fb7dd7d3 --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class SkyReelsV2PipelineOutput(BaseOutput): + r""" + Output class for SkyReelsV2 pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2.py b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2.py new file mode 100644 index 0000000000000000000000000000000000000000..faad0fb14086f4a5992c248cd2bf9c1f8ee474fe --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2.py @@ -0,0 +1,616 @@ +# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Any, Callable + +import regex as re +import torch +from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import SkyReelsV2LoraLoaderMixin +from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel +from ...schedulers import UniPCMultistepScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SkyReelsV2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """\ + Examples: + ```py + >>> import torch + >>> from diffusers import ( + ... SkyReelsV2Pipeline, + ... UniPCMultistepScheduler, + ... AutoencoderKLWan, + ... ) + >>> from diffusers.utils import export_to_video + + >>> # Load the pipeline + >>> # Available models: + >>> # - Skywork/SkyReels-V2-T2V-14B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-T2V-14B-720P-Diffusers + >>> vae = AutoencoderKLWan.from_pretrained( + ... "Skywork/SkyReels-V2-T2V-14B-720P-Diffusers", + ... subfolder="vae", + ... torch_dtype=torch.float32, + ... ) + >>> pipe = SkyReelsV2Pipeline.from_pretrained( + ... "Skywork/SkyReels-V2-T2V-14B-720P-Diffusers", + ... vae=vae, + ... torch_dtype=torch.bfloat16, + ... ) + >>> flow_shift = 8.0 # 8.0 for T2V, 5.0 for I2V + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + + >>> output = pipe( + ... prompt=prompt, + ... num_inference_steps=50, + ... height=544, + ... width=960, + ... guidance_scale=6.0, # 6.0 for T2V, 5.0 for I2V + ... num_frames=97, + ... ).frames[0] + >>> export_to_video(output, "video.mp4", fps=24, quality=8) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +class SkyReelsV2Pipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin): + r""" + Pipeline for Text-to-Video (t2v) generation using SkyReels-V2. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`SkyReelsV2Transformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: T5EncoderModel | UMT5EncoderModel, + transformer: SkyReelsV2Transformer3DModel, + vae: AutoencoderKLWan, + scheduler: UniPCMultistepScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.prepare_latents + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 544, + width: int = 960, + num_frames: int = 97, + num_inference_steps: int = 50, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + height (`int`, defaults to `544`): + The height in pixels of the generated image. + width (`int`, defaults to `960`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `97`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `6.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`SkyReelsV2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `512`): + The maximum sequence length for the text encoder. + + Examples: + + Returns: + [`~SkyReelsV2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SkyReelsV2PipelineOutput(frames=video) diff --git a/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing.py b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing.py new file mode 100644 index 0000000000000000000000000000000000000000..8751240a1af9ef8413500092e392b5dae3f1fccc --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing.py @@ -0,0 +1,978 @@ +# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import math +import re +from copy import deepcopy +from typing import Any, Callable + +import torch +from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import SkyReelsV2LoraLoaderMixin +from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel +from ...schedulers import UniPCMultistepScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SkyReelsV2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """\ + Examples: + ```py + >>> import torch + >>> from diffusers import ( + ... SkyReelsV2DiffusionForcingPipeline, + ... UniPCMultistepScheduler, + ... AutoencoderKLWan, + ... ) + >>> from diffusers.utils import export_to_video + + >>> # Load the pipeline + >>> # Available models: + >>> # - Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-DF-14B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-DF-14B-720P-Diffusers + >>> vae = AutoencoderKLWan.from_pretrained( + ... "Skywork/SkyReels-V2-DF-14B-720P-Diffusers", + ... subfolder="vae", + ... torch_dtype=torch.float32, + ... ) + >>> pipe = SkyReelsV2DiffusionForcingPipeline.from_pretrained( + ... "Skywork/SkyReels-V2-DF-14B-720P-Diffusers", + ... vae=vae, + ... torch_dtype=torch.bfloat16, + ... ) + >>> flow_shift = 8.0 # 8.0 for T2V, 5.0 for I2V + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + + >>> output = pipe( + ... prompt=prompt, + ... num_inference_steps=30, + ... height=544, + ... width=960, + ... guidance_scale=6.0, # 6.0 for T2V, 5.0 for I2V + ... num_frames=97, + ... ar_step=5, # Controls asynchronous inference (0 for synchronous mode) + ... causal_block_size=5, # Number of frames processed together in a causal block + ... overlap_history=None, # Number of frames to overlap for smooth transitions in long videos + ... addnoise_condition=20, # Improves consistency in long video generation + ... ).frames[0] + >>> export_to_video(output, "video.mp4", fps=24, quality=8) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class SkyReelsV2DiffusionForcingPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin): + """ + Pipeline for Text-to-Video (t2v) generation using SkyReels-V2 with diffusion forcing. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a specific device, etc.). + + Args: + tokenizer ([`AutoTokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`UMT5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`SkyReelsV2Transformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: T5EncoderModel | UMT5EncoderModel, + transformer: SkyReelsV2Transformer3DModel, + vae: AutoencoderKLWan, + scheduler: UniPCMultistepScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + overlap_history=None, + num_frames=None, + base_num_frames=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if num_frames > base_num_frames and overlap_history is None: + raise ValueError( + "`overlap_history` is required when `num_frames` exceeds `base_num_frames` to ensure smooth transitions in long video generation. " + "Please specify a value for `overlap_history`. Recommended values are 17 or 37." + ) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 97, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + base_latent_num_frames: int | None = None, + video_latents: torch.Tensor | None = None, + causal_block_size: int | None = None, + overlap_history_latent_frames: int | None = None, + long_video_iter: int | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + prefix_video_latents = None + prefix_video_latents_frames = 0 + + if video_latents is not None: # long video generation at the iterations other than the first one + prefix_video_latents = video_latents[:, :, -overlap_history_latent_frames:] + + if prefix_video_latents.shape[2] % causal_block_size != 0: + truncate_len_latents = prefix_video_latents.shape[2] % causal_block_size + logger.warning( + f"The length of prefix video latents is truncated by {truncate_len_latents} frames for the causal block size alignment. " + f"This truncation ensures compatibility with the causal block size, which is required for proper processing. " + f"However, it may slightly affect the continuity of the generated video at the truncation boundary." + ) + prefix_video_latents = prefix_video_latents[:, :, :-truncate_len_latents] + prefix_video_latents_frames = prefix_video_latents.shape[2] + + finished_frame_num = ( + long_video_iter * (base_latent_num_frames - overlap_history_latent_frames) + + overlap_history_latent_frames + ) + left_frame_num = num_latent_frames - finished_frame_num + num_latent_frames = min(left_frame_num + overlap_history_latent_frames, base_latent_num_frames) + elif base_latent_num_frames is not None: # long video generation at the first iteration + num_latent_frames = base_latent_num_frames + else: # short video generation + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + latent_height, + latent_width, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + return latents, num_latent_frames, prefix_video_latents, prefix_video_latents_frames + + def generate_timestep_matrix( + self, + num_latent_frames: int, + step_template: torch.Tensor, + base_num_latent_frames: int, + ar_step: int = 5, + num_pre_ready: int = 0, + causal_block_size: int = 1, + shrink_interval_with_mask: bool = False, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[tuple]]: + """ + This function implements the core diffusion forcing algorithm that creates a coordinated denoising schedule + across temporal frames. It supports both synchronous and asynchronous generation modes: + + **Synchronous Mode** (ar_step=0, causal_block_size=1): + - All frames are denoised simultaneously at each timestep + - Each frame follows the same denoising trajectory: [1000, 800, 600, ..., 0] + - Simpler but may have less temporal consistency for long videos + + **Asynchronous Mode** (ar_step>0, causal_block_size>1): + - Frames are grouped into causal blocks and processed block/chunk-wise + - Each block is denoised in a staggered pattern creating a "denoising wave" + - Earlier blocks are more denoised, later blocks lag behind by ar_step timesteps + - Creates stronger temporal dependencies and better consistency + + Args: + num_latent_frames (int): Total number of latent frames to generate + step_template (torch.Tensor): Base timestep schedule (e.g., [1000, 800, 600, ..., 0]) + base_num_latent_frames (int): Maximum frames the model can process in one forward pass + ar_step (int, optional): Autoregressive step size for temporal lag. + 0 = synchronous, >0 = asynchronous. Defaults to 5. + num_pre_ready (int, optional): + Number of frames already denoised (e.g., from prefix in a video2video task). + Defaults to 0. + causal_block_size (int, optional): Number of frames processed as a causal block. + Defaults to 1. + shrink_interval_with_mask (bool, optional): Whether to optimize processing intervals. + Defaults to False. + + Returns: + tuple containing: + - step_matrix (torch.Tensor): Matrix of timesteps for each frame at each iteration Shape: + [num_iterations, num_latent_frames] + - step_index (torch.Tensor): Index matrix for timestep lookup Shape: [num_iterations, + num_latent_frames] + - step_update_mask (torch.Tensor): Boolean mask indicating which frames to update Shape: + [num_iterations, num_latent_frames] + - valid_interval (list[tuple]): list of (start, end) intervals for each iteration + + Raises: + ValueError: If ar_step is too small for the given configuration + """ + # Initialize lists to store the scheduling matrices and metadata + step_matrix, step_index = [], [] # Will store timestep values and indices for each iteration + update_mask, valid_interval = [], [] # Will store update masks and processing intervals + + # Calculate total number of denoising iterations (add 1 for initial noise state) + num_iterations = len(step_template) + 1 + + # Convert frame counts to block counts for causal processing + # Each block contains causal_block_size frames that are processed together + # E.g.: 25 frames ÷ 5 = 5 blocks total + num_blocks = num_latent_frames // causal_block_size + base_num_blocks = base_num_latent_frames // causal_block_size + + # Validate ar_step is sufficient for the given configuration + # In asynchronous mode, we need enough timesteps to create the staggered pattern + if base_num_blocks < num_blocks: + min_ar_step = len(step_template) / base_num_blocks + if ar_step < min_ar_step: + raise ValueError(f"`ar_step` should be at least {math.ceil(min_ar_step)} in your setting") + + # Extend step_template with boundary values for easier indexing + # 999: dummy value for counter starting from 1 + # 0: final timestep (completely denoised) + step_template = torch.cat( + [ + torch.tensor([999], dtype=torch.int64, device=step_template.device), + step_template.long(), + torch.tensor([0], dtype=torch.int64, device=step_template.device), + ] + ) + + # Initialize the previous row state (tracks denoising progress for each block) + # 0 means not started, num_iterations means fully denoised + pre_row = torch.zeros(num_blocks, dtype=torch.long) + + # Mark pre-ready frames (e.g., from prefix video for a video2video task) as already at final denoising state + if num_pre_ready > 0: + pre_row[: num_pre_ready // causal_block_size] = num_iterations + + # Main loop: Generate denoising schedule until all frames are fully denoised + while not torch.all(pre_row >= (num_iterations - 1)): + # Create new row representing the next denoising step + new_row = torch.zeros(num_blocks, dtype=torch.long) + + # Apply diffusion forcing logic for each block + for i in range(num_blocks): + if i == 0 or pre_row[i - 1] >= ( + num_iterations - 1 + ): # the first frame or the last frame is completely denoised + new_row[i] = pre_row[i] + 1 + else: + # Asynchronous mode: lag behind previous block by ar_step timesteps + # This creates the "diffusion forcing" staggered pattern + new_row[i] = new_row[i - 1] - ar_step + + # Clamp values to valid range [0, num_iterations] + new_row = new_row.clamp(0, num_iterations) + + # Create update mask: True for blocks that need denoising update at this iteration + # Exclude blocks that haven't started (new_row != pre_row) or are finished (new_row != num_iterations) + # Final state example: [False, ..., False, True, True, True, True, True] + # where first 20 frames are done (False) and last 5 frames still need updates (True) + update_mask.append((new_row != pre_row) & (new_row != num_iterations)) + + # Store the iteration state + step_index.append(new_row) # Index into step_template + step_matrix.append(step_template[new_row]) # Actual timestep values + pre_row = new_row # Update for next iteration + + # For videos longer than model capacity, we process in sliding windows + terminal_flag = base_num_blocks + + # Optional optimization: shrink interval based on first update mask + if shrink_interval_with_mask: + idx_sequence = torch.arange(num_blocks, dtype=torch.int64) + update_mask = update_mask[0] + update_mask_idx = idx_sequence[update_mask] + last_update_idx = update_mask_idx[-1].item() + terminal_flag = last_update_idx + 1 + + # Each interval defines which frames to process in the current forward pass + for curr_mask in update_mask: + # Extend terminal flag if current mask has updates beyond current terminal + if terminal_flag < num_blocks and curr_mask[terminal_flag]: + terminal_flag += 1 + # Create interval: [start, end) where start ensures we don't exceed model capacity + valid_interval.append((max(terminal_flag - base_num_blocks, 0), terminal_flag)) + + # Convert lists to tensors for efficient processing + step_update_mask = torch.stack(update_mask, dim=0) + step_index = torch.stack(step_index, dim=0) + step_matrix = torch.stack(step_matrix, dim=0) + + # Each block's schedule is replicated to all frames within that block + if causal_block_size > 1: + # Expand each block to causal_block_size frames + step_update_mask = step_update_mask.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + step_index = step_index.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + step_matrix = step_matrix.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + # Scale intervals from block-level to frame-level + valid_interval = [(s * causal_block_size, e * causal_block_size) for s, e in valid_interval] + + return step_matrix, step_index, step_update_mask, valid_interval + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str], + negative_prompt: str | list[str] = None, + height: int = 544, + width: int = 960, + num_frames: int = 97, + num_inference_steps: int = 50, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + overlap_history: int | None = None, + addnoise_condition: float = 0, + base_num_frames: int = 97, + ar_step: int = 0, + causal_block_size: int | None = None, + fps: int = 24, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `544`): + The height of the generated video. + width (`int`, defaults to `960`): + The width of the generated video. + num_frames (`int`, defaults to `97`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `6.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. (**6.0 for T2V**, **5.0 for I2V**) + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`SkyReelsV2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `512`): + The maximum sequence length of the prompt. + overlap_history (`int`, *optional*, defaults to `None`): + Number of frames to overlap for smooth transitions in long videos. If `None`, the pipeline assumes + short video generation mode, and no overlap is applied. 17 and 37 are recommended to set. + addnoise_condition (`float`, *optional*, defaults to `0`): + This is used to help smooth the long video generation by adding some noise to the clean condition. Too + large noise can cause the inconsistency as well. 20 is a recommended value, and you may try larger + ones, but it is recommended to not exceed 50. + base_num_frames (`int`, *optional*, defaults to `97`): + 97 or 121 | Base frame count (**97 for 540P**, **121 for 720P**) + ar_step (`int`, *optional*, defaults to `0`): + Controls asynchronous inference (0 for synchronous mode) You can set `ar_step=5` to enable asynchronous + inference. When asynchronous inference, `causal_block_size=5` is recommended while it is not supposed + to be set for synchronous generation. Asynchronous inference will take more steps to diffuse the whole + sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous + inference may improve the instruction following and visual consistent performance. + causal_block_size (`int`, *optional*, defaults to `None`): + The number of frames in each block/chunk. Recommended when using asynchronous inference (when ar_step > + 0) + fps (`int`, *optional*, defaults to `24`): + Frame rate of the generated video + + Examples: + + Returns: + [`~SkyReelsV2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + overlap_history, + num_frames, + base_num_frames, + ) + + if addnoise_condition > 60: + logger.warning( + f"The value of 'addnoise_condition' is too large ({addnoise_condition}) and may cause inconsistencies in long video generation. A value of 20 is recommended." + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + if causal_block_size is None: + causal_block_size = self.transformer.config.num_frame_per_block + else: + self.transformer._set_ar_attention(causal_block_size) + + fps_embeds = [fps] * prompt_embeds.shape[0] + fps_embeds = [0 if i == 16 else 1 for i in fps_embeds] + + # Determine if we're doing long video generation + is_long_video = overlap_history is not None and base_num_frames is not None and num_frames > base_num_frames + # Initialize accumulated_latents to store all latents in one tensor + accumulated_latents = None + if is_long_video: + # Long video generation setup + overlap_history_latent_frames = (overlap_history - 1) // self.vae_scale_factor_temporal + 1 + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + base_latent_num_frames = ( + (base_num_frames - 1) // self.vae_scale_factor_temporal + 1 + if base_num_frames is not None + else num_latent_frames + ) + n_iter = ( + 1 + + (num_latent_frames - base_latent_num_frames - 1) + // (base_latent_num_frames - overlap_history_latent_frames) + + 1 + ) + else: + # Short video generation setup + n_iter = 1 + base_latent_num_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + + # Loop through iterations (multiple iterations only for long videos) + for iter_idx in range(n_iter): + if is_long_video: + logger.debug(f"Processing iteration {iter_idx + 1}/{n_iter} for long video generation...") + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents, current_num_latent_frames, prefix_video_latents, prefix_video_latents_frames = ( + self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents if iter_idx == 0 else None, + video_latents=accumulated_latents, # Pass latents directly instead of decoded video + base_latent_num_frames=base_latent_num_frames if is_long_video else None, + causal_block_size=causal_block_size, + overlap_history_latent_frames=overlap_history_latent_frames if is_long_video else None, + long_video_iter=iter_idx if is_long_video else None, + ) + ) + + if prefix_video_latents_frames > 0: + latents[:, :, :prefix_video_latents_frames, :, :] = prefix_video_latents.to(transformer_dtype) + + # 6. Prepare sample schedulers and timestep matrix + sample_schedulers = [] + for _ in range(current_num_latent_frames): + sample_scheduler = deepcopy(self.scheduler) + sample_scheduler.set_timesteps(num_inference_steps, device=device) + sample_schedulers.append(sample_scheduler) + + # Different matrix generation for short vs long video + step_matrix, _, step_update_mask, valid_interval = self.generate_timestep_matrix( + current_num_latent_frames, + timesteps, + current_num_latent_frames if is_long_video else base_latent_num_frames, + ar_step, + prefix_video_latents_frames, + causal_block_size, + ) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(step_matrix) + + with self.progress_bar(total=len(step_matrix)) as progress_bar: + for i, t in enumerate(step_matrix): + if self.interrupt: + continue + + self._current_timestep = t + valid_interval_start, valid_interval_end = valid_interval[i] + latent_model_input = ( + latents[:, :, valid_interval_start:valid_interval_end, :, :].to(transformer_dtype).clone() + ) + timestep = t.expand(latents.shape[0], -1)[:, valid_interval_start:valid_interval_end].clone() + + if addnoise_condition > 0 and valid_interval_start < prefix_video_latents_frames: + noise_factor = 0.001 * addnoise_condition + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] = ( + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] + * (1.0 - noise_factor) + + torch.randn_like( + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] + ) + * noise_factor + ) + timestep[:, valid_interval_start:prefix_video_latents_frames] = addnoise_condition + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + enable_diffusion_forcing=True, + fps=fps_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + enable_diffusion_forcing=True, + fps=fps_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + update_mask_i = step_update_mask[i] + for idx in range(valid_interval_start, valid_interval_end): + if update_mask_i[idx].item(): + latents[:, :, idx, :, :] = sample_schedulers[idx].step( + noise_pred[:, :, idx - valid_interval_start, :, :], + t[idx], + latents[:, :, idx, :, :], + return_dict=False, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(step_matrix) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0 + ): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # Handle latent accumulation for long videos or use the current latents for short videos + if is_long_video: + if accumulated_latents is None: + accumulated_latents = latents + else: + # Keep overlap frames for conditioning but don't include them in final output + accumulated_latents = torch.cat( + [accumulated_latents, latents[:, :, overlap_history_latent_frames:]], dim=2 + ) + + if is_long_video: + latents = accumulated_latents + + self._current_timestep = None + + # Final decoding step - convert latents to pixels + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SkyReelsV2PipelineOutput(frames=video) diff --git a/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_i2v.py b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_i2v.py new file mode 100644 index 0000000000000000000000000000000000000000..a8f1b3a84a4aa770dd07b30fc396db4b7adac5c3 --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_i2v.py @@ -0,0 +1,1059 @@ +# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import math +import re +from copy import deepcopy +from typing import Any, Callable + +import PIL +import torch +from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel + +from diffusers.image_processor import PipelineImageInput +from diffusers.utils.torch_utils import randn_tensor +from diffusers.video_processor import VideoProcessor + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import SkyReelsV2LoraLoaderMixin +from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel +from ...schedulers import UniPCMultistepScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SkyReelsV2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """\ + Examples: + ```py + >>> import torch + >>> from diffusers import ( + ... SkyReelsV2DiffusionForcingImageToVideoPipeline, + ... UniPCMultistepScheduler, + ... AutoencoderKLWan, + ... ) + >>> from diffusers.utils import export_to_video + >>> from PIL import Image + + >>> # Load the pipeline + >>> # Available models: + >>> # - Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-DF-14B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-DF-14B-720P-Diffusers + >>> vae = AutoencoderKLWan.from_pretrained( + ... "Skywork/SkyReels-V2-DF-14B-720P-Diffusers", + ... subfolder="vae", + ... torch_dtype=torch.float32, + ... ) + >>> pipe = SkyReelsV2DiffusionForcingImageToVideoPipeline.from_pretrained( + ... "Skywork/SkyReels-V2-DF-14B-720P-Diffusers", + ... vae=vae, + ... torch_dtype=torch.bfloat16, + ... ) + >>> flow_shift = 5.0 # 8.0 for T2V, 5.0 for I2V + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + >>> image = Image.open("path/to/image.png") + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... num_inference_steps=50, + ... height=544, + ... width=960, + ... guidance_scale=5.0, # 6.0 for T2V, 5.0 for I2V + ... num_frames=97, + ... ar_step=0, # Controls asynchronous inference (0 for synchronous mode) + ... overlap_history=None, # Number of frames to overlap for smooth transitions in long videos + ... addnoise_condition=20, # Improves consistency in long video generation + ... ).frames[0] + >>> export_to_video(output, "video.mp4", fps=24, quality=8) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class SkyReelsV2DiffusionForcingImageToVideoPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin): + """ + Pipeline for Image-to-Video (i2v) generation using SkyReels-V2 with diffusion forcing. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a specific device, etc.). + + Args: + tokenizer ([`AutoTokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`UMT5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`SkyReelsV2Transformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: T5EncoderModel | UMT5EncoderModel, + transformer: SkyReelsV2Transformer3DModel, + vae: AutoencoderKLWan, + scheduler: UniPCMultistepScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + callback_on_step_end_tensor_inputs=None, + overlap_history=None, + num_frames=None, + base_num_frames=None, + ): + if image is not None and image_embeds is not None: + raise ValueError( + f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to" + " only forward one of the two." + ) + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `image_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}") + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if num_frames > base_num_frames and overlap_history is None: + raise ValueError( + "`overlap_history` is required when `num_frames` exceeds `base_num_frames` to ensure smooth transitions in long video generation. " + "Please specify a value for `overlap_history`. Recommended values are 17 or 37." + ) + + def prepare_latents( + self, + image: PipelineImageInput | None, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 97, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + last_image: torch.Tensor | None = None, + video_latents: torch.Tensor | None = None, + base_latent_num_frames: int | None = None, + causal_block_size: int | None = None, + overlap_history_latent_frames: int | None = None, + long_video_iter: int | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + prefix_video_latents_frames = 0 + + if video_latents is not None: # long video generation at the iterations other than the first one + condition = video_latents[:, :, -overlap_history_latent_frames:] + + if condition.shape[2] % causal_block_size != 0: + truncate_len_latents = condition.shape[2] % causal_block_size + logger.warning( + f"The length of prefix video latents is truncated by {truncate_len_latents} frames for the causal block size alignment. " + f"This truncation ensures compatibility with the causal block size, which is required for proper processing. " + f"However, it may slightly affect the continuity of the generated video at the truncation boundary." + ) + condition = condition[:, :, :-truncate_len_latents] + prefix_video_latents_frames = condition.shape[2] + + finished_frame_num = ( + long_video_iter * (base_latent_num_frames - overlap_history_latent_frames) + + overlap_history_latent_frames + ) + left_frame_num = num_latent_frames - finished_frame_num + num_latent_frames = min(left_frame_num + overlap_history_latent_frames, base_latent_num_frames) + elif base_latent_num_frames is not None: # long video generation at the first iteration + num_latent_frames = base_latent_num_frames + else: # short video generation + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + if image is not None: + image = image.unsqueeze(2) + if last_image is not None: + last_image = last_image.unsqueeze(2) + video_condition = torch.cat([image, last_image], dim=0) + else: + video_condition = image + + video_condition = video_condition.to(device=device, dtype=self.vae.dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + + if isinstance(generator, list): + latent_condition = [ + retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") for _ in generator + ] + latent_condition = torch.cat(latent_condition) + else: + latent_condition = retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") + latent_condition = latent_condition.repeat_interleave(batch_size, dim=0) + + latent_condition = latent_condition.to(dtype) + condition = (latent_condition - latents_mean) * latents_std + prefix_video_latents_frames = condition.shape[2] + + return latents, num_latent_frames, condition, prefix_video_latents_frames + + # Copied from diffusers.pipelines.skyreels_v2.pipeline_skyreels_v2_diffusion_forcing.SkyReelsV2DiffusionForcingPipeline.generate_timestep_matrix + def generate_timestep_matrix( + self, + num_latent_frames: int, + step_template: torch.Tensor, + base_num_latent_frames: int, + ar_step: int = 5, + num_pre_ready: int = 0, + causal_block_size: int = 1, + shrink_interval_with_mask: bool = False, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[tuple]]: + """ + This function implements the core diffusion forcing algorithm that creates a coordinated denoising schedule + across temporal frames. It supports both synchronous and asynchronous generation modes: + + **Synchronous Mode** (ar_step=0, causal_block_size=1): + - All frames are denoised simultaneously at each timestep + - Each frame follows the same denoising trajectory: [1000, 800, 600, ..., 0] + - Simpler but may have less temporal consistency for long videos + + **Asynchronous Mode** (ar_step>0, causal_block_size>1): + - Frames are grouped into causal blocks and processed block/chunk-wise + - Each block is denoised in a staggered pattern creating a "denoising wave" + - Earlier blocks are more denoised, later blocks lag behind by ar_step timesteps + - Creates stronger temporal dependencies and better consistency + + Args: + num_latent_frames (int): Total number of latent frames to generate + step_template (torch.Tensor): Base timestep schedule (e.g., [1000, 800, 600, ..., 0]) + base_num_latent_frames (int): Maximum frames the model can process in one forward pass + ar_step (int, optional): Autoregressive step size for temporal lag. + 0 = synchronous, >0 = asynchronous. Defaults to 5. + num_pre_ready (int, optional): + Number of frames already denoised (e.g., from prefix in a video2video task). + Defaults to 0. + causal_block_size (int, optional): Number of frames processed as a causal block. + Defaults to 1. + shrink_interval_with_mask (bool, optional): Whether to optimize processing intervals. + Defaults to False. + + Returns: + tuple containing: + - step_matrix (torch.Tensor): Matrix of timesteps for each frame at each iteration Shape: + [num_iterations, num_latent_frames] + - step_index (torch.Tensor): Index matrix for timestep lookup Shape: [num_iterations, + num_latent_frames] + - step_update_mask (torch.Tensor): Boolean mask indicating which frames to update Shape: + [num_iterations, num_latent_frames] + - valid_interval (list[tuple]): list of (start, end) intervals for each iteration + + Raises: + ValueError: If ar_step is too small for the given configuration + """ + # Initialize lists to store the scheduling matrices and metadata + step_matrix, step_index = [], [] # Will store timestep values and indices for each iteration + update_mask, valid_interval = [], [] # Will store update masks and processing intervals + + # Calculate total number of denoising iterations (add 1 for initial noise state) + num_iterations = len(step_template) + 1 + + # Convert frame counts to block counts for causal processing + # Each block contains causal_block_size frames that are processed together + # E.g.: 25 frames ÷ 5 = 5 blocks total + num_blocks = num_latent_frames // causal_block_size + base_num_blocks = base_num_latent_frames // causal_block_size + + # Validate ar_step is sufficient for the given configuration + # In asynchronous mode, we need enough timesteps to create the staggered pattern + if base_num_blocks < num_blocks: + min_ar_step = len(step_template) / base_num_blocks + if ar_step < min_ar_step: + raise ValueError(f"`ar_step` should be at least {math.ceil(min_ar_step)} in your setting") + + # Extend step_template with boundary values for easier indexing + # 999: dummy value for counter starting from 1 + # 0: final timestep (completely denoised) + step_template = torch.cat( + [ + torch.tensor([999], dtype=torch.int64, device=step_template.device), + step_template.long(), + torch.tensor([0], dtype=torch.int64, device=step_template.device), + ] + ) + + # Initialize the previous row state (tracks denoising progress for each block) + # 0 means not started, num_iterations means fully denoised + pre_row = torch.zeros(num_blocks, dtype=torch.long) + + # Mark pre-ready frames (e.g., from prefix video for a video2video task) as already at final denoising state + if num_pre_ready > 0: + pre_row[: num_pre_ready // causal_block_size] = num_iterations + + # Main loop: Generate denoising schedule until all frames are fully denoised + while not torch.all(pre_row >= (num_iterations - 1)): + # Create new row representing the next denoising step + new_row = torch.zeros(num_blocks, dtype=torch.long) + + # Apply diffusion forcing logic for each block + for i in range(num_blocks): + if i == 0 or pre_row[i - 1] >= ( + num_iterations - 1 + ): # the first frame or the last frame is completely denoised + new_row[i] = pre_row[i] + 1 + else: + # Asynchronous mode: lag behind previous block by ar_step timesteps + # This creates the "diffusion forcing" staggered pattern + new_row[i] = new_row[i - 1] - ar_step + + # Clamp values to valid range [0, num_iterations] + new_row = new_row.clamp(0, num_iterations) + + # Create update mask: True for blocks that need denoising update at this iteration + # Exclude blocks that haven't started (new_row != pre_row) or are finished (new_row != num_iterations) + # Final state example: [False, ..., False, True, True, True, True, True] + # where first 20 frames are done (False) and last 5 frames still need updates (True) + update_mask.append((new_row != pre_row) & (new_row != num_iterations)) + + # Store the iteration state + step_index.append(new_row) # Index into step_template + step_matrix.append(step_template[new_row]) # Actual timestep values + pre_row = new_row # Update for next iteration + + # For videos longer than model capacity, we process in sliding windows + terminal_flag = base_num_blocks + + # Optional optimization: shrink interval based on first update mask + if shrink_interval_with_mask: + idx_sequence = torch.arange(num_blocks, dtype=torch.int64) + update_mask = update_mask[0] + update_mask_idx = idx_sequence[update_mask] + last_update_idx = update_mask_idx[-1].item() + terminal_flag = last_update_idx + 1 + + # Each interval defines which frames to process in the current forward pass + for curr_mask in update_mask: + # Extend terminal flag if current mask has updates beyond current terminal + if terminal_flag < num_blocks and curr_mask[terminal_flag]: + terminal_flag += 1 + # Create interval: [start, end) where start ensures we don't exceed model capacity + valid_interval.append((max(terminal_flag - base_num_blocks, 0), terminal_flag)) + + # Convert lists to tensors for efficient processing + step_update_mask = torch.stack(update_mask, dim=0) + step_index = torch.stack(step_index, dim=0) + step_matrix = torch.stack(step_matrix, dim=0) + + # Each block's schedule is replicated to all frames within that block + if causal_block_size > 1: + # Expand each block to causal_block_size frames + step_update_mask = step_update_mask.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + step_index = step_index.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + step_matrix = step_matrix.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + # Scale intervals from block-level to frame-level + valid_interval = [(s * causal_block_size, e * causal_block_size) for s, e in valid_interval] + + return step_matrix, step_index, step_update_mask, valid_interval + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 544, + width: int = 960, + num_frames: int = 97, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + image_embeds: torch.Tensor | None = None, + last_image: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + overlap_history: int | None = None, + addnoise_condition: float = 0, + base_num_frames: int = 97, + ar_step: int = 0, + causal_block_size: int | None = None, + fps: int = 24, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `544`): + The height of the generated video. + width (`int`, defaults to `960`): + The width of the generated video. + num_frames (`int`, defaults to `97`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. (**6.0 for T2V**, **5.0 for I2V**) + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided, + image embeddings are generated from the `image` input argument. + last_image (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided, + image embeddings are generated from the `image` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`SkyReelsV2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `512`): + The maximum sequence length of the prompt. + overlap_history (`int`, *optional*, defaults to `None`): + Number of frames to overlap for smooth transitions in long videos. If `None`, the pipeline assumes + short video generation mode, and no overlap is applied. 17 and 37 are recommended to set. + addnoise_condition (`float`, *optional*, defaults to `0`): + This is used to help smooth the long video generation by adding some noise to the clean condition. Too + large noise can cause the inconsistency as well. 20 is a recommended value, and you may try larger + ones, but it is recommended to not exceed 50. + base_num_frames (`int`, *optional*, defaults to `97`): + 97 or 121 | Base frame count (**97 for 540P**, **121 for 720P**) + ar_step (`int`, *optional*, defaults to `0`): + Controls asynchronous inference (0 for synchronous mode) You can set `ar_step=5` to enable asynchronous + inference. When asynchronous inference, `causal_block_size=5` is recommended while it is not supposed + to be set for synchronous generation. Asynchronous inference will take more steps to diffuse the whole + sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous + inference may improve the instruction following and visual consistent performance. + causal_block_size (`int`, *optional*, defaults to `None`): + The number of frames in each block/chunk. Recommended when using asynchronous inference (when ar_step > + 0) + fps (`int`, *optional*, defaults to `24`): + Frame rate of the generated video + + Examples: + + Returns: + [`~SkyReelsV2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds, + negative_prompt_embeds, + image_embeds, + callback_on_step_end_tensor_inputs, + overlap_history, + num_frames, + base_num_frames, + ) + + if addnoise_condition > 60: + logger.warning( + f"The value of 'addnoise_condition' is too large ({addnoise_condition}) and may cause inconsistencies in long video generation. A value of 20 is recommended." + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + if causal_block_size is None: + causal_block_size = self.transformer.config.num_frame_per_block + else: + self.transformer._set_ar_attention(causal_block_size) + + fps_embeds = [fps] * prompt_embeds.shape[0] + fps_embeds = [0 if i == 16 else 1 for i in fps_embeds] + + # Determine if we're doing long video generation + is_long_video = overlap_history is not None and base_num_frames is not None and num_frames > base_num_frames + # Initialize accumulated_latents to store all latents in one tensor + accumulated_latents = None + if is_long_video: + # Long video generation setup + overlap_history_latent_frames = (overlap_history - 1) // self.vae_scale_factor_temporal + 1 + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + base_latent_num_frames = ( + (base_num_frames - 1) // self.vae_scale_factor_temporal + 1 + if base_num_frames is not None + else num_latent_frames + ) + n_iter = ( + 1 + + (num_latent_frames - base_latent_num_frames - 1) + // (base_latent_num_frames - overlap_history_latent_frames) + + 1 + ) + else: + # Short video generation setup + n_iter = 1 + base_latent_num_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + + image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.float32) + + if last_image is not None: + last_image = self.video_processor.preprocess(last_image, height=height, width=width).to( + device, dtype=torch.float32 + ) + + # Loop through iterations (multiple iterations only for long videos) + for iter_idx in range(n_iter): + if is_long_video: + logger.debug(f"Processing iteration {iter_idx + 1}/{n_iter} for long video generation...") + + num_channels_latents = self.vae.config.z_dim + latents, current_num_latent_frames, condition, prefix_video_latents_frames = self.prepare_latents( + image if iter_idx == 0 else None, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents if iter_idx == 0 else None, + last_image, + video_latents=accumulated_latents, # Pass latents directly instead of decoded video + base_latent_num_frames=base_latent_num_frames if is_long_video else None, + causal_block_size=causal_block_size, + overlap_history_latent_frames=overlap_history_latent_frames if is_long_video else None, + long_video_iter=iter_idx if is_long_video else None, + ) + + if iter_idx == 0: + latents[:, :, :prefix_video_latents_frames, :, :] = condition[: (condition.shape[0] + 1) // 2].to( + transformer_dtype + ) + else: + latents[:, :, :prefix_video_latents_frames, :, :] = condition.to(transformer_dtype) + + if iter_idx == 0 and last_image is not None: + end_video_latents = condition[condition.shape[0] // 2 :].to(transformer_dtype) + + if last_image is not None and iter_idx + 1 == n_iter: + latents = torch.cat([latents, end_video_latents], dim=2) + base_latent_num_frames += prefix_video_latents_frames + current_num_latent_frames += prefix_video_latents_frames + + # 4. Prepare sample schedulers and timestep matrix + sample_schedulers = [] + for _ in range(current_num_latent_frames): + sample_scheduler = deepcopy(self.scheduler) + sample_scheduler.set_timesteps(num_inference_steps, device=device) + sample_schedulers.append(sample_scheduler) + step_matrix, _, step_update_mask, valid_interval = self.generate_timestep_matrix( + current_num_latent_frames, + timesteps, + base_latent_num_frames, + ar_step, + prefix_video_latents_frames, + causal_block_size, + ) + + if last_image is not None and iter_idx + 1 == n_iter: + step_matrix[:, -prefix_video_latents_frames:] = 0 + step_update_mask[:, -prefix_video_latents_frames:] = False + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(step_matrix) + + with self.progress_bar(total=len(step_matrix)) as progress_bar: + for i, t in enumerate(step_matrix): + if self.interrupt: + continue + + self._current_timestep = t + valid_interval_start, valid_interval_end = valid_interval[i] + latent_model_input = ( + latents[:, :, valid_interval_start:valid_interval_end, :, :].to(transformer_dtype).clone() + ) + timestep = t.expand(latents.shape[0], -1)[:, valid_interval_start:valid_interval_end].clone() + + if addnoise_condition > 0 and valid_interval_start < prefix_video_latents_frames: + noise_factor = 0.001 * addnoise_condition + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] = ( + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] + * (1.0 - noise_factor) + + torch.randn_like( + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] + ) + * noise_factor + ) + timestep[:, valid_interval_start:prefix_video_latents_frames] = addnoise_condition + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + enable_diffusion_forcing=True, + fps=fps_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + enable_diffusion_forcing=True, + fps=fps_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + update_mask_i = step_update_mask[i] + for idx in range(valid_interval_start, valid_interval_end): + if update_mask_i[idx].item(): + latents[:, :, idx, :, :] = sample_schedulers[idx].step( + noise_pred[:, :, idx - valid_interval_start, :, :], + t[idx], + latents[:, :, idx, :, :], + return_dict=False, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(step_matrix) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0 + ): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + # Handle latent accumulation for long videos or use the current latents for short videos + if is_long_video: + if accumulated_latents is None: + accumulated_latents = latents + else: + # Keep overlap frames for conditioning but don't include them in final output + accumulated_latents = torch.cat( + [accumulated_latents, latents[:, :, overlap_history_latent_frames:]], + dim=2, + ) + + if is_long_video: + latents = accumulated_latents + + self._current_timestep = None + + # Final decoding step - convert latents to pixels + if not output_type == "latent": + if last_image is not None: + latents = latents[:, :, :-prefix_video_latents_frames, :, :].to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SkyReelsV2PipelineOutput(frames=video) diff --git a/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_v2v.py b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_v2v.py new file mode 100644 index 0000000000000000000000000000000000000000..924acb850d0908557a1919edbc305a1957ddac0a --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_diffusion_forcing_v2v.py @@ -0,0 +1,1063 @@ +# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +import math +import re +from copy import deepcopy +from typing import Any, Callable + +import torch +from PIL import Image +from transformers import AutoTokenizer, T5EncoderModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import SkyReelsV2LoraLoaderMixin +from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel +from ...schedulers import UniPCMultistepScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SkyReelsV2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """\ + Examples: + ```py + >>> import torch + >>> from diffusers import ( + ... SkyReelsV2DiffusionForcingVideoToVideoPipeline, + ... UniPCMultistepScheduler, + ... AutoencoderKLWan, + ... ) + >>> from diffusers.utils import export_to_video + + >>> # Load the pipeline + >>> # Available models: + >>> # - Skywork/SkyReels-V2-DF-1.3B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-DF-14B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-DF-14B-720P-Diffusers + >>> vae = AutoencoderKLWan.from_pretrained( + ... "Skywork/SkyReels-V2-DF-14B-720P-Diffusers", + ... subfolder="vae", + ... torch_dtype=torch.float32, + ... ) + >>> pipe = SkyReelsV2DiffusionForcingVideoToVideoPipeline.from_pretrained( + ... "Skywork/SkyReels-V2-DF-14B-720P-Diffusers", + ... vae=vae, + ... torch_dtype=torch.bfloat16, + ... ) + >>> flow_shift = 8.0 # 8.0 for T2V, 5.0 for I2V + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + + >>> output = pipe( + ... prompt=prompt, + ... num_inference_steps=50, + ... height=544, + ... width=960, + ... guidance_scale=6.0, # 6.0 for T2V, 5.0 for I2V + ... num_frames=97, + ... ar_step=0, # Controls asynchronous inference (0 for synchronous mode) + ... overlap_history=None, # Number of frames to overlap for smooth transitions in long videos + ... addnoise_condition=20, # Improves consistency in long video generation + ... ).frames[0] + >>> export_to_video(output, "video.mp4", fps=24, quality=8) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class SkyReelsV2DiffusionForcingVideoToVideoPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin): + """ + Pipeline for Video-to-Video (v2v) generation using SkyReels-V2 with diffusion forcing. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a specific device, etc.). + + Args: + tokenizer ([`AutoTokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`UMT5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`SkyReelsV2Transformer3DModel`]): + Conditional Transformer to denoise the encoded image latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: T5EncoderModel | UMT5EncoderModel, + transformer: SkyReelsV2Transformer3DModel, + vae: AutoencoderKLWan, + scheduler: UniPCMultistepScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + video=None, + latents=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + overlap_history=None, + num_frames=None, + base_num_frames=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if video is not None and latents is not None: + raise ValueError("Only one of `video` or `latents` should be provided") + + if num_frames > base_num_frames and overlap_history is None: + raise ValueError( + "`overlap_history` is required when `num_frames` exceeds `base_num_frames` to ensure smooth transitions in long video generation. " + "Please specify a value for `overlap_history`. Recommended values are 17 or 37." + ) + + def prepare_latents( + self, + video: torch.Tensor, + batch_size: int = 1, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 97, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + video_latents: torch.Tensor | None = None, + base_latent_num_frames: int | None = None, + overlap_history: int | None = None, + causal_block_size: int | None = None, + overlap_history_latent_frames: int | None = None, + long_video_iter: int | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = ( + (num_frames - 1) // self.vae_scale_factor_temporal + 1 if latents is None else latents.shape[2] + ) + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + if long_video_iter == 0: + prefix_video_latents = [ + retrieve_latents( + self.vae.encode( + vid.unsqueeze(0)[:, :, -overlap_history:] if vid.dim() == 4 else vid[:, :, -overlap_history:] + ), + sample_mode="argmax", + ) + for vid in video + ] + prefix_video_latents = torch.cat(prefix_video_latents, dim=0).to(dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(device, self.vae.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device, self.vae.dtype + ) + prefix_video_latents = (prefix_video_latents - latents_mean) * latents_std + else: + prefix_video_latents = video_latents[:, :, -overlap_history_latent_frames:] + + if prefix_video_latents.shape[2] % causal_block_size != 0: + truncate_len_latents = prefix_video_latents.shape[2] % causal_block_size + logger.warning( + f"The length of prefix video latents is truncated by {truncate_len_latents} frames for the causal block size alignment. " + f"This truncation ensures compatibility with the causal block size, which is required for proper processing. " + f"However, it may slightly affect the continuity of the generated video at the truncation boundary." + ) + prefix_video_latents = prefix_video_latents[:, :, :-truncate_len_latents] + prefix_video_latents_frames = prefix_video_latents.shape[2] + + finished_frame_num = ( + long_video_iter * (base_latent_num_frames - overlap_history_latent_frames) + overlap_history_latent_frames + ) + left_frame_num = num_latent_frames - finished_frame_num + num_latent_frames = min(left_frame_num + overlap_history_latent_frames, base_latent_num_frames) + + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + latent_height, + latent_width, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + return latents, num_latent_frames, prefix_video_latents, prefix_video_latents_frames + + # Copied from diffusers.pipelines.skyreels_v2.pipeline_skyreels_v2_diffusion_forcing.SkyReelsV2DiffusionForcingPipeline.generate_timestep_matrix + def generate_timestep_matrix( + self, + num_latent_frames: int, + step_template: torch.Tensor, + base_num_latent_frames: int, + ar_step: int = 5, + num_pre_ready: int = 0, + causal_block_size: int = 1, + shrink_interval_with_mask: bool = False, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[tuple]]: + """ + This function implements the core diffusion forcing algorithm that creates a coordinated denoising schedule + across temporal frames. It supports both synchronous and asynchronous generation modes: + + **Synchronous Mode** (ar_step=0, causal_block_size=1): + - All frames are denoised simultaneously at each timestep + - Each frame follows the same denoising trajectory: [1000, 800, 600, ..., 0] + - Simpler but may have less temporal consistency for long videos + + **Asynchronous Mode** (ar_step>0, causal_block_size>1): + - Frames are grouped into causal blocks and processed block/chunk-wise + - Each block is denoised in a staggered pattern creating a "denoising wave" + - Earlier blocks are more denoised, later blocks lag behind by ar_step timesteps + - Creates stronger temporal dependencies and better consistency + + Args: + num_latent_frames (int): Total number of latent frames to generate + step_template (torch.Tensor): Base timestep schedule (e.g., [1000, 800, 600, ..., 0]) + base_num_latent_frames (int): Maximum frames the model can process in one forward pass + ar_step (int, optional): Autoregressive step size for temporal lag. + 0 = synchronous, >0 = asynchronous. Defaults to 5. + num_pre_ready (int, optional): + Number of frames already denoised (e.g., from prefix in a video2video task). + Defaults to 0. + causal_block_size (int, optional): Number of frames processed as a causal block. + Defaults to 1. + shrink_interval_with_mask (bool, optional): Whether to optimize processing intervals. + Defaults to False. + + Returns: + tuple containing: + - step_matrix (torch.Tensor): Matrix of timesteps for each frame at each iteration Shape: + [num_iterations, num_latent_frames] + - step_index (torch.Tensor): Index matrix for timestep lookup Shape: [num_iterations, + num_latent_frames] + - step_update_mask (torch.Tensor): Boolean mask indicating which frames to update Shape: + [num_iterations, num_latent_frames] + - valid_interval (list[tuple]): list of (start, end) intervals for each iteration + + Raises: + ValueError: If ar_step is too small for the given configuration + """ + # Initialize lists to store the scheduling matrices and metadata + step_matrix, step_index = [], [] # Will store timestep values and indices for each iteration + update_mask, valid_interval = [], [] # Will store update masks and processing intervals + + # Calculate total number of denoising iterations (add 1 for initial noise state) + num_iterations = len(step_template) + 1 + + # Convert frame counts to block counts for causal processing + # Each block contains causal_block_size frames that are processed together + # E.g.: 25 frames ÷ 5 = 5 blocks total + num_blocks = num_latent_frames // causal_block_size + base_num_blocks = base_num_latent_frames // causal_block_size + + # Validate ar_step is sufficient for the given configuration + # In asynchronous mode, we need enough timesteps to create the staggered pattern + if base_num_blocks < num_blocks: + min_ar_step = len(step_template) / base_num_blocks + if ar_step < min_ar_step: + raise ValueError(f"`ar_step` should be at least {math.ceil(min_ar_step)} in your setting") + + # Extend step_template with boundary values for easier indexing + # 999: dummy value for counter starting from 1 + # 0: final timestep (completely denoised) + step_template = torch.cat( + [ + torch.tensor([999], dtype=torch.int64, device=step_template.device), + step_template.long(), + torch.tensor([0], dtype=torch.int64, device=step_template.device), + ] + ) + + # Initialize the previous row state (tracks denoising progress for each block) + # 0 means not started, num_iterations means fully denoised + pre_row = torch.zeros(num_blocks, dtype=torch.long) + + # Mark pre-ready frames (e.g., from prefix video for a video2video task) as already at final denoising state + if num_pre_ready > 0: + pre_row[: num_pre_ready // causal_block_size] = num_iterations + + # Main loop: Generate denoising schedule until all frames are fully denoised + while not torch.all(pre_row >= (num_iterations - 1)): + # Create new row representing the next denoising step + new_row = torch.zeros(num_blocks, dtype=torch.long) + + # Apply diffusion forcing logic for each block + for i in range(num_blocks): + if i == 0 or pre_row[i - 1] >= ( + num_iterations - 1 + ): # the first frame or the last frame is completely denoised + new_row[i] = pre_row[i] + 1 + else: + # Asynchronous mode: lag behind previous block by ar_step timesteps + # This creates the "diffusion forcing" staggered pattern + new_row[i] = new_row[i - 1] - ar_step + + # Clamp values to valid range [0, num_iterations] + new_row = new_row.clamp(0, num_iterations) + + # Create update mask: True for blocks that need denoising update at this iteration + # Exclude blocks that haven't started (new_row != pre_row) or are finished (new_row != num_iterations) + # Final state example: [False, ..., False, True, True, True, True, True] + # where first 20 frames are done (False) and last 5 frames still need updates (True) + update_mask.append((new_row != pre_row) & (new_row != num_iterations)) + + # Store the iteration state + step_index.append(new_row) # Index into step_template + step_matrix.append(step_template[new_row]) # Actual timestep values + pre_row = new_row # Update for next iteration + + # For videos longer than model capacity, we process in sliding windows + terminal_flag = base_num_blocks + + # Optional optimization: shrink interval based on first update mask + if shrink_interval_with_mask: + idx_sequence = torch.arange(num_blocks, dtype=torch.int64) + update_mask = update_mask[0] + update_mask_idx = idx_sequence[update_mask] + last_update_idx = update_mask_idx[-1].item() + terminal_flag = last_update_idx + 1 + + # Each interval defines which frames to process in the current forward pass + for curr_mask in update_mask: + # Extend terminal flag if current mask has updates beyond current terminal + if terminal_flag < num_blocks and curr_mask[terminal_flag]: + terminal_flag += 1 + # Create interval: [start, end) where start ensures we don't exceed model capacity + valid_interval.append((max(terminal_flag - base_num_blocks, 0), terminal_flag)) + + # Convert lists to tensors for efficient processing + step_update_mask = torch.stack(update_mask, dim=0) + step_index = torch.stack(step_index, dim=0) + step_matrix = torch.stack(step_matrix, dim=0) + + # Each block's schedule is replicated to all frames within that block + if causal_block_size > 1: + # Expand each block to causal_block_size frames + step_update_mask = step_update_mask.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + step_index = step_index.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + step_matrix = step_matrix.unsqueeze(-1).repeat(1, 1, causal_block_size).flatten(1).contiguous() + # Scale intervals from block-level to frame-level + valid_interval = [(s * causal_block_size, e * causal_block_size) for s, e in valid_interval] + + return step_matrix, step_index, step_update_mask, valid_interval + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + video: list[Image.Image], + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 544, + width: int = 960, + num_frames: int = 120, + num_inference_steps: int = 50, + guidance_scale: float = 6.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + overlap_history: int | None = None, + addnoise_condition: float = 0, + base_num_frames: int = 97, + ar_step: int = 0, + causal_block_size: int | None = None, + fps: int = 24, + ): + r""" + The call function to the pipeline for generation. + + Args: + video (`list[Image.Image]`): + The video to guide the video generation. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the video generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the video generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `544`): + The height of the generated video. + width (`int`, defaults to `960`): + The width of the generated video. + num_frames (`int`, defaults to `120`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `6.0`): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. (**6.0 for T2V**, **5.0 for I2V**) + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`SkyReelsV2PipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `512`): + The maximum sequence length of the prompt. + overlap_history (`int`, *optional*, defaults to `None`): + Number of frames to overlap for smooth transitions in long videos. If `None`, the pipeline assumes + short video generation mode, and no overlap is applied. 17 and 37 are recommended to set. + addnoise_condition (`float`, *optional*, defaults to `0`): + This is used to help smooth the long video generation by adding some noise to the clean condition. Too + large noise can cause the inconsistency as well. 20 is a recommended value, and you may try larger + ones, but it is recommended to not exceed 50. + base_num_frames (`int`, *optional*, defaults to `97`): + 97 or 121 | Base frame count (**97 for 540P**, **121 for 720P**) + ar_step (`int`, *optional*, defaults to `0`): + Controls asynchronous inference (0 for synchronous mode) You can set `ar_step=5` to enable asynchronous + inference. When asynchronous inference, `causal_block_size=5` is recommended while it is not supposed + to be set for synchronous generation. Asynchronous inference will take more steps to diffuse the whole + sequence which means it will be SLOWER than synchronous mode. In our experiments, asynchronous + inference may improve the instruction following and visual consistent performance. + causal_block_size (`int`, *optional*, defaults to `None`): + The number of frames in each block/chunk. Recommended when using asynchronous inference (when ar_step > + 0) + fps (`int`, *optional*, defaults to `24`): + Frame rate of the generated video + + Examples: + + Returns: + [`~SkyReelsV2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + height = height or self.transformer.config.sample_height * self.vae_scale_factor_spatial + width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial + num_videos_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + video, + latents, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + overlap_history, + num_frames, + base_num_frames, + ) + + if addnoise_condition > 60: + logger.warning( + f"The value of 'addnoise_condition' is too large ({addnoise_condition}) and may cause inconsistencies in long video generation. A value of 20 is recommended." + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + if latents is None: + video_original = self.video_processor.preprocess_video(video, height=height, width=width).to( + device, dtype=torch.float32 + ) + + if causal_block_size is None: + causal_block_size = self.transformer.config.num_frame_per_block + else: + self.transformer._set_ar_attention(causal_block_size) + + fps_embeds = [fps] * prompt_embeds.shape[0] + fps_embeds = [0 if i == 16 else 1 for i in fps_embeds] + + # Long video generation + accumulated_latents = None + overlap_history_latent_frames = (overlap_history - 1) // self.vae_scale_factor_temporal + 1 + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + base_latent_num_frames = ( + (base_num_frames - 1) // self.vae_scale_factor_temporal + 1 + if base_num_frames is not None + else num_latent_frames + ) + n_iter = ( + 1 + + (num_latent_frames - base_latent_num_frames - 1) + // (base_latent_num_frames - overlap_history_latent_frames) + + 1 + ) + for long_video_iter in range(n_iter): + logger.debug(f"Processing iteration {long_video_iter + 1}/{n_iter} for long video generation...") + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents, current_num_latent_frames, prefix_video_latents, prefix_video_latents_frames = ( + self.prepare_latents( + video_original, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents if long_video_iter == 0 else None, + video_latents=accumulated_latents, # Pass latents directly instead of decoded video + overlap_history=overlap_history, + base_latent_num_frames=base_latent_num_frames, + causal_block_size=causal_block_size, + overlap_history_latent_frames=overlap_history_latent_frames, + long_video_iter=long_video_iter, + ) + ) + + if prefix_video_latents_frames > 0: + latents[:, :, :prefix_video_latents_frames, :, :] = prefix_video_latents.to(transformer_dtype) + + # 4. Prepare sample schedulers and timestep matrix + sample_schedulers = [] + for _ in range(current_num_latent_frames): + sample_scheduler = deepcopy(self.scheduler) + sample_scheduler.set_timesteps(num_inference_steps, device=device) + sample_schedulers.append(sample_scheduler) + step_matrix, _, step_update_mask, valid_interval = self.generate_timestep_matrix( + current_num_latent_frames, + timesteps, + current_num_latent_frames, + ar_step, + prefix_video_latents_frames, + causal_block_size, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(step_matrix) + + with self.progress_bar(total=len(step_matrix)) as progress_bar: + for i, t in enumerate(step_matrix): + if self.interrupt: + continue + + self._current_timestep = t + valid_interval_start, valid_interval_end = valid_interval[i] + latent_model_input = ( + latents[:, :, valid_interval_start:valid_interval_end, :, :].to(transformer_dtype).clone() + ) + timestep = t.expand(latents.shape[0], -1)[:, valid_interval_start:valid_interval_end].clone() + + if addnoise_condition > 0 and valid_interval_start < prefix_video_latents_frames: + noise_factor = 0.001 * addnoise_condition + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] = ( + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] + * (1.0 - noise_factor) + + torch.randn_like( + latent_model_input[:, :, valid_interval_start:prefix_video_latents_frames, :, :] + ) + * noise_factor + ) + timestep[:, valid_interval_start:prefix_video_latents_frames] = addnoise_condition + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + enable_diffusion_forcing=True, + fps=fps_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + enable_diffusion_forcing=True, + fps=fps_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + update_mask_i = step_update_mask[i] + for idx in range(valid_interval_start, valid_interval_end): + if update_mask_i[idx].item(): + latents[:, :, idx, :, :] = sample_schedulers[idx].step( + noise_pred[:, :, idx - valid_interval_start, :, :], + t[idx], + latents[:, :, idx, :, :], + return_dict=False, + )[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(step_matrix) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0 + ): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if accumulated_latents is None: + accumulated_latents = latents + else: + # Keep overlap frames for conditioning but don't include them in final output + accumulated_latents = torch.cat( + [accumulated_latents, latents[:, :, overlap_history_latent_frames:]], dim=2 + ) + + latents = accumulated_latents + + self._current_timestep = None + + # Final decoding step - convert latents to pixels + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video_generated = self.vae.decode(latents, return_dict=False)[0] + video = torch.cat([video_original, video_generated], dim=2) + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SkyReelsV2PipelineOutput(frames=video) diff --git a/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_i2v.py b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_i2v.py new file mode 100644 index 0000000000000000000000000000000000000000..91c09a56fcfbdc97bac39fbc5e1d21e475502732 --- /dev/null +++ b/diffusers/pipelines/skyreels_v2/pipeline_skyreels_v2_i2v.py @@ -0,0 +1,747 @@ +# Copyright 2025 The SkyReels-V2 Team, The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Any, Callable + +import PIL +import regex as re +import torch +from transformers import AutoTokenizer, CLIPProcessor, CLIPVisionModelWithProjection, T5EncoderModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import SkyReelsV2LoraLoaderMixin +from ...models import AutoencoderKLWan, SkyReelsV2Transformer3DModel +from ...schedulers import UniPCMultistepScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import SkyReelsV2PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """\ + Examples: + ```py + >>> import torch + >>> from diffusers import ( + ... SkyReelsV2ImageToVideoPipeline, + ... UniPCMultistepScheduler, + ... AutoencoderKLWan, + ... ) + >>> from diffusers.utils import export_to_video + >>> from PIL import Image + + >>> # Load the pipeline + >>> # Available models: + >>> # - Skywork/SkyReels-V2-I2V-1.3B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-I2V-14B-540P-Diffusers + >>> # - Skywork/SkyReels-V2-I2V-14B-720P-Diffusers + >>> vae = AutoencoderKLWan.from_pretrained( + ... "Skywork/SkyReels-V2-I2V-14B-720P-Diffusers", + ... subfolder="vae", + ... torch_dtype=torch.float32, + ... ) + >>> pipe = SkyReelsV2ImageToVideoPipeline.from_pretrained( + ... "Skywork/SkyReels-V2-I2V-14B-720P-Diffusers", + ... vae=vae, + ... torch_dtype=torch.bfloat16, + ... ) + >>> flow_shift = 5.0 # 8.0 for T2V, 5.0 for I2V + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe = pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + >>> image = Image.open("path/to/image.png") + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... num_inference_steps=50, + ... height=544, + ... width=960, + ... guidance_scale=5.0, # 6.0 for T2V, 5.0 for I2V + ... num_frames=97, + ... ).frames[0] + >>> export_to_video(output, "video.mp4", fps=24, quality=8) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class SkyReelsV2ImageToVideoPipeline(DiffusionPipeline, SkyReelsV2LoraLoaderMixin): + r""" + Pipeline for Image-to-Video (i2v) generation using SkyReels-V2. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + image_encoder ([`CLIPVisionModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModelWithProjection), + specifically the + [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large) + variant. + transformer ([`SkyReelsV2Transformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: T5EncoderModel | UMT5EncoderModel, + image_encoder: CLIPVisionModelWithProjection, + image_processor: CLIPProcessor, + transformer: SkyReelsV2Transformer3DModel, + vae: AutoencoderKLWan, + scheduler: UniPCMultistepScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + image_encoder=image_encoder, + transformer=transformer, + scheduler=scheduler, + image_processor=image_processor, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.image_processor = image_processor + + # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.encode_image + def encode_image( + self, + image: PipelineImageInput, + device: torch.device | None = None, + ): + device = device or self._execution_device + image = self.image_processor(images=image, return_tensors="pt").to(device) + image_embeds = self.image_encoder(**image, output_hidden_states=True) + return image_embeds.hidden_states[-2] + + # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if image is not None and image_embeds is not None: + raise ValueError( + f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to" + " only forward one of the two." + ) + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `prompt_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}") + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + def prepare_latents( + self, + image: PipelineImageInput, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + last_image: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + image = image.unsqueeze(2) + if last_image is None: + video_condition = torch.cat( + [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 1, height, width)], dim=2 + ) + else: + last_image = last_image.unsqueeze(2) + video_condition = torch.cat( + [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 2, height, width), last_image], + dim=2, + ) + video_condition = video_condition.to(device=device, dtype=self.vae.dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + + if isinstance(generator, list): + latent_condition = [ + retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") for _ in generator + ] + latent_condition = torch.cat(latent_condition) + else: + latent_condition = retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") + latent_condition = latent_condition.repeat(batch_size, 1, 1, 1, 1) + + latent_condition = latent_condition.to(dtype) + latent_condition = (latent_condition - latents_mean) * latents_std + + mask_lat_size = torch.ones(batch_size, 1, num_frames, latent_height, latent_width) + + if last_image is None: + mask_lat_size[:, :, list(range(1, num_frames))] = 0 + else: + mask_lat_size[:, :, list(range(1, num_frames - 1))] = 0 + first_frame_mask = mask_lat_size[:, :, 0:1] + first_frame_mask = torch.repeat_interleave(first_frame_mask, dim=2, repeats=self.vae_scale_factor_temporal) + mask_lat_size = torch.concat([first_frame_mask, mask_lat_size[:, :, 1:, :]], dim=2) + mask_lat_size = mask_lat_size.view(batch_size, -1, self.vae_scale_factor_temporal, latent_height, latent_width) + mask_lat_size = mask_lat_size.transpose(1, 2) + mask_lat_size = mask_lat_size.to(latent_condition.device) + + return latents, torch.concat([mask_lat_size, latent_condition], dim=1) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 544, + width: int = 960, + num_frames: int = 97, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + image_embeds: torch.Tensor | None = None, + last_image: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `544`): + The height of the generated video. + width (`int`, defaults to `960`): + The width of the generated video. + num_frames (`int`, defaults to `97`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided, + image embeddings are generated from the `image` input argument. + last_image (`torch.Tensor`, *optional*): + Optional last image for image-to-video conditioning that anchors the end of the generated video. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to `512`): + The maximum sequence length of the prompt. + + Examples: + + Returns: + [`~SkyReelsV2PipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`SkyReelsV2PipelineOutput`] is returned, otherwise a `tuple` is returned + where the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds, + negative_prompt_embeds, + image_embeds, + callback_on_step_end_tensor_inputs, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + # Encode image embedding + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + if image_embeds is None: + if last_image is None: + image_embeds = self.encode_image(image, device) + else: + image_embeds = self.encode_image([image, last_image], device) + image_embeds = image_embeds.repeat(batch_size, 1, 1) + image_embeds = image_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.z_dim + image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.float32) + if last_image is not None: + last_image = self.video_processor.preprocess(last_image, height=height, width=width).to( + device, dtype=torch.float32 + ) + latents, condition = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + last_image, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = torch.cat([latents, condition], dim=1).to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_image=image_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_hidden_states_image=image_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return SkyReelsV2PipelineOutput(frames=video) diff --git a/diffusers/pipelines/stable_audio/__init__.py b/diffusers/pipelines/stable_audio/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..dfdd419ae9914e64c7fdcf7c152ac308b75d75d2 --- /dev/null +++ b/diffusers/pipelines/stable_audio/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["modeling_stable_audio"] = ["StableAudioProjectionModel"] + _import_structure["pipeline_stable_audio"] = ["StableAudioPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + + else: + from .modeling_stable_audio import StableAudioProjectionModel + from .pipeline_stable_audio import StableAudioPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/stable_audio/modeling_stable_audio.py b/diffusers/pipelines/stable_audio/modeling_stable_audio.py new file mode 100644 index 0000000000000000000000000000000000000000..d40269411bc01503be84c6744d5605efb32ab2eb --- /dev/null +++ b/diffusers/pipelines/stable_audio/modeling_stable_audio.py @@ -0,0 +1,156 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from math import pi + +import torch +import torch.nn as nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin +from ...utils import BaseOutput, logging + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableAudioPositionalEmbedding(nn.Module): + """Used for continuous time""" + + def __init__(self, dim: int): + super().__init__() + assert (dim % 2) == 0 + half_dim = dim // 2 + self.weights = nn.Parameter(torch.randn(half_dim)) + + def forward(self, times: torch.Tensor) -> torch.Tensor: + times = times[..., None] + freqs = times * self.weights[None] * 2 * pi + fouriered = torch.cat((freqs.sin(), freqs.cos()), dim=-1) + fouriered = torch.cat((times, fouriered), dim=-1) + return fouriered + + +@dataclass +class StableAudioProjectionModelOutput(BaseOutput): + """ + Args: + Class for StableAudio projection layer's outputs. + text_hidden_states (`torch.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Sequence of hidden-states obtained by linearly projecting the hidden-states for the text encoder. + seconds_start_hidden_states (`torch.Tensor` of shape `(batch_size, 1, hidden_size)`, *optional*): + Sequence of hidden-states obtained by linearly projecting the audio start hidden states. + seconds_end_hidden_states (`torch.Tensor` of shape `(batch_size, 1, hidden_size)`, *optional*): + Sequence of hidden-states obtained by linearly projecting the audio end hidden states. + """ + + text_hidden_states: torch.Tensor | None = None + seconds_start_hidden_states: torch.Tensor | None = None + seconds_end_hidden_states: torch.Tensor | None = None + + +class StableAudioNumberConditioner(nn.Module): + """ + A simple linear projection model to map numbers to a latent space. + + Args: + number_embedding_dim (`int`): + Dimensionality of the number embeddings. + min_value (`int`): + The minimum value of the seconds number conditioning modules. + max_value (`int`): + The maximum value of the seconds number conditioning modules + internal_dim (`int`): + Dimensionality of the intermediate number hidden states. + """ + + def __init__( + self, + number_embedding_dim, + min_value, + max_value, + internal_dim: int | None = 256, + ): + super().__init__() + self.time_positional_embedding = nn.Sequential( + StableAudioPositionalEmbedding(internal_dim), + nn.Linear(in_features=internal_dim + 1, out_features=number_embedding_dim), + ) + + self.number_embedding_dim = number_embedding_dim + self.min_value = min_value + self.max_value = max_value + + def forward( + self, + floats: torch.Tensor, + ): + floats = floats.clamp(self.min_value, self.max_value) + + normalized_floats = (floats - self.min_value) / (self.max_value - self.min_value) + + # Cast floats to same type as embedder + embedder_dtype = next(self.time_positional_embedding.parameters()).dtype + normalized_floats = normalized_floats.to(embedder_dtype) + + embedding = self.time_positional_embedding(normalized_floats) + float_embeds = embedding.view(-1, 1, self.number_embedding_dim) + + return float_embeds + + +class StableAudioProjectionModel(ModelMixin, ConfigMixin): + """ + A simple linear projection model to map the conditioning values to a shared latent space. + + Args: + text_encoder_dim (`int`): + Dimensionality of the text embeddings from the text encoder (T5). + conditioning_dim (`int`): + Dimensionality of the output conditioning tensors. + min_value (`int`): + The minimum value of the seconds number conditioning modules. + max_value (`int`): + The maximum value of the seconds number conditioning modules + """ + + @register_to_config + def __init__(self, text_encoder_dim, conditioning_dim, min_value, max_value): + super().__init__() + self.text_projection = ( + nn.Identity() if conditioning_dim == text_encoder_dim else nn.Linear(text_encoder_dim, conditioning_dim) + ) + self.start_number_conditioner = StableAudioNumberConditioner(conditioning_dim, min_value, max_value) + self.end_number_conditioner = StableAudioNumberConditioner(conditioning_dim, min_value, max_value) + + def forward( + self, + text_hidden_states: torch.Tensor | None = None, + start_seconds: torch.Tensor | None = None, + end_seconds: torch.Tensor | None = None, + ): + text_hidden_states = ( + text_hidden_states if text_hidden_states is None else self.text_projection(text_hidden_states) + ) + seconds_start_hidden_states = ( + start_seconds if start_seconds is None else self.start_number_conditioner(start_seconds) + ) + seconds_end_hidden_states = end_seconds if end_seconds is None else self.end_number_conditioner(end_seconds) + + return StableAudioProjectionModelOutput( + text_hidden_states=text_hidden_states, + seconds_start_hidden_states=seconds_start_hidden_states, + seconds_end_hidden_states=seconds_end_hidden_states, + ) diff --git a/diffusers/pipelines/stable_audio/pipeline_stable_audio.py b/diffusers/pipelines/stable_audio/pipeline_stable_audio.py new file mode 100644 index 0000000000000000000000000000000000000000..475f4032edabb2072699a61073177a802574a815 --- /dev/null +++ b/diffusers/pipelines/stable_audio/pipeline_stable_audio.py @@ -0,0 +1,736 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import torch +from transformers import ( + T5EncoderModel, + T5Tokenizer, + T5TokenizerFast, +) + +from ...models import AutoencoderOobleck, StableAudioDiTModel +from ...models.embeddings import get_1d_rotary_pos_embed +from ...schedulers import EDMDPMSolverMultistepScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline +from .modeling_stable_audio import StableAudioProjectionModel + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import scipy + >>> import torch + >>> import soundfile as sf + >>> from diffusers import StableAudioPipeline + + >>> repo_id = "stabilityai/stable-audio-open-1.0" + >>> pipe = StableAudioPipeline.from_pretrained(repo_id, torch_dtype=torch.float16) + >>> pipe = pipe.to("cuda") + + >>> # define the prompts + >>> prompt = "The sound of a hammer hitting a wooden surface." + >>> negative_prompt = "Low quality." + + >>> # set the seed for generator + >>> generator = torch.Generator("cuda").manual_seed(0) + + >>> # run the generation + >>> audio = pipe( + ... prompt, + ... negative_prompt=negative_prompt, + ... num_inference_steps=200, + ... audio_end_in_s=10.0, + ... num_waveforms_per_prompt=3, + ... generator=generator, + ... ).audios + + >>> output = audio[0].T.float().cpu().numpy() + >>> sf.write("hammer.wav", output, pipe.vae.sampling_rate) + ``` +""" + + +class StableAudioPipeline(DiffusionPipeline): + r""" + Pipeline for text-to-audio generation using StableAudio. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderOobleck`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.T5EncoderModel`]): + Frozen text-encoder. StableAudio uses the encoder of + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [google-t5/t5-base](https://huggingface.co/google-t5/t5-base) variant. + projection_model ([`StableAudioProjectionModel`]): + A trained model used to linearly project the hidden-states from the text encoder model and the start and + end seconds. The projected hidden-states from the encoder and the conditional seconds are concatenated to + give the input to the transformer model. + tokenizer ([`~transformers.T5Tokenizer`]): + Tokenizer to tokenize text for the frozen text-encoder. + transformer ([`StableAudioDiTModel`]): + A `StableAudioDiTModel` to denoise the encoded audio latents. + scheduler ([`EDMDPMSolverMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded audio latents. + """ + + model_cpu_offload_seq = "text_encoder->projection_model->transformer->vae" + + def __init__( + self, + vae: AutoencoderOobleck, + text_encoder: T5EncoderModel, + projection_model: StableAudioProjectionModel, + tokenizer: T5Tokenizer | T5TokenizerFast, + transformer: StableAudioDiTModel, + scheduler: EDMDPMSolverMultistepScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + projection_model=projection_model, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + self.rotary_embed_dim = self.transformer.config.attention_head_dim // 2 + + def encode_prompt( + self, + prompt, + device, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + attention_mask: torch.LongTensor | None = None, + negative_attention_mask: torch.LongTensor | None = None, + ): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # 1. Tokenize text + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + f"The following part of your input was truncated because {self.text_encoder.config.model_type} can " + f"only handle sequences up to {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_input_ids = text_input_ids.to(device) + attention_mask = attention_mask.to(device) + + # 2. Text encoder forward + self.text_encoder.eval() + prompt_embeds = self.text_encoder( + text_input_ids, + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + + if do_classifier_free_guidance and negative_prompt is not None: + uncond_tokens: list[str] + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # 1. Tokenize text + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + uncond_input_ids = uncond_input.input_ids.to(device) + negative_attention_mask = uncond_input.attention_mask.to(device) + + # 2. Text encoder forward + self.text_encoder.eval() + negative_prompt_embeds = self.text_encoder( + uncond_input_ids, + attention_mask=negative_attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if negative_attention_mask is not None: + # set the masked tokens to the null embed + negative_prompt_embeds = torch.where( + negative_attention_mask.to(torch.bool).unsqueeze(2), negative_prompt_embeds, 0.0 + ) + + # 3. Project prompt_embeds and negative_prompt_embeds + if do_classifier_free_guidance and negative_prompt_embeds is not None: + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the negative and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + if attention_mask is not None and negative_attention_mask is None: + negative_attention_mask = torch.ones_like(attention_mask) + elif attention_mask is None and negative_attention_mask is not None: + attention_mask = torch.ones_like(negative_attention_mask) + + if attention_mask is not None: + attention_mask = torch.cat([negative_attention_mask, attention_mask]) + + prompt_embeds = self.projection_model( + text_hidden_states=prompt_embeds, + ).text_hidden_states + if attention_mask is not None: + prompt_embeds = prompt_embeds * attention_mask.unsqueeze(-1).to(prompt_embeds.dtype) + prompt_embeds = prompt_embeds * attention_mask.unsqueeze(-1).to(prompt_embeds.dtype) + + return prompt_embeds + + def encode_duration( + self, + audio_start_in_s, + audio_end_in_s, + device, + do_classifier_free_guidance, + batch_size, + ): + audio_start_in_s = audio_start_in_s if isinstance(audio_start_in_s, list) else [audio_start_in_s] + audio_end_in_s = audio_end_in_s if isinstance(audio_end_in_s, list) else [audio_end_in_s] + + if len(audio_start_in_s) == 1: + audio_start_in_s = audio_start_in_s * batch_size + if len(audio_end_in_s) == 1: + audio_end_in_s = audio_end_in_s * batch_size + + # Cast the inputs to floats + audio_start_in_s = [float(x) for x in audio_start_in_s] + audio_start_in_s = torch.tensor(audio_start_in_s).to(device) + + audio_end_in_s = [float(x) for x in audio_end_in_s] + audio_end_in_s = torch.tensor(audio_end_in_s).to(device) + + projection_output = self.projection_model( + start_seconds=audio_start_in_s, + end_seconds=audio_end_in_s, + ) + seconds_start_hidden_states = projection_output.seconds_start_hidden_states + seconds_end_hidden_states = projection_output.seconds_end_hidden_states + + # For classifier free guidance, we need to do two forward passes. + # Here we repeat the audio hidden states to avoid doing two forward passes + if do_classifier_free_guidance: + seconds_start_hidden_states = torch.cat([seconds_start_hidden_states, seconds_start_hidden_states], dim=0) + seconds_end_hidden_states = torch.cat([seconds_end_hidden_states, seconds_end_hidden_states], dim=0) + + return seconds_start_hidden_states, seconds_end_hidden_states + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + audio_start_in_s, + audio_end_in_s, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + attention_mask=None, + negative_attention_mask=None, + initial_audio_waveforms=None, + initial_audio_sampling_rate=None, + ): + if audio_end_in_s < audio_start_in_s: + raise ValueError( + f"`audio_end_in_s={audio_end_in_s}' must be higher than 'audio_start_in_s={audio_start_in_s}` but " + ) + + if ( + audio_start_in_s < self.projection_model.config.min_value + or audio_start_in_s > self.projection_model.config.max_value + ): + raise ValueError( + f"`audio_start_in_s` must be greater than or equal to {self.projection_model.config.min_value}, and lower than or equal to {self.projection_model.config.max_value} but " + f"is {audio_start_in_s}." + ) + + if ( + audio_end_in_s < self.projection_model.config.min_value + or audio_end_in_s > self.projection_model.config.max_value + ): + raise ValueError( + f"`audio_end_in_s` must be greater than or equal to {self.projection_model.config.min_value}, and lower than or equal to {self.projection_model.config.max_value} but " + f"is {audio_end_in_s}." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and (prompt_embeds is None): + raise ValueError( + "Provide either `prompt`, or `prompt_embeds`. Cannot leave" + "`prompt` undefined without specifying `prompt_embeds`." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if attention_mask is not None and attention_mask.shape != prompt_embeds.shape[:2]: + raise ValueError( + "`attention_mask should have the same batch size and sequence length as `prompt_embeds`, but got:" + f"`attention_mask: {attention_mask.shape} != `prompt_embeds` {prompt_embeds.shape}" + ) + + if initial_audio_sampling_rate is None and initial_audio_waveforms is not None: + raise ValueError( + "`initial_audio_waveforms' is provided but the sampling rate is not. Make sure to pass `initial_audio_sampling_rate`." + ) + + if initial_audio_sampling_rate is not None and initial_audio_sampling_rate != self.vae.sampling_rate: + raise ValueError( + f"`initial_audio_sampling_rate` must be {self.vae.hop_length}' but is `{initial_audio_sampling_rate}`." + "Make sure to resample the `initial_audio_waveforms` and to correct the sampling rate. " + ) + + def prepare_latents( + self, + batch_size, + num_channels_vae, + sample_size, + dtype, + device, + generator, + latents=None, + initial_audio_waveforms=None, + num_waveforms_per_prompt=None, + audio_channels=None, + ): + shape = (batch_size, num_channels_vae, sample_size) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + + # encode the initial audio for use by the model + if initial_audio_waveforms is not None: + # check dimension + if initial_audio_waveforms.ndim == 2: + initial_audio_waveforms = initial_audio_waveforms.unsqueeze(1) + elif initial_audio_waveforms.ndim != 3: + raise ValueError( + f"`initial_audio_waveforms` must be of shape `(batch_size, num_channels, audio_length)` or `(batch_size, audio_length)` but has `{initial_audio_waveforms.ndim}` dimensions" + ) + + audio_vae_length = int(self.transformer.config.sample_size) * self.vae.hop_length + audio_shape = (batch_size // num_waveforms_per_prompt, audio_channels, audio_vae_length) + + # check num_channels + if initial_audio_waveforms.shape[1] == 1 and audio_channels == 2: + initial_audio_waveforms = initial_audio_waveforms.repeat(1, 2, 1) + elif initial_audio_waveforms.shape[1] == 2 and audio_channels == 1: + initial_audio_waveforms = initial_audio_waveforms.mean(1, keepdim=True) + + if initial_audio_waveforms.shape[:2] != audio_shape[:2]: + raise ValueError( + f"`initial_audio_waveforms` must be of shape `(batch_size, num_channels, audio_length)` or `(batch_size, audio_length)` but is of shape `{initial_audio_waveforms.shape}`" + ) + + # crop or pad + audio_length = initial_audio_waveforms.shape[-1] + if audio_length < audio_vae_length: + logger.warning( + f"The provided input waveform is shorter ({audio_length}) than the required audio length ({audio_vae_length}) of the model and will thus be padded." + ) + elif audio_length > audio_vae_length: + logger.warning( + f"The provided input waveform is longer ({audio_length}) than the required audio length ({audio_vae_length}) of the model and will thus be cropped." + ) + + audio = initial_audio_waveforms.new_zeros(audio_shape) + audio[:, :, : min(audio_length, audio_vae_length)] = initial_audio_waveforms[:, :, :audio_vae_length] + + encoded_audio = self.vae.encode(audio).latent_dist.sample(generator) + encoded_audio = encoded_audio.repeat((num_waveforms_per_prompt, 1, 1)) + latents = encoded_audio + latents + return latents + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + audio_end_in_s: float | None = None, + audio_start_in_s: float | None = 0.0, + num_inference_steps: int = 100, + guidance_scale: float = 7.0, + negative_prompt: str | list[str] | None = None, + num_waveforms_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + initial_audio_waveforms: torch.Tensor | None = None, + initial_audio_sampling_rate: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + attention_mask: torch.LongTensor | None = None, + negative_attention_mask: torch.LongTensor | None = None, + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int | None = 1, + output_type: str | None = "pt", + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide audio generation. If not defined, you need to pass `prompt_embeds`. + audio_end_in_s (`float`, *optional*, defaults to 47.55): + Audio end index in seconds. + audio_start_in_s (`float`, *optional*, defaults to 0): + Audio start index in seconds. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality audio at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.0): + A higher guidance scale value encourages the model to generate audio that is closely linked to the text + `prompt` at the expense of lower sound quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in audio generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_waveforms_per_prompt (`int`, *optional*, defaults to 1): + The number of waveforms to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for audio + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + initial_audio_waveforms (`torch.Tensor`, *optional*): + Optional initial audio waveforms to use as the initial audio waveform for generation. Must be of shape + `(batch_size, num_channels, audio_length)` or `(batch_size, audio_length)`, where `batch_size` + corresponds to the number of prompts passed to the model. + initial_audio_sampling_rate (`int`, *optional*): + Sampling rate of the `initial_audio_waveforms`, if they are provided. Must be the same as the model. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-computed text embeddings from the text encoder model. Can be used to easily tweak text inputs, + *e.g.* prompt weighting. If not provided, text embeddings will be computed from `prompt` input + argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-computed negative text embeddings from the text encoder model. Can be used to easily tweak text + inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be computed from + `negative_prompt` input argument. + attention_mask (`torch.LongTensor`, *optional*): + Pre-computed attention mask to be applied to the `prompt_embeds`. If not provided, attention mask will + be computed from `prompt` input argument. + negative_attention_mask (`torch.LongTensor`, *optional*): + Pre-computed attention mask to be applied to the `negative_text_audio_duration_embeds`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + output_type (`str`, *optional*, defaults to `"pt"`): + The output format of the generated audio. Choose between `"np"` to return a NumPy `np.ndarray` or + `"pt"` to return a PyTorch `torch.Tensor` object. Set to `"latent"` to return the latent diffusion + model (LDM) output. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated audio. + """ + # 0. Convert audio input length from seconds to latent length + downsample_ratio = self.vae.hop_length + + max_audio_length_in_s = self.transformer.config.sample_size * downsample_ratio / self.vae.config.sampling_rate + if audio_end_in_s is None: + audio_end_in_s = max_audio_length_in_s + + if audio_end_in_s - audio_start_in_s > max_audio_length_in_s: + raise ValueError( + f"The total audio length requested ({audio_end_in_s - audio_start_in_s}s) is longer than the model maximum possible length ({max_audio_length_in_s}). Make sure that 'audio_end_in_s-audio_start_in_s<={max_audio_length_in_s}'." + ) + + waveform_start = int(audio_start_in_s * self.vae.config.sampling_rate) + waveform_end = int(audio_end_in_s * self.vae.config.sampling_rate) + waveform_length = int(self.transformer.config.sample_size) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + audio_start_in_s, + audio_end_in_s, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + attention_mask, + negative_attention_mask, + initial_audio_waveforms, + initial_audio_sampling_rate, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + prompt_embeds = self.encode_prompt( + prompt, + device, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + attention_mask, + negative_attention_mask, + ) + + # Encode duration + seconds_start_hidden_states, seconds_end_hidden_states = self.encode_duration( + audio_start_in_s, + audio_end_in_s, + device, + do_classifier_free_guidance and (negative_prompt is not None or negative_prompt_embeds is not None), + batch_size, + ) + + # Create text_audio_duration_embeds and audio_duration_embeds + text_audio_duration_embeds = torch.cat( + [prompt_embeds, seconds_start_hidden_states, seconds_end_hidden_states], dim=1 + ) + + audio_duration_embeds = torch.cat([seconds_start_hidden_states, seconds_end_hidden_states], dim=2) + + # In case of classifier free guidance without negative prompt, we need to create unconditional embeddings and + # to concatenate it to the embeddings + if do_classifier_free_guidance and negative_prompt_embeds is None and negative_prompt is None: + negative_text_audio_duration_embeds = torch.zeros_like( + text_audio_duration_embeds, device=text_audio_duration_embeds.device + ) + text_audio_duration_embeds = torch.cat( + [negative_text_audio_duration_embeds, text_audio_duration_embeds], dim=0 + ) + audio_duration_embeds = torch.cat([audio_duration_embeds, audio_duration_embeds], dim=0) + + bs_embed, seq_len, hidden_size = text_audio_duration_embeds.shape + # duplicate audio_duration_embeds and text_audio_duration_embeds for each generation per prompt, using mps friendly method + text_audio_duration_embeds = text_audio_duration_embeds.repeat(1, num_waveforms_per_prompt, 1) + text_audio_duration_embeds = text_audio_duration_embeds.view( + bs_embed * num_waveforms_per_prompt, seq_len, hidden_size + ) + + audio_duration_embeds = audio_duration_embeds.repeat(1, num_waveforms_per_prompt, 1) + audio_duration_embeds = audio_duration_embeds.view( + bs_embed * num_waveforms_per_prompt, -1, audio_duration_embeds.shape[-1] + ) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_vae = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_waveforms_per_prompt, + num_channels_vae, + waveform_length, + text_audio_duration_embeds.dtype, + device, + generator, + latents, + initial_audio_waveforms, + num_waveforms_per_prompt, + audio_channels=self.vae.config.audio_channels, + ) + + # 6. Prepare extra step kwargs + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare rotary positional embedding + rotary_embedding = get_1d_rotary_pos_embed( + self.rotary_embed_dim, + latents.shape[2] + audio_duration_embeds.shape[1], + use_real=True, + repeat_interleave_real=False, + ) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.transformer( + latent_model_input, + t.unsqueeze(0), + encoder_hidden_states=text_audio_duration_embeds, + global_hidden_states=audio_duration_embeds, + rotary_embedding=rotary_embedding, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing + if not output_type == "latent": + audio = self.vae.decode(latents).sample + else: + return AudioPipelineOutput(audios=latents) + + audio = audio[:, :, waveform_start:waveform_end] + + if output_type == "np": + audio = audio.cpu().float().numpy() + + self.maybe_free_model_hooks() + + if not return_dict: + return (audio,) + + return AudioPipelineOutput(audios=audio) diff --git a/diffusers/pipelines/stable_cascade/__init__.py b/diffusers/pipelines/stable_cascade/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5270cb94af01fd94ed6c8e76c243a86cad8ec348 --- /dev/null +++ b/diffusers/pipelines/stable_cascade/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_cascade"] = ["StableCascadeDecoderPipeline"] + _import_structure["pipeline_stable_cascade_combined"] = ["StableCascadeCombinedPipeline"] + _import_structure["pipeline_stable_cascade_prior"] = ["StableCascadePriorPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_cascade import StableCascadeDecoderPipeline + from .pipeline_stable_cascade_combined import StableCascadeCombinedPipeline + from .pipeline_stable_cascade_prior import StableCascadePriorPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/stable_cascade/pipeline_stable_cascade.py b/diffusers/pipelines/stable_cascade/pipeline_stable_cascade.py new file mode 100644 index 0000000000000000000000000000000000000000..0961d1c46e94a7d9ab415d7c2f6b8341e0203e54 --- /dev/null +++ b/diffusers/pipelines/stable_cascade/pipeline_stable_cascade.py @@ -0,0 +1,541 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Callable + +import torch +from transformers import CLIPTextModelWithProjection, CLIPTokenizer + +from ...models import StableCascadeUNet +from ...schedulers import DDPMWuerstchenScheduler +from ...utils import is_torch_version, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..deprecated.wuerstchen.modeling_paella_vq_model import PaellaVQModel +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableCascadePriorPipeline, StableCascadeDecoderPipeline + + >>> prior_pipe = StableCascadePriorPipeline.from_pretrained( + ... "stabilityai/stable-cascade-prior", torch_dtype=torch.bfloat16 + ... ).to("cuda") + >>> gen_pipe = StableCascadeDecoderPipeline.from_pretrain( + ... "stabilityai/stable-cascade", torch_dtype=torch.float16 + ... ).to("cuda") + + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> prior_output = pipe(prompt) + >>> images = gen_pipe(prior_output.image_embeddings, prompt=prompt) + ``` +""" + + +class StableCascadeDecoderPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline for generating images from the Stable Cascade model. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + tokenizer (`CLIPTokenizer`): + The CLIP tokenizer. + text_encoder (`CLIPTextModelWithProjection`): + The CLIP text encoder. + decoder ([`StableCascadeUNet`]): + The Stable Cascade decoder unet. + vqgan ([`PaellaVQModel`]): + The VQGAN model. + scheduler ([`DDPMWuerstchenScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + latent_dim_scale (float, `optional`, defaults to 10.67): + Multiplier to determine the VQ latent space size from the image embeddings. If the image embeddings are + height=24 and width=24, the VQ latent shape needs to be height=int(24*10.67)=256 and + width=int(24*10.67)=256 in order to match the training conditions. + """ + + _last_supported_version = "0.35.2" + + unet_name = "decoder" + text_encoder_name = "text_encoder" + model_cpu_offload_seq = "text_encoder->decoder->vqgan" + _callback_tensor_inputs = [ + "latents", + "prompt_embeds_pooled", + "negative_prompt_embeds", + "image_embeddings", + ] + + def __init__( + self, + decoder: StableCascadeUNet, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModelWithProjection, + scheduler: DDPMWuerstchenScheduler, + vqgan: PaellaVQModel, + latent_dim_scale: float = 10.67, + ) -> None: + super().__init__() + self.register_modules( + decoder=decoder, + tokenizer=tokenizer, + text_encoder=text_encoder, + scheduler=scheduler, + vqgan=vqgan, + ) + self.register_to_config(latent_dim_scale=latent_dim_scale) + + def prepare_latents( + self, batch_size, image_embeddings, num_images_per_prompt, dtype, device, generator, latents, scheduler + ): + _, channels, height, width = image_embeddings.shape + latents_shape = ( + batch_size * num_images_per_prompt, + 4, + int(height * self.config.latent_dim_scale), + int(width * self.config.latent_dim_scale), + ) + + if latents is None: + latents = randn_tensor(latents_shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def encode_prompt( + self, + device, + batch_size, + num_images_per_prompt, + do_classifier_free_guidance, + prompt=None, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_pooled: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_pooled: torch.Tensor | None = None, + ): + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + attention_mask = attention_mask[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask.to(device), output_hidden_states=True + ) + prompt_embeds = text_encoder_output.hidden_states[-1] + if prompt_embeds_pooled is None: + prompt_embeds_pooled = text_encoder_output.text_embeds.unsqueeze(1) + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds_pooled = prompt_embeds_pooled.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + prompt_embeds_pooled = prompt_embeds_pooled.repeat_interleave(num_images_per_prompt, dim=0) + + if negative_prompt_embeds is None and do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds_text_encoder_output = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=uncond_input.attention_mask.to(device), + output_hidden_states=True, + ) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.hidden_states[-1] + negative_prompt_embeds_pooled = negative_prompt_embeds_text_encoder_output.text_embeds.unsqueeze(1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + seq_len = negative_prompt_embeds_pooled.shape[1] + negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.to( + dtype=self.text_encoder.dtype, device=device + ) + negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + # done duplicates + + return prompt_embeds, prompt_embeds_pooled, negative_prompt_embeds, negative_prompt_embeds_pooled + + def check_inputs( + self, + prompt, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + def get_timestep_ratio_conditioning(self, t, alphas_cumprod): + s = torch.tensor([0.008]) + clamp_range = [0, 1] + min_var = torch.cos(s / (1 + s) * torch.pi * 0.5) ** 2 + var = alphas_cumprod[t] + var = var.clamp(*clamp_range) + s, min_var = s.to(var.device), min_var.to(var.device) + ratio = (((var * min_var) ** 0.5).acos() / (torch.pi * 0.5)) * (1 + s) - s + return ratio + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image_embeddings: torch.Tensor | list[torch.Tensor], + prompt: str | list[str] = None, + num_inference_steps: int = 10, + guidance_scale: float = 0.0, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_pooled: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_pooled: torch.Tensor | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + image_embeddings (`torch.Tensor` or `list[torch.Tensor]`): + Image Embeddings either extracted from an image or generated by a Prior Model. + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + num_inference_steps (`int`, *optional*, defaults to 12): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 0.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `decoder_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `decoder_guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_pooled (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_pooled (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds_pooled will be generated from `negative_prompt` + input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is a list with the generated image + embeddings. + """ + + # 0. Define commonly used variables + device = self._execution_device + dtype = self.decoder.dtype + self._guidance_scale = guidance_scale + if is_torch_version("<", "2.2.0") and dtype == torch.bfloat16: + raise ValueError("`StableCascadeDecoderPipeline` requires torch>=2.2.0 when using `torch.bfloat16` dtype.") + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + if isinstance(image_embeddings, list): + image_embeddings = torch.cat(image_embeddings, dim=0) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Compute the effective number of images per prompt + # We must account for the fact that the image embeddings from the prior can be generated with num_images_per_prompt > 1 + # This results in a case where a single prompt is associated with multiple image embeddings + # Divide the number of image embeddings by the batch size to determine if this is the case. + num_images_per_prompt = num_images_per_prompt * (image_embeddings.shape[0] // batch_size) + + # 2. Encode caption + if prompt_embeds is None and negative_prompt_embeds is None: + _, prompt_embeds_pooled, _, negative_prompt_embeds_pooled = self.encode_prompt( + prompt=prompt, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_pooled=prompt_embeds_pooled, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_pooled=negative_prompt_embeds_pooled, + ) + + # The pooled embeds from the prior are pooled again before being passed to the decoder + prompt_embeds_pooled = ( + torch.cat([prompt_embeds_pooled, negative_prompt_embeds_pooled]) + if self.do_classifier_free_guidance + else prompt_embeds_pooled + ) + effnet = ( + torch.cat([image_embeddings, torch.zeros_like(image_embeddings)]) + if self.do_classifier_free_guidance + else image_embeddings + ) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + latents = self.prepare_latents( + batch_size, image_embeddings, num_images_per_prompt, dtype, device, generator, latents, self.scheduler + ) + + if isinstance(self.scheduler, DDPMWuerstchenScheduler): + timesteps = timesteps[:-1] + else: + if hasattr(self.scheduler.config, "clip_sample") and self.scheduler.config.clip_sample: + self.scheduler.config.clip_sample = False # disample sample clipping + logger.warning(" set `clip_sample` to be False") + + # 6. Run denoising loop + if hasattr(self.scheduler, "betas"): + alphas = 1.0 - self.scheduler.betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + else: + alphas_cumprod = [] + + self._num_timesteps = len(timesteps) + for i, t in enumerate(self.progress_bar(timesteps)): + if not isinstance(self.scheduler, DDPMWuerstchenScheduler): + if len(alphas_cumprod) > 0: + timestep_ratio = self.get_timestep_ratio_conditioning(t.long().cpu(), alphas_cumprod) + timestep_ratio = timestep_ratio.expand(latents.size(0)).to(dtype).to(device) + else: + timestep_ratio = t.float().div(self.scheduler.timesteps[-1]).expand(latents.size(0)).to(dtype) + else: + timestep_ratio = t.expand(latents.size(0)).to(dtype) + + # 7. Denoise latents + predicted_latents = self.decoder( + sample=torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, + timestep_ratio=torch.cat([timestep_ratio] * 2) if self.do_classifier_free_guidance else timestep_ratio, + clip_text_pooled=prompt_embeds_pooled, + effnet=effnet, + return_dict=False, + )[0] + + # 8. Check for classifier free guidance and apply it + if self.do_classifier_free_guidance: + predicted_latents_text, predicted_latents_uncond = predicted_latents.chunk(2) + predicted_latents = torch.lerp(predicted_latents_uncond, predicted_latents_text, self.guidance_scale) + + # 9. Renoise latents to next timestep + if not isinstance(self.scheduler, DDPMWuerstchenScheduler): + timestep_ratio = t + latents = self.scheduler.step( + model_output=predicted_latents, + timestep=timestep_ratio, + sample=latents, + generator=generator, + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type not in ["pt", "np", "pil", "latent"]: + raise ValueError( + f"Only the output types `pt`, `np`, `pil` and `latent` are supported not output_type={output_type}" + ) + + if not output_type == "latent": + # 10. Scale and decode the image latents with vq-vae + latents = self.vqgan.config.scale_factor * latents + images = self.vqgan.decode(latents).sample.clamp(0, 1) + if output_type == "np": + images = images.permute(0, 2, 3, 1).cpu().float().numpy() # float() as bfloat16-> numpy doesn't work + elif output_type == "pil": + images = images.permute(0, 2, 3, 1).cpu().float().numpy() # float() as bfloat16-> numpy doesn't work + images = self.numpy_to_pil(images) + else: + images = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return images + return ImagePipelineOutput(images) diff --git a/diffusers/pipelines/stable_cascade/pipeline_stable_cascade_combined.py b/diffusers/pipelines/stable_cascade/pipeline_stable_cascade_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..71bfc3f7bab653be460b79e2bd9f30362720ca25 --- /dev/null +++ b/diffusers/pipelines/stable_cascade/pipeline_stable_cascade_combined.py @@ -0,0 +1,321 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Callable + +import PIL +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...models import StableCascadeUNet +from ...schedulers import DDPMWuerstchenScheduler +from ...utils import is_torch_version, replace_example_docstring +from ..deprecated.wuerstchen.modeling_paella_vq_model import PaellaVQModel +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline +from .pipeline_stable_cascade import StableCascadeDecoderPipeline +from .pipeline_stable_cascade_prior import StableCascadePriorPipeline + + +TEXT2IMAGE_EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableCascadeCombinedPipeline + + >>> pipe = StableCascadeCombinedPipeline.from_pretrained( + ... "stabilityai/stable-cascade", variant="bf16", torch_dtype=torch.bfloat16 + ... ) + >>> pipe.enable_model_cpu_offload() + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> images = pipe(prompt=prompt) + ``` +""" + + +class StableCascadeCombinedPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Combined Pipeline for text-to-image generation using Stable Cascade. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + tokenizer (`CLIPTokenizer`): + The decoder tokenizer to be used for text inputs. + text_encoder (`CLIPTextModelWithProjection`): + The decoder text encoder to be used for text inputs. + decoder (`StableCascadeUNet`): + The decoder model to be used for decoder image generation pipeline. + scheduler (`DDPMWuerstchenScheduler`): + The scheduler to be used for decoder image generation pipeline. + vqgan (`PaellaVQModel`): + The VQGAN model to be used for decoder image generation pipeline. + prior_prior (`StableCascadeUNet`): + The prior model to be used for prior pipeline. + prior_text_encoder (`CLIPTextModelWithProjection`): + The prior text encoder to be used for text inputs. + prior_tokenizer (`CLIPTokenizer`): + The prior tokenizer to be used for text inputs. + prior_scheduler (`DDPMWuerstchenScheduler`): + The scheduler to be used for prior pipeline. + prior_feature_extractor ([`~transformers.CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `image_encoder`. + prior_image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + """ + + _last_supported_version = "0.35.2" + + _load_connected_pipes = True + _optional_components = ["prior_feature_extractor", "prior_image_encoder"] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModelWithProjection, + decoder: StableCascadeUNet, + scheduler: DDPMWuerstchenScheduler, + vqgan: PaellaVQModel, + prior_prior: StableCascadeUNet, + prior_text_encoder: CLIPTextModelWithProjection, + prior_tokenizer: CLIPTokenizer, + prior_scheduler: DDPMWuerstchenScheduler, + prior_feature_extractor: CLIPImageProcessor | None = None, + prior_image_encoder: CLIPVisionModelWithProjection | None = None, + ): + super().__init__() + + self.register_modules( + text_encoder=text_encoder, + tokenizer=tokenizer, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + prior_text_encoder=prior_text_encoder, + prior_tokenizer=prior_tokenizer, + prior_prior=prior_prior, + prior_scheduler=prior_scheduler, + prior_feature_extractor=prior_feature_extractor, + prior_image_encoder=prior_image_encoder, + ) + self.prior_pipe = StableCascadePriorPipeline( + prior=prior_prior, + text_encoder=prior_text_encoder, + tokenizer=prior_tokenizer, + scheduler=prior_scheduler, + image_encoder=prior_image_encoder, + feature_extractor=prior_feature_extractor, + ) + self.decoder_pipe = StableCascadeDecoderPipeline( + text_encoder=text_encoder, + tokenizer=tokenizer, + decoder=decoder, + scheduler=scheduler, + vqgan=vqgan, + ) + + def enable_xformers_memory_efficient_attention(self, attention_op: Callable | None = None): + self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op) + + def enable_model_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id, device=device) + + def enable_sequential_cpu_offload(self, gpu_id: int | None = None, device: torch.device | str = None): + r""" + Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗 + Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a + GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis. + Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower. + """ + self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id, device=device) + + def progress_bar(self, iterable=None, total=None): + self.prior_pipe.progress_bar(iterable=iterable, total=total) + self.decoder_pipe.progress_bar(iterable=iterable, total=total) + + def set_progress_bar_config(self, **kwargs): + self.prior_pipe.set_progress_bar_config(**kwargs) + self.decoder_pipe.set_progress_bar_config(**kwargs) + + @torch.no_grad() + @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + images: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image] = None, + height: int = 512, + width: int = 512, + prior_num_inference_steps: int = 60, + prior_guidance_scale: float = 4.0, + num_inference_steps: int = 12, + decoder_guidance_scale: float = 0.0, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_pooled: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_pooled: torch.Tensor | None = None, + num_images_per_prompt: int = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + prior_callback_on_step_end: Callable[[int, int], None] | None = None, + prior_callback_on_step_end_tensor_inputs: list[str] = ["latents"], + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation for the prior and decoder. + images (`torch.Tensor`, `PIL.Image.Image`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, *optional*): + The images to guide the image generation for the prior. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_pooled (`torch.Tensor`, *optional*): + Pre-generated text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* + prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` + input argument. + negative_prompt_embeds_pooled (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* + prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` + input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `prior_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `prior_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + prior_num_inference_steps (`int | dict[float, int]`, *optional*, defaults to 60): + The number of prior denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. For more specific timestep spacing, you can pass customized + `prior_timesteps` + num_inference_steps (`int`, *optional*, defaults to 12): + The number of decoder denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. For more specific timestep spacing, you can pass customized + `timesteps` + decoder_guidance_scale (`float`, *optional*, defaults to 0.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + prior_callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep: + int, callback_kwargs: Dict)`. + prior_callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the + list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in + the `._callback_tensor_inputs` attribute of your pipeline class. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True, + otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + dtype = self.decoder_pipe.decoder.dtype + if is_torch_version("<", "2.2.0") and dtype == torch.bfloat16: + raise ValueError( + "`StableCascadeCombinedPipeline` requires torch>=2.2.0 when using `torch.bfloat16` dtype." + ) + + prior_outputs = self.prior_pipe( + prompt=prompt if prompt_embeds is None else None, + images=images, + height=height, + width=width, + num_inference_steps=prior_num_inference_steps, + guidance_scale=prior_guidance_scale, + negative_prompt=negative_prompt if negative_prompt_embeds is None else None, + prompt_embeds=prompt_embeds, + prompt_embeds_pooled=prompt_embeds_pooled, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_pooled=negative_prompt_embeds_pooled, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + output_type="pt", + return_dict=True, + callback_on_step_end=prior_callback_on_step_end, + callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs, + ) + image_embeddings = prior_outputs.image_embeddings + prompt_embeds = prior_outputs.get("prompt_embeds", None) + prompt_embeds_pooled = prior_outputs.get("prompt_embeds_pooled", None) + negative_prompt_embeds = prior_outputs.get("negative_prompt_embeds", None) + negative_prompt_embeds_pooled = prior_outputs.get("negative_prompt_embeds_pooled", None) + + outputs = self.decoder_pipe( + image_embeddings=image_embeddings, + prompt=prompt if prompt_embeds is None else None, + num_inference_steps=num_inference_steps, + guidance_scale=decoder_guidance_scale, + negative_prompt=negative_prompt if negative_prompt_embeds is None else None, + prompt_embeds=prompt_embeds, + prompt_embeds_pooled=prompt_embeds_pooled, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_pooled=negative_prompt_embeds_pooled, + generator=generator, + output_type=output_type, + return_dict=return_dict, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + return outputs diff --git a/diffusers/pipelines/stable_cascade/pipeline_stable_cascade_prior.py b/diffusers/pipelines/stable_cascade/pipeline_stable_cascade_prior.py new file mode 100644 index 0000000000000000000000000000000000000000..fb58094f964b1ec1ed3217800d8262c7fd41b152 --- /dev/null +++ b/diffusers/pipelines/stable_cascade/pipeline_stable_cascade_prior.py @@ -0,0 +1,659 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from math import ceil +from typing import Callable + +import numpy as np +import PIL +import torch +from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...models import StableCascadeUNet +from ...schedulers import DDPMWuerstchenScheduler +from ...utils import BaseOutput, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DeprecatedPipelineMixin, DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +DEFAULT_STAGE_C_TIMESTEPS = list(np.linspace(1.0, 2 / 3, 20)) + list(np.linspace(2 / 3, 0.0, 11))[1:] + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableCascadePriorPipeline + + >>> prior_pipe = StableCascadePriorPipeline.from_pretrained( + ... "stabilityai/stable-cascade-prior", torch_dtype=torch.bfloat16 + ... ).to("cuda") + + >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet" + >>> prior_output = pipe(prompt) + ``` +""" + + +@dataclass +class StableCascadePriorPipelineOutput(BaseOutput): + """ + Output class for WuerstchenPriorPipeline. + + Args: + image_embeddings (`torch.Tensor` or `np.ndarray`) + Prior image embeddings for text prompt + prompt_embeds (`torch.Tensor`): + Text embeddings for the prompt. + negative_prompt_embeds (`torch.Tensor`): + Text embeddings for the negative prompt. + """ + + image_embeddings: torch.Tensor | np.ndarray + prompt_embeds: torch.Tensor | np.ndarray + prompt_embeds_pooled: torch.Tensor | np.ndarray + negative_prompt_embeds: torch.Tensor | np.ndarray + negative_prompt_embeds_pooled: torch.Tensor | np.ndarray + + +class StableCascadePriorPipeline(DeprecatedPipelineMixin, DiffusionPipeline): + """ + Pipeline for generating image prior for Stable Cascade. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + prior ([`StableCascadeUNet`]): + The Stable Cascade prior to approximate the image embedding from the text and/or image embedding. + text_encoder ([`CLIPTextModelWithProjection`]): + Frozen text-encoder + ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)). + feature_extractor ([`~transformers.CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `image_encoder`. + image_encoder ([`CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + scheduler ([`DDPMWuerstchenScheduler`]): + A scheduler to be used in combination with `prior` to generate image embedding. + resolution_multiple ('float', *optional*, defaults to 42.67): + Default resolution for multiple images generated. + """ + + _last_supported_version = "0.35.2" + + unet_name = "prior" + text_encoder_name = "text_encoder" + model_cpu_offload_seq = "image_encoder->text_encoder->prior" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "text_encoder_hidden_states", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModelWithProjection, + prior: StableCascadeUNet, + scheduler: DDPMWuerstchenScheduler, + resolution_multiple: float = 42.67, + feature_extractor: CLIPImageProcessor | None = None, + image_encoder: CLIPVisionModelWithProjection | None = None, + ) -> None: + super().__init__() + self.register_modules( + tokenizer=tokenizer, + text_encoder=text_encoder, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + prior=prior, + scheduler=scheduler, + ) + self.register_to_config(resolution_multiple=resolution_multiple) + + def prepare_latents( + self, batch_size, height, width, num_images_per_prompt, dtype, device, generator, latents, scheduler + ): + latent_shape = ( + num_images_per_prompt * batch_size, + self.prior.config.in_channels, + ceil(height / self.config.resolution_multiple), + ceil(width / self.config.resolution_multiple), + ) + + if latents is None: + latents = randn_tensor(latent_shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != latent_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latent_shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def encode_prompt( + self, + device, + batch_size, + num_images_per_prompt, + do_classifier_free_guidance, + prompt=None, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_pooled: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_pooled: torch.Tensor | None = None, + ): + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + attention_mask = text_inputs.attention_mask + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] + attention_mask = attention_mask[:, : self.tokenizer.model_max_length] + + text_encoder_output = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask.to(device), output_hidden_states=True + ) + prompt_embeds = text_encoder_output.hidden_states[-1] + if prompt_embeds_pooled is None: + prompt_embeds_pooled = text_encoder_output.text_embeds.unsqueeze(1) + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds_pooled = prompt_embeds_pooled.to(dtype=self.text_encoder.dtype, device=device) + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + prompt_embeds_pooled = prompt_embeds_pooled.repeat_interleave(num_images_per_prompt, dim=0) + + if negative_prompt_embeds is None and do_classifier_free_guidance: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + negative_prompt_embeds_text_encoder_output = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=uncond_input.attention_mask.to(device), + output_hidden_states=True, + ) + + negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.hidden_states[-1] + negative_prompt_embeds_pooled = negative_prompt_embeds_text_encoder_output.text_embeds.unsqueeze(1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + seq_len = negative_prompt_embeds_pooled.shape[1] + negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.to( + dtype=self.text_encoder.dtype, device=device + ) + negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + # done duplicates + + return prompt_embeds, prompt_embeds_pooled, negative_prompt_embeds, negative_prompt_embeds_pooled + + def encode_image(self, images, device, dtype, batch_size, num_images_per_prompt): + image_embeds = [] + for image in images: + image = self.feature_extractor(image, return_tensors="pt").pixel_values + image = image.to(device=device, dtype=dtype) + image_embed = self.image_encoder(image).image_embeds.unsqueeze(1) + image_embeds.append(image_embed) + image_embeds = torch.cat(image_embeds, dim=1) + + image_embeds = image_embeds.repeat(batch_size * num_images_per_prompt, 1, 1) + negative_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, negative_image_embeds + + def check_inputs( + self, + prompt, + images=None, + image_embeds=None, + negative_prompt=None, + prompt_embeds=None, + prompt_embeds_pooled=None, + negative_prompt_embeds=None, + negative_prompt_embeds_pooled=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and prompt_embeds_pooled is None: + raise ValueError( + "If `prompt_embeds` are provided, `prompt_embeds_pooled` must also be provided. Make sure to generate `prompt_embeds_pooled` from the same text encoder that was used to generate `prompt_embeds`" + ) + + if negative_prompt_embeds is not None and negative_prompt_embeds_pooled is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_prompt_embeds_pooled` must also be provided. Make sure to generate `prompt_embeds_pooled` from the same text encoder that was used to generate `prompt_embeds`" + ) + + if prompt_embeds_pooled is not None and negative_prompt_embeds_pooled is not None: + if prompt_embeds_pooled.shape != negative_prompt_embeds_pooled.shape: + raise ValueError( + "`prompt_embeds_pooled` and `negative_prompt_embeds_pooled` must have the same shape when passed" + f"directly, but got: `prompt_embeds_pooled` {prompt_embeds_pooled.shape} !=" + f"`negative_prompt_embeds_pooled` {negative_prompt_embeds_pooled.shape}." + ) + + if image_embeds is not None and images is not None: + raise ValueError( + f"Cannot forward both `images`: {images} and `image_embeds`: {image_embeds}. Please make sure to" + " only forward one of the two." + ) + + if images: + for i, image in enumerate(images): + if not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise TypeError( + f"'images' must contain images of type 'torch.Tensor' or 'PIL.Image.Image, but got" + f"{type(image)} for image number {i}." + ) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + def get_timestep_ratio_conditioning(self, t, alphas_cumprod): + s = torch.tensor([0.008]) + clamp_range = [0, 1] + min_var = torch.cos(s / (1 + s) * torch.pi * 0.5) ** 2 + var = alphas_cumprod[t] + var = var.clamp(*clamp_range) + s, min_var = s.to(var.device), min_var.to(var.device) + ratio = (((var * min_var) ** 0.5).acos() / (torch.pi * 0.5)) * (1 + s) - s + return ratio + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] | None = None, + images: torch.Tensor | PIL.Image.Image | list[torch.Tensor] | list[PIL.Image.Image] = None, + height: int = 1024, + width: int = 1024, + num_inference_steps: int = 20, + timesteps: list[float] = None, + guidance_scale: float = 4.0, + negative_prompt: str | list[str] | None = None, + prompt_embeds: torch.Tensor | None = None, + prompt_embeds_pooled: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds_pooled: torch.Tensor | None = None, + image_embeds: torch.Tensor | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pt", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + ): + """ + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + images (`torch.Tensor`, `PIL.Image.Image`, `list[torch.Tensor]` or `list[PIL.Image.Image]`, *optional*): + Reference image(s) used to condition the prior generation. When provided, image embeddings are derived + from the image and combined with the text prompt. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 60): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[float]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 8.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `decoder_guidance_scale` is defined as `w` of + equation 2. of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by + setting `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are + closely linked to the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `decoder_guidance_scale` is less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + prompt_embeds_pooled (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + negative_prompt_embeds_pooled (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds_pooled will be generated from `negative_prompt` + input argument. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs, *e.g.* prompt weighting. If + not provided, image embeddings will be generated from `image` input argument if existing. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"` + (`np.array`) or `"pt"` (`torch.Tensor`). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`StableCascadePriorPipelineOutput`] or `tuple` [`StableCascadePriorPipelineOutput`] if `return_dict` is + True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated image + embeddings. + """ + + # 0. Define commonly used variables + device = self._execution_device + dtype = next(self.prior.parameters()).dtype + self._guidance_scale = guidance_scale + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + images=images, + image_embeds=image_embeds, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_pooled=prompt_embeds_pooled, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_pooled=negative_prompt_embeds_pooled, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + # 2. Encode caption + images + ( + prompt_embeds, + prompt_embeds_pooled, + negative_prompt_embeds, + negative_prompt_embeds_pooled, + ) = self.encode_prompt( + prompt=prompt, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + prompt_embeds_pooled=prompt_embeds_pooled, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_pooled=negative_prompt_embeds_pooled, + ) + + if images is not None: + image_embeds_pooled, uncond_image_embeds_pooled = self.encode_image( + images=images, + device=device, + dtype=dtype, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + ) + elif image_embeds is not None: + image_embeds_pooled = image_embeds.repeat(batch_size * num_images_per_prompt, 1, 1) + uncond_image_embeds_pooled = torch.zeros_like(image_embeds_pooled) + else: + image_embeds_pooled = torch.zeros( + batch_size * num_images_per_prompt, + 1, + self.prior.config.clip_image_in_channels, + device=device, + dtype=dtype, + ) + uncond_image_embeds_pooled = torch.zeros( + batch_size * num_images_per_prompt, + 1, + self.prior.config.clip_image_in_channels, + device=device, + dtype=dtype, + ) + + if self.do_classifier_free_guidance: + image_embeds = torch.cat([image_embeds_pooled, uncond_image_embeds_pooled], dim=0) + else: + image_embeds = image_embeds_pooled + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + text_encoder_hidden_states = ( + torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds + ) + text_encoder_pooled = ( + torch.cat([prompt_embeds_pooled, negative_prompt_embeds_pooled]) + if negative_prompt_embeds is not None + else prompt_embeds_pooled + ) + + # 4. Prepare and set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latents + latents = self.prepare_latents( + batch_size, height, width, num_images_per_prompt, dtype, device, generator, latents, self.scheduler + ) + + if isinstance(self.scheduler, DDPMWuerstchenScheduler): + timesteps = timesteps[:-1] + else: + if hasattr(self.scheduler.config, "clip_sample") and self.scheduler.config.clip_sample: + self.scheduler.config.clip_sample = False # disample sample clipping + logger.warning(" set `clip_sample` to be False") + # 6. Run denoising loop + if hasattr(self.scheduler, "betas"): + alphas = 1.0 - self.scheduler.betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + else: + alphas_cumprod = [] + + self._num_timesteps = len(timesteps) + for i, t in enumerate(self.progress_bar(timesteps)): + if not isinstance(self.scheduler, DDPMWuerstchenScheduler): + if len(alphas_cumprod) > 0: + timestep_ratio = self.get_timestep_ratio_conditioning(t.long().cpu(), alphas_cumprod) + timestep_ratio = timestep_ratio.expand(latents.size(0)).to(dtype).to(device) + else: + timestep_ratio = t.float().div(self.scheduler.timesteps[-1]).expand(latents.size(0)).to(dtype) + else: + timestep_ratio = t.expand(latents.size(0)).to(dtype) + # 7. Denoise image embeddings + predicted_image_embedding = self.prior( + sample=torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, + timestep_ratio=torch.cat([timestep_ratio] * 2) if self.do_classifier_free_guidance else timestep_ratio, + clip_text_pooled=text_encoder_pooled, + clip_text=text_encoder_hidden_states, + clip_img=image_embeds, + return_dict=False, + )[0] + + # 8. Check for classifier free guidance and apply it + if self.do_classifier_free_guidance: + predicted_image_embedding_text, predicted_image_embedding_uncond = predicted_image_embedding.chunk(2) + predicted_image_embedding = torch.lerp( + predicted_image_embedding_uncond, predicted_image_embedding_text, self.guidance_scale + ) + + # 9. Renoise latents to next timestep + if not isinstance(self.scheduler, DDPMWuerstchenScheduler): + timestep_ratio = t + latents = self.scheduler.step( + model_output=predicted_image_embedding, timestep=timestep_ratio, sample=latents, generator=generator + ).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + if XLA_AVAILABLE: + xm.mark_step() + + # Offload all models + self.maybe_free_model_hooks() + + if output_type == "np": + latents = latents.cpu().float().numpy() # float() as bfloat16-> numpy doesn't work + prompt_embeds = prompt_embeds.cpu().float().numpy() # float() as bfloat16-> numpy doesn't work + negative_prompt_embeds = ( + negative_prompt_embeds.cpu().float().numpy() if negative_prompt_embeds is not None else None + ) # float() as bfloat16-> numpy doesn't work + + if not return_dict: + return ( + latents, + prompt_embeds, + prompt_embeds_pooled, + negative_prompt_embeds, + negative_prompt_embeds_pooled, + ) + + return StableCascadePriorPipelineOutput( + image_embeddings=latents, + prompt_embeds=prompt_embeds, + prompt_embeds_pooled=prompt_embeds_pooled, + negative_prompt_embeds=negative_prompt_embeds, + negative_prompt_embeds_pooled=negative_prompt_embeds_pooled, + ) diff --git a/diffusers/pipelines/stable_diffusion/README.md b/diffusers/pipelines/stable_diffusion/README.md new file mode 100644 index 0000000000000000000000000000000000000000..164baeb0a4d3b69c108d5957ae08a934b04b90c8 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/README.md @@ -0,0 +1,176 @@ +# Stable Diffusion + +## Overview + +Stable Diffusion was proposed in [Stable Diffusion Announcement](https://stability.ai/blog/stable-diffusion-announcement) by Patrick Esser and Robin Rombach and the Stability AI team. + +The summary of the model is the following: + +*Stable Diffusion is a text-to-image model that will empower billions of people to create stunning art within seconds. It is a breakthrough in speed and quality meaning that it can run on consumer GPUs. You can see some of the amazing output that has been created by this model without pre or post-processing on this page. The model itself builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain and many others. We are delighted that AI media generation is a cooperative field and hope it can continue this way to bring the gift of creativity to all.* + +## Tips: + +- Stable Diffusion has the same architecture as [Latent Diffusion](https://huggingface.co/papers/2112.10752) but uses a frozen CLIP Text Encoder instead of training the text encoder jointly with the diffusion model. +- An in-detail explanation of the Stable Diffusion model can be found under [Stable Diffusion with 🧨 Diffusers](https://huggingface.co/blog/stable_diffusion). +- If you don't want to rely on the Hugging Face Hub and having to pass a authentication token, you can +download the weights with `git lfs install; git clone https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5` and instead pass the local path to the cloned folder to `from_pretrained` as shown below. +- Stable Diffusion can work with a variety of different samplers as is shown below. + +## Available Pipelines: + +| Pipeline | Tasks | Colab +|---|---|:---:| +| [pipeline_stable_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) | *Text-to-Image Generation* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb) +| [pipeline_stable_diffusion_img2img](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) | *Image-to-Image Text-Guided Generation* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb) +| [pipeline_stable_diffusion_inpaint](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | *Text-Guided Image Inpainting* | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb) + +## Examples: + +### Using Stable Diffusion without being logged into the Hub. + +If you want to download the model weights using a single Python line, you need to be logged in via `hf auth login`. + +```python +from diffusers import DiffusionPipeline + +pipeline = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") +``` + +This however can make it difficult to build applications on top of `diffusers` as you will always have to pass the token around. A potential way to solve this issue is by downloading the weights to a local path `"./stable-diffusion-v1-5"`: + +``` +git lfs install +git clone https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 +``` + +and simply passing the local path to `from_pretrained`: + +```python +from diffusers import StableDiffusionPipeline + +pipe = StableDiffusionPipeline.from_pretrained("./stable-diffusion-v1-5") +``` + +### Text-to-Image with default PLMS scheduler + +```python +# make sure you're logged in with `hf auth login` +from diffusers import StableDiffusionPipeline + +pipe = StableDiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5") +pipe = pipe.to("cuda") + +prompt = "a photo of an astronaut riding a horse on mars" +image = pipe(prompt).images[0] + +image.save("astronaut_rides_horse.png") +``` + +### Text-to-Image with DDIM scheduler + +```python +# make sure you're logged in with `hf auth login` +from diffusers import StableDiffusionPipeline, DDIMScheduler + +scheduler = DDIMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler") + +pipe = StableDiffusionPipeline.from_pretrained( + "stable-diffusion-v1-5/stable-diffusion-v1-5", + scheduler=scheduler, +).to("cuda") + +prompt = "a photo of an astronaut riding a horse on mars" +image = pipe(prompt).images[0] + +image.save("astronaut_rides_horse.png") +``` + +### Text-to-Image with K-LMS scheduler + +```python +# make sure you're logged in with `hf auth login` +from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler + +lms = LMSDiscreteScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler") + +pipe = StableDiffusionPipeline.from_pretrained( + "stable-diffusion-v1-5/stable-diffusion-v1-5", + scheduler=lms, +).to("cuda") + +prompt = "a photo of an astronaut riding a horse on mars" +image = pipe(prompt).images[0] + +image.save("astronaut_rides_horse.png") +``` + +### CycleDiffusion using Stable Diffusion and DDIM scheduler + +```python +import requests +import torch +from PIL import Image +from io import BytesIO + +from diffusers import CycleDiffusionPipeline, DDIMScheduler + + +# load the scheduler. CycleDiffusion only supports stochastic schedulers. + +# load the pipeline +# make sure you're logged in with `hf auth login` +model_id_or_path = "CompVis/stable-diffusion-v1-4" +scheduler = DDIMScheduler.from_pretrained(model_id_or_path, subfolder="scheduler") +pipe = CycleDiffusionPipeline.from_pretrained(model_id_or_path, scheduler=scheduler).to("cuda") + +# let's download an initial image +url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/An%20astronaut%20riding%20a%20horse.png" +response = requests.get(url) +init_image = Image.open(BytesIO(response.content)).convert("RGB") +init_image = init_image.resize((512, 512)) +init_image.save("horse.png") + +# let's specify a prompt +source_prompt = "An astronaut riding a horse" +prompt = "An astronaut riding an elephant" + +# call the pipeline +image = pipe( + prompt=prompt, + source_prompt=source_prompt, + image=init_image, + num_inference_steps=100, + eta=0.1, + strength=0.8, + guidance_scale=2, + source_guidance_scale=1, +).images[0] + +image.save("horse_to_elephant.png") + +# let's try another example +# See more samples at the original repo: https://github.com/ChenWu98/cycle-diffusion +url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png" +response = requests.get(url) +init_image = Image.open(BytesIO(response.content)).convert("RGB") +init_image = init_image.resize((512, 512)) +init_image.save("black.png") + +source_prompt = "A black colored car" +prompt = "A blue colored car" + +# call the pipeline +torch.manual_seed(0) +image = pipe( + prompt=prompt, + source_prompt=source_prompt, + image=init_image, + num_inference_steps=100, + eta=0.1, + strength=0.85, + guidance_scale=3, + source_guidance_scale=1, +).images[0] + +image.save("black_to_blue.png") +``` diff --git a/diffusers/pipelines/stable_diffusion/__init__.py b/diffusers/pipelines/stable_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..263457a76489867dfb9b1c0ffb66a8b12bf4ce9d --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/__init__.py @@ -0,0 +1,165 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_onnx_available, + is_torch_available, + is_transformers_available, + is_transformers_version, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["StableDiffusionPipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["clip_image_project_model"] = ["CLIPImageProjection"] + _import_structure["pipeline_stable_diffusion"] = ["StableDiffusionPipeline"] + _import_structure["pipeline_stable_diffusion_img2img"] = ["StableDiffusionImg2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_inpaint"] = ["StableDiffusionInpaintPipeline"] + _import_structure["pipeline_stable_diffusion_instruct_pix2pix"] = ["StableDiffusionInstructPix2PixPipeline"] + _import_structure["pipeline_stable_diffusion_latent_upscale"] = ["StableDiffusionLatentUpscalePipeline"] + _import_structure["pipeline_stable_diffusion_upscale"] = ["StableDiffusionUpscalePipeline"] + _import_structure["pipeline_stable_unclip"] = ["StableUnCLIPPipeline"] + _import_structure["pipeline_stable_unclip_img2img"] = ["StableUnCLIPImg2ImgPipeline"] + _import_structure["safety_checker"] = ["StableDiffusionSafetyChecker"] + _import_structure["stable_unclip_image_normalizer"] = ["StableUnCLIPImageNormalizer"] +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + StableDiffusionImageVariationPipeline, + ) + + _dummy_objects.update({"StableDiffusionImageVariationPipeline": StableDiffusionImageVariationPipeline}) +else: + _import_structure["pipeline_stable_diffusion_image_variation"] = ["StableDiffusionImageVariationPipeline"] +try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.26.0")): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + StableDiffusionDepth2ImgPipeline, + ) + + _dummy_objects.update( + { + "StableDiffusionDepth2ImgPipeline": StableDiffusionDepth2ImgPipeline, + } + ) +else: + _import_structure["pipeline_stable_diffusion_depth2img"] = ["StableDiffusionDepth2ImgPipeline"] + +try: + if not (is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_onnx_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_onnx_objects)) +else: + _import_structure["pipeline_onnx_stable_diffusion"] = [ + "OnnxStableDiffusionPipeline", + "StableDiffusionOnnxPipeline", + ] + _import_structure["pipeline_onnx_stable_diffusion_img2img"] = ["OnnxStableDiffusionImg2ImgPipeline"] + _import_structure["pipeline_onnx_stable_diffusion_inpaint"] = ["OnnxStableDiffusionInpaintPipeline"] + _import_structure["pipeline_onnx_stable_diffusion_inpaint_legacy"] = ["OnnxStableDiffusionInpaintPipelineLegacy"] + _import_structure["pipeline_onnx_stable_diffusion_upscale"] = ["OnnxStableDiffusionUpscalePipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + + else: + from .clip_image_project_model import CLIPImageProjection + from .pipeline_stable_diffusion import ( + StableDiffusionPipeline, + StableDiffusionPipelineOutput, + ) + from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline + from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline + from .pipeline_stable_diffusion_instruct_pix2pix import ( + StableDiffusionInstructPix2PixPipeline, + ) + from .pipeline_stable_diffusion_latent_upscale import ( + StableDiffusionLatentUpscalePipeline, + ) + from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline + from .pipeline_stable_unclip import StableUnCLIPPipeline + from .pipeline_stable_unclip_img2img import StableUnCLIPImg2ImgPipeline + from .safety_checker import StableDiffusionSafetyChecker + from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import ( + StableDiffusionImageVariationPipeline, + ) + else: + from .pipeline_stable_diffusion_image_variation import ( + StableDiffusionImageVariationPipeline, + ) + + try: + if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.26.0")): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import StableDiffusionDepth2ImgPipeline + else: + from .pipeline_stable_diffusion_depth2img import ( + StableDiffusionDepth2ImgPipeline, + ) + + try: + if not (is_transformers_available() and is_onnx_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_onnx_objects import * + else: + from .pipeline_onnx_stable_diffusion import ( + OnnxStableDiffusionPipeline, + StableDiffusionOnnxPipeline, + ) + from .pipeline_onnx_stable_diffusion_img2img import ( + OnnxStableDiffusionImg2ImgPipeline, + ) + from .pipeline_onnx_stable_diffusion_inpaint import ( + OnnxStableDiffusionInpaintPipeline, + ) + from .pipeline_onnx_stable_diffusion_upscale import ( + OnnxStableDiffusionUpscalePipeline, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/stable_diffusion/clip_image_project_model.py b/diffusers/pipelines/stable_diffusion/clip_image_project_model.py new file mode 100644 index 0000000000000000000000000000000000000000..30dd90242d07d249e62e976c148a9d2cde02271e --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/clip_image_project_model.py @@ -0,0 +1,29 @@ +# Copyright 2025 The GLIGEN Authors and HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin + + +class CLIPImageProjection(ModelMixin, ConfigMixin): + @register_to_config + def __init__(self, hidden_size: int = 768): + super().__init__() + self.hidden_size = hidden_size + self.project = nn.Linear(self.hidden_size, self.hidden_size, bias=False) + + def forward(self, x): + return self.project(x) diff --git a/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py b/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py new file mode 100644 index 0000000000000000000000000000000000000000..0c8fd842fcbab26dbfe0e4dc8252dc22512e3093 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/convert_from_ckpt.py @@ -0,0 +1,1876 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Conversion script for the Stable Diffusion checkpoints.""" + +import re +from contextlib import nullcontext +from io import BytesIO + +import requests +import torch +import yaml +from transformers import ( + AutoFeatureExtractor, + BertTokenizerFast, + CLIPImageProcessor, + CLIPTextConfig, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionConfig, + CLIPVisionModelWithProjection, +) + +from ...models import ( + AutoencoderKL, + ControlNetModel, + PriorTransformer, + UNet2DConditionModel, +) +from ...schedulers import ( + DDIMScheduler, + DDPMScheduler, + DPMSolverMultistepScheduler, + EulerAncestralDiscreteScheduler, + EulerDiscreteScheduler, + HeunDiscreteScheduler, + LMSDiscreteScheduler, + PNDMScheduler, + UnCLIPScheduler, +) +from ...utils import is_accelerate_available, logging +from ...utils.constants import DIFFUSERS_REQUEST_TIMEOUT +from ...utils.torch_utils import get_device +from ..deprecated.paint_by_example import PaintByExampleImageEncoder +from ..latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel +from ..pipeline_utils import DiffusionPipeline +from .safety_checker import StableDiffusionSafetyChecker +from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + +if is_accelerate_available(): + from accelerate import init_empty_weights + from accelerate.utils import set_module_tensor_to_device + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def shave_segments(path, n_shave_prefix_segments=1): + """ + Removes segments. Positive values shave the first segments, negative shave the last segments. + """ + if n_shave_prefix_segments >= 0: + return ".".join(path.split(".")[n_shave_prefix_segments:]) + else: + return ".".join(path.split(".")[:n_shave_prefix_segments]) + + +def renew_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item.replace("in_layers.0", "norm1") + new_item = new_item.replace("in_layers.2", "conv1") + + new_item = new_item.replace("out_layers.0", "norm2") + new_item = new_item.replace("out_layers.3", "conv2") + + new_item = new_item.replace("emb_layers.1", "time_emb_proj") + new_item = new_item.replace("skip_connection", "conv_shortcut") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside resnets to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("nin_shortcut", "conv_shortcut") + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + # new_item = new_item.replace('norm.weight', 'group_norm.weight') + # new_item = new_item.replace('norm.bias', 'group_norm.bias') + + # new_item = new_item.replace('proj_out.weight', 'proj_attn.weight') + # new_item = new_item.replace('proj_out.bias', 'proj_attn.bias') + + # new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0): + """ + Updates paths inside attentions to the new naming scheme (local renaming) + """ + mapping = [] + for old_item in old_list: + new_item = old_item + + new_item = new_item.replace("norm.weight", "group_norm.weight") + new_item = new_item.replace("norm.bias", "group_norm.bias") + + new_item = new_item.replace("q.weight", "to_q.weight") + new_item = new_item.replace("q.bias", "to_q.bias") + + new_item = new_item.replace("k.weight", "to_k.weight") + new_item = new_item.replace("k.bias", "to_k.bias") + + new_item = new_item.replace("v.weight", "to_v.weight") + new_item = new_item.replace("v.bias", "to_v.bias") + + new_item = new_item.replace("proj_out.weight", "to_out.0.weight") + new_item = new_item.replace("proj_out.bias", "to_out.0.bias") + + new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments) + + mapping.append({"old": old_item, "new": new_item}) + + return mapping + + +def assign_to_checkpoint( + paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None +): + """ + This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits + attention layers, and takes into account additional replacements that may arise. + + Assigns the weights to the new checkpoint. + """ + assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys." + + # Splits the attention layers into three variables. + if attention_paths_to_split is not None: + for path, path_map in attention_paths_to_split.items(): + old_tensor = old_checkpoint[path] + channels = old_tensor.shape[0] // 3 + + target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1) + + num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3 + + old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:]) + query, key, value = old_tensor.split(channels // num_heads, dim=1) + + checkpoint[path_map["query"]] = query.reshape(target_shape) + checkpoint[path_map["key"]] = key.reshape(target_shape) + checkpoint[path_map["value"]] = value.reshape(target_shape) + + for path in paths: + new_path = path["new"] + + # These have already been assigned + if attention_paths_to_split is not None and new_path in attention_paths_to_split: + continue + + # Global renaming happens here + new_path = new_path.replace("middle_block.0", "mid_block.resnets.0") + new_path = new_path.replace("middle_block.1", "mid_block.attentions.0") + new_path = new_path.replace("middle_block.2", "mid_block.resnets.1") + + if additional_replacements is not None: + for replacement in additional_replacements: + new_path = new_path.replace(replacement["old"], replacement["new"]) + + # proj_attn.weight has to be converted from conv 1D to linear + is_attn_weight = "proj_attn.weight" in new_path or ("attentions" in new_path and "to_" in new_path) + shape = old_checkpoint[path["old"]].shape + if is_attn_weight and len(shape) == 3: + checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0] + elif is_attn_weight and len(shape) == 4: + checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0, 0] + else: + checkpoint[new_path] = old_checkpoint[path["old"]] + + +def conv_attn_to_linear(checkpoint): + keys = list(checkpoint.keys()) + attn_keys = ["query.weight", "key.weight", "value.weight"] + for key in keys: + if ".".join(key.split(".")[-2:]) in attn_keys: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0, 0] + elif "proj_attn.weight" in key: + if checkpoint[key].ndim > 2: + checkpoint[key] = checkpoint[key][:, :, 0] + + +def create_unet_diffusers_config(original_config, image_size: int, controlnet=False): + """ + Creates a config for the diffusers based on the config of the LDM model. + """ + if controlnet: + unet_params = original_config["model"]["params"]["control_stage_config"]["params"] + else: + if ( + "unet_config" in original_config["model"]["params"] + and original_config["model"]["params"]["unet_config"] is not None + ): + unet_params = original_config["model"]["params"]["unet_config"]["params"] + else: + unet_params = original_config["model"]["params"]["network_config"]["params"] + + vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"] + + block_out_channels = [unet_params["model_channels"] * mult for mult in unet_params["channel_mult"]] + + down_block_types = [] + resolution = 1 + for i in range(len(block_out_channels)): + block_type = "CrossAttnDownBlock2D" if resolution in unet_params["attention_resolutions"] else "DownBlock2D" + down_block_types.append(block_type) + if i != len(block_out_channels) - 1: + resolution *= 2 + + up_block_types = [] + for i in range(len(block_out_channels)): + block_type = "CrossAttnUpBlock2D" if resolution in unet_params["attention_resolutions"] else "UpBlock2D" + up_block_types.append(block_type) + resolution //= 2 + + if unet_params["transformer_depth"] is not None: + transformer_layers_per_block = ( + unet_params["transformer_depth"] + if isinstance(unet_params["transformer_depth"], int) + else list(unet_params["transformer_depth"]) + ) + else: + transformer_layers_per_block = 1 + + vae_scale_factor = 2 ** (len(vae_params["ch_mult"]) - 1) + + head_dim = unet_params["num_heads"] if "num_heads" in unet_params else None + use_linear_projection = ( + unet_params["use_linear_in_transformer"] if "use_linear_in_transformer" in unet_params else False + ) + if use_linear_projection: + # stable diffusion 2-base-512 and 2-768 + if head_dim is None: + head_dim_mult = unet_params["model_channels"] // unet_params["num_head_channels"] + head_dim = [head_dim_mult * c for c in list(unet_params["channel_mult"])] + + class_embed_type = None + addition_embed_type = None + addition_time_embed_dim = None + projection_class_embeddings_input_dim = None + context_dim = None + + if unet_params["context_dim"] is not None: + context_dim = ( + unet_params["context_dim"] + if isinstance(unet_params["context_dim"], int) + else unet_params["context_dim"][0] + ) + + if "num_classes" in unet_params: + if unet_params["num_classes"] == "sequential": + if context_dim in [2048, 1280]: + # SDXL + addition_embed_type = "text_time" + addition_time_embed_dim = 256 + else: + class_embed_type = "projection" + assert "adm_in_channels" in unet_params + projection_class_embeddings_input_dim = unet_params["adm_in_channels"] + + config = { + "sample_size": image_size // vae_scale_factor, + "in_channels": unet_params["in_channels"], + "down_block_types": tuple(down_block_types), + "block_out_channels": tuple(block_out_channels), + "layers_per_block": unet_params["num_res_blocks"], + "cross_attention_dim": context_dim, + "attention_head_dim": head_dim, + "use_linear_projection": use_linear_projection, + "class_embed_type": class_embed_type, + "addition_embed_type": addition_embed_type, + "addition_time_embed_dim": addition_time_embed_dim, + "projection_class_embeddings_input_dim": projection_class_embeddings_input_dim, + "transformer_layers_per_block": transformer_layers_per_block, + } + + if "disable_self_attentions" in unet_params: + config["only_cross_attention"] = unet_params["disable_self_attentions"] + + if "num_classes" in unet_params and isinstance(unet_params["num_classes"], int): + config["num_class_embeds"] = unet_params["num_classes"] + + if controlnet: + config["conditioning_channels"] = unet_params["hint_channels"] + else: + config["out_channels"] = unet_params["out_channels"] + config["up_block_types"] = tuple(up_block_types) + + return config + + +def create_vae_diffusers_config(original_config, image_size: int): + """ + Creates a config for the diffusers based on the config of the LDM model. + """ + vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"] + _ = original_config["model"]["params"]["first_stage_config"]["params"]["embed_dim"] + + block_out_channels = [vae_params["ch"] * mult for mult in vae_params["ch_mult"]] + down_block_types = [ + "DownEncoderBlock2D" if image_size // 2**i not in vae_params["attn_resolutions"] else "AttnDownEncoderBlock2D" + for i, _ in enumerate(block_out_channels) + ] + up_block_types = [ + "UpDecoderBlock2D" if image_size // 2**i not in vae_params["attn_resolutions"] else "AttnUpDecoderBlock2D" + for i, _ in enumerate(block_out_channels) + ][::-1] + + config = { + "sample_size": image_size, + "in_channels": vae_params["in_channels"], + "out_channels": vae_params["out_ch"], + "down_block_types": tuple(down_block_types), + "up_block_types": tuple(up_block_types), + "block_out_channels": tuple(block_out_channels), + "latent_channels": vae_params["z_channels"], + "layers_per_block": vae_params["num_res_blocks"], + } + return config + + +def create_diffusers_schedular(original_config): + schedular = DDIMScheduler( + num_train_timesteps=original_config["model"]["params"]["timesteps"], + beta_start=original_config["model"]["params"]["linear_start"], + beta_end=original_config["model"]["params"]["linear_end"], + beta_schedule="scaled_linear", + ) + return schedular + + +def create_ldm_bert_config(original_config): + bert_params = original_config["model"]["params"]["cond_stage_config"]["params"] + config = LDMBertConfig( + d_model=bert_params.n_embed, + encoder_layers=bert_params.n_layer, + encoder_ffn_dim=bert_params.n_embed * 4, + ) + return config + + +def convert_ldm_unet_checkpoint( + checkpoint, config, path=None, extract_ema=False, controlnet=False, skip_extract_state_dict=False +): + """ + Takes a state dict and a config, and returns a converted checkpoint. + """ + + if skip_extract_state_dict: + unet_state_dict = checkpoint + else: + # extract state_dict for UNet + unet_state_dict = {} + keys = list(checkpoint.keys()) + + if controlnet: + unet_key = "control_model." + else: + unet_key = "model.diffusion_model." + + # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA + if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema: + logger.warning(f"Checkpoint {path} has both EMA and non-EMA weights.") + logger.warning( + "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA" + " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag." + ) + for key in keys: + if key.startswith("model.diffusion_model"): + flat_ema_key = "model_ema." + "".join(key.split(".")[1:]) + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key) + else: + if sum(k.startswith("model_ema") for k in keys) > 100: + logger.warning( + "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA" + " weights (usually better for inference), please make sure to add the `--extract_ema` flag." + ) + + for key in keys: + if key.startswith(unet_key): + unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key) + + new_checkpoint = {} + + new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"] + new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"] + new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"] + new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"] + + if config["class_embed_type"] is None: + # No parameters to port + ... + elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection": + new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"] + new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"] + new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"] + new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"] + else: + raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}") + + if config["addition_embed_type"] == "text_time": + new_checkpoint["add_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"] + new_checkpoint["add_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"] + new_checkpoint["add_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"] + new_checkpoint["add_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"] + + # Relevant to StableDiffusionUpscalePipeline + if "num_class_embeds" in config: + if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict): + new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"] + + new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"] + new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"] + + if not controlnet: + new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"] + new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"] + new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"] + new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"] + + # Retrieves the keys for the input blocks only + num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer}) + input_blocks = { + layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key] + for layer_id in range(num_input_blocks) + } + + # Retrieves the keys for the middle blocks only + num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer}) + middle_blocks = { + layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key] + for layer_id in range(num_middle_blocks) + } + + # Retrieves the keys for the output blocks only + num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer}) + output_blocks = { + layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key] + for layer_id in range(num_output_blocks) + } + + for i in range(1, num_input_blocks): + block_id = (i - 1) // (config["layers_per_block"] + 1) + layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1) + + resnets = [ + key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key + ] + attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key] + + if f"input_blocks.{i}.0.op.weight" in unet_state_dict: + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.weight" + ) + new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop( + f"input_blocks.{i}.0.op.bias" + ) + + paths = renew_resnet_paths(resnets) + meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + if len(attentions): + paths = renew_attention_paths(attentions) + + meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + resnet_0 = middle_blocks[0] + attentions = middle_blocks[1] + resnet_1 = middle_blocks[2] + + resnet_0_paths = renew_resnet_paths(resnet_0) + assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config) + + resnet_1_paths = renew_resnet_paths(resnet_1) + assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config) + + attentions_paths = renew_attention_paths(attentions) + meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"} + assign_to_checkpoint( + attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + for i in range(num_output_blocks): + block_id = i // (config["layers_per_block"] + 1) + layer_in_block_id = i % (config["layers_per_block"] + 1) + output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]] + output_block_list = {} + + for layer in output_block_layers: + layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1) + if layer_id in output_block_list: + output_block_list[layer_id].append(layer_name) + else: + output_block_list[layer_id] = [layer_name] + + if len(output_block_list) > 1: + resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key] + attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key] + + resnet_0_paths = renew_resnet_paths(resnets) + paths = renew_resnet_paths(resnets) + + meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"} + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + + output_block_list = {k: sorted(v) for k, v in sorted(output_block_list.items())} + if ["conv.bias", "conv.weight"] in output_block_list.values(): + index = list(output_block_list.values()).index(["conv.bias", "conv.weight"]) + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.weight" + ] + new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[ + f"output_blocks.{i}.{index}.conv.bias" + ] + + # Clear attentions as they have been attributed above. + if len(attentions) == 2: + attentions = [] + + if len(attentions): + paths = renew_attention_paths(attentions) + meta_path = { + "old": f"output_blocks.{i}.1", + "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}", + } + assign_to_checkpoint( + paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config + ) + else: + resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1) + for path in resnet_0_paths: + old_path = ".".join(["output_blocks", str(i), path["old"]]) + new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]]) + + new_checkpoint[new_path] = unet_state_dict[old_path] + + if controlnet: + # conditioning embedding + + orig_index = 0 + + new_checkpoint["controlnet_cond_embedding.conv_in.weight"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.weight" + ) + new_checkpoint["controlnet_cond_embedding.conv_in.bias"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.bias" + ) + + orig_index += 2 + + diffusers_index = 0 + + while diffusers_index < 6: + new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.weight"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.weight" + ) + new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.bias"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.bias" + ) + diffusers_index += 1 + orig_index += 2 + + new_checkpoint["controlnet_cond_embedding.conv_out.weight"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.weight" + ) + new_checkpoint["controlnet_cond_embedding.conv_out.bias"] = unet_state_dict.pop( + f"input_hint_block.{orig_index}.bias" + ) + + # down blocks + for i in range(num_input_blocks): + new_checkpoint[f"controlnet_down_blocks.{i}.weight"] = unet_state_dict.pop(f"zero_convs.{i}.0.weight") + new_checkpoint[f"controlnet_down_blocks.{i}.bias"] = unet_state_dict.pop(f"zero_convs.{i}.0.bias") + + # mid block + new_checkpoint["controlnet_mid_block.weight"] = unet_state_dict.pop("middle_block_out.0.weight") + new_checkpoint["controlnet_mid_block.bias"] = unet_state_dict.pop("middle_block_out.0.bias") + + return new_checkpoint + + +def convert_ldm_vae_checkpoint(checkpoint, config): + # extract state dict for VAE + vae_state_dict = {} + keys = list(checkpoint.keys()) + vae_key = "first_stage_model." if any(k.startswith("first_stage_model.") for k in keys) else "" + for key in keys: + if key.startswith(vae_key): + vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key) + + new_checkpoint = {} + + new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"] + new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"] + new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"] + new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"] + new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"] + new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"] + + new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"] + new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"] + new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"] + new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"] + new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"] + new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"] + + new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"] + new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"] + new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"] + new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"] + + # Retrieves the keys for the encoder down blocks only + num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer}) + down_blocks = { + layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks) + } + + # Retrieves the keys for the decoder up blocks only + num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer}) + up_blocks = { + layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks) + } + + for i in range(num_down_blocks): + resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key] + + if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict: + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.weight" + ) + new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop( + f"encoder.down.{i}.downsample.conv.bias" + ) + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + + for i in range(num_up_blocks): + block_id = num_up_blocks - 1 - i + resnets = [ + key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key + ] + + if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict: + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.weight" + ] + new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[ + f"decoder.up.{block_id}.upsample.conv.bias" + ] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key] + num_mid_res_blocks = 2 + for i in range(1, num_mid_res_blocks + 1): + resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key] + + paths = renew_vae_resnet_paths(resnets) + meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + + mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key] + paths = renew_vae_attention_paths(mid_attentions) + meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"} + assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config) + conv_attn_to_linear(new_checkpoint) + return new_checkpoint + + +def convert_ldm_bert_checkpoint(checkpoint, config): + def _copy_attn_layer(hf_attn_layer, pt_attn_layer): + hf_attn_layer.q_proj.weight.data = pt_attn_layer.to_q.weight + hf_attn_layer.k_proj.weight.data = pt_attn_layer.to_k.weight + hf_attn_layer.v_proj.weight.data = pt_attn_layer.to_v.weight + + hf_attn_layer.out_proj.weight = pt_attn_layer.to_out.weight + hf_attn_layer.out_proj.bias = pt_attn_layer.to_out.bias + + def _copy_linear(hf_linear, pt_linear): + hf_linear.weight = pt_linear.weight + hf_linear.bias = pt_linear.bias + + def _copy_layer(hf_layer, pt_layer): + # copy layer norms + _copy_linear(hf_layer.self_attn_layer_norm, pt_layer[0][0]) + _copy_linear(hf_layer.final_layer_norm, pt_layer[1][0]) + + # copy attn + _copy_attn_layer(hf_layer.self_attn, pt_layer[0][1]) + + # copy MLP + pt_mlp = pt_layer[1][1] + _copy_linear(hf_layer.fc1, pt_mlp.net[0][0]) + _copy_linear(hf_layer.fc2, pt_mlp.net[2]) + + def _copy_layers(hf_layers, pt_layers): + for i, hf_layer in enumerate(hf_layers): + if i != 0: + i += i + pt_layer = pt_layers[i : i + 2] + _copy_layer(hf_layer, pt_layer) + + hf_model = LDMBertModel(config).eval() + + # copy embeds + hf_model.model.embed_tokens.weight = checkpoint.transformer.token_emb.weight + hf_model.model.embed_positions.weight.data = checkpoint.transformer.pos_emb.emb.weight + + # copy layer norm + _copy_linear(hf_model.model.layer_norm, checkpoint.transformer.norm) + + # copy hidden layers + _copy_layers(hf_model.model.layers, checkpoint.transformer.attn_layers.layers) + + _copy_linear(hf_model.to_logits, checkpoint.transformer.to_logits) + + return hf_model + + +def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder=None): + if text_encoder is None: + config_name = "openai/clip-vit-large-patch14" + try: + config = CLIPTextConfig.from_pretrained(config_name, local_files_only=local_files_only) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: 'openai/clip-vit-large-patch14'." + ) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + text_model = CLIPTextModel(config) + else: + text_model = text_encoder + + keys = list(checkpoint.keys()) + + text_model_dict = {} + + remove_prefixes = ["cond_stage_model.transformer", "conditioner.embedders.0.transformer"] + + for key in keys: + for prefix in remove_prefixes: + if key.startswith(prefix): + text_model_dict[key[len(prefix + ".") :]] = checkpoint[key] + + if is_accelerate_available(): + for param_name, param in text_model_dict.items(): + set_module_tensor_to_device(text_model, param_name, "cpu", value=param) + else: + if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)): + text_model_dict.pop("text_model.embeddings.position_ids", None) + + text_model.load_state_dict(text_model_dict) + + return text_model + + +textenc_conversion_lst = [ + ("positional_embedding", "text_model.embeddings.position_embedding.weight"), + ("token_embedding.weight", "text_model.embeddings.token_embedding.weight"), + ("ln_final.weight", "text_model.final_layer_norm.weight"), + ("ln_final.bias", "text_model.final_layer_norm.bias"), + ("text_projection", "text_projection.weight"), +] +textenc_conversion_map = {x[0]: x[1] for x in textenc_conversion_lst} + +textenc_transformer_conversion_lst = [ + # (stable-diffusion, HF Diffusers) + ("resblocks.", "text_model.encoder.layers."), + ("ln_1", "layer_norm1"), + ("ln_2", "layer_norm2"), + (".c_fc.", ".fc1."), + (".c_proj.", ".fc2."), + (".attn", ".self_attn"), + ("ln_final.", "transformer.text_model.final_layer_norm."), + ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"), + ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"), +] +protected = {re.escape(x[0]): x[1] for x in textenc_transformer_conversion_lst} +textenc_pattern = re.compile("|".join(protected.keys())) + + +def convert_paint_by_example_checkpoint(checkpoint, local_files_only=False): + config = CLIPVisionConfig.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only) + model = PaintByExampleImageEncoder(config) + + keys = list(checkpoint.keys()) + + text_model_dict = {} + + for key in keys: + if key.startswith("cond_stage_model.transformer"): + text_model_dict[key[len("cond_stage_model.transformer.") :]] = checkpoint[key] + + # load clip vision + model.model.load_state_dict(text_model_dict) + + # load mapper + keys_mapper = { + k[len("cond_stage_model.mapper.res") :]: v + for k, v in checkpoint.items() + if k.startswith("cond_stage_model.mapper") + } + + MAPPING = { + "attn.c_qkv": ["attn1.to_q", "attn1.to_k", "attn1.to_v"], + "attn.c_proj": ["attn1.to_out.0"], + "ln_1": ["norm1"], + "ln_2": ["norm3"], + "mlp.c_fc": ["ff.net.0.proj"], + "mlp.c_proj": ["ff.net.2"], + } + + mapped_weights = {} + for key, value in keys_mapper.items(): + prefix = key[: len("blocks.i")] + suffix = key.split(prefix)[-1].split(".")[-1] + name = key.split(prefix)[-1].split(suffix)[0][1:-1] + mapped_names = MAPPING[name] + + num_splits = len(mapped_names) + for i, mapped_name in enumerate(mapped_names): + new_name = ".".join([prefix, mapped_name, suffix]) + shape = value.shape[0] // num_splits + mapped_weights[new_name] = value[i * shape : (i + 1) * shape] + + model.mapper.load_state_dict(mapped_weights) + + # load final layer norm + model.final_layer_norm.load_state_dict( + { + "bias": checkpoint["cond_stage_model.final_ln.bias"], + "weight": checkpoint["cond_stage_model.final_ln.weight"], + } + ) + + # load final proj + model.proj_out.load_state_dict( + { + "bias": checkpoint["proj_out.bias"], + "weight": checkpoint["proj_out.weight"], + } + ) + + # load uncond vector + model.uncond_vector.data = torch.nn.Parameter(checkpoint["learnable_vector"]) + return model + + +def convert_open_clip_checkpoint( + checkpoint, + config_name, + prefix="cond_stage_model.model.", + has_projection=False, + local_files_only=False, + **config_kwargs, +): + # text_model = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="text_encoder") + # text_model = CLIPTextModelWithProjection.from_pretrained( + # "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280 + # ) + try: + config = CLIPTextConfig.from_pretrained(config_name, **config_kwargs, local_files_only=local_files_only) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: '{config_name}'." + ) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + text_model = CLIPTextModelWithProjection(config) if has_projection else CLIPTextModel(config) + + keys = list(checkpoint.keys()) + + keys_to_ignore = [] + if config_name == "stabilityai/stable-diffusion-2" and config.num_hidden_layers == 23: + # make sure to remove all keys > 22 + keys_to_ignore += [k for k in keys if k.startswith("cond_stage_model.model.transformer.resblocks.23")] + keys_to_ignore += ["cond_stage_model.model.text_projection"] + + text_model_dict = {} + + if prefix + "text_projection" in checkpoint: + d_model = int(checkpoint[prefix + "text_projection"].shape[0]) + else: + d_model = 1024 + + text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids") + + for key in keys: + if key in keys_to_ignore: + continue + if key[len(prefix) :] in textenc_conversion_map: + if key.endswith("text_projection"): + value = checkpoint[key].T.contiguous() + else: + value = checkpoint[key] + + text_model_dict[textenc_conversion_map[key[len(prefix) :]]] = value + + if key.startswith(prefix + "transformer."): + new_key = key[len(prefix + "transformer.") :] + if new_key.endswith(".in_proj_weight"): + new_key = new_key[: -len(".in_proj_weight")] + new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key) + text_model_dict[new_key + ".q_proj.weight"] = checkpoint[key][:d_model, :] + text_model_dict[new_key + ".k_proj.weight"] = checkpoint[key][d_model : d_model * 2, :] + text_model_dict[new_key + ".v_proj.weight"] = checkpoint[key][d_model * 2 :, :] + elif new_key.endswith(".in_proj_bias"): + new_key = new_key[: -len(".in_proj_bias")] + new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key) + text_model_dict[new_key + ".q_proj.bias"] = checkpoint[key][:d_model] + text_model_dict[new_key + ".k_proj.bias"] = checkpoint[key][d_model : d_model * 2] + text_model_dict[new_key + ".v_proj.bias"] = checkpoint[key][d_model * 2 :] + else: + new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key) + + text_model_dict[new_key] = checkpoint[key] + + if is_accelerate_available(): + for param_name, param in text_model_dict.items(): + set_module_tensor_to_device(text_model, param_name, "cpu", value=param) + else: + if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)): + text_model_dict.pop("text_model.embeddings.position_ids", None) + + text_model.load_state_dict(text_model_dict) + + return text_model + + +def stable_unclip_image_encoder(original_config, local_files_only=False): + """ + Returns the image processor and clip image encoder for the img2img unclip pipeline. + + We currently know of two types of stable unclip models which separately use the clip and the openclip image + encoders. + """ + + image_embedder_config = original_config["model"]["params"]["embedder_config"] + + sd_clip_image_embedder_class = image_embedder_config["target"] + sd_clip_image_embedder_class = sd_clip_image_embedder_class.split(".")[-1] + + if sd_clip_image_embedder_class == "ClipImageEmbedder": + clip_model_name = image_embedder_config.params.model + + if clip_model_name == "ViT-L/14": + feature_extractor = CLIPImageProcessor() + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + else: + raise NotImplementedError(f"Unknown CLIP checkpoint name in stable diffusion checkpoint {clip_model_name}") + + elif sd_clip_image_embedder_class == "FrozenOpenCLIPImageEmbedder": + feature_extractor = CLIPImageProcessor() + image_encoder = CLIPVisionModelWithProjection.from_pretrained( + "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", local_files_only=local_files_only + ) + else: + raise NotImplementedError( + f"Unknown CLIP image embedder class in stable diffusion checkpoint {sd_clip_image_embedder_class}" + ) + + return feature_extractor, image_encoder + + +def stable_unclip_image_noising_components( + original_config, clip_stats_path: str | None = None, device: str | None = None +): + """ + Returns the noising components for the img2img and txt2img unclip pipelines. + + Converts the stability noise augmentor into + 1. a `StableUnCLIPImageNormalizer` for holding the CLIP stats + 2. a `DDPMScheduler` for holding the noise schedule + + If the noise augmentor config specifies a clip stats path, the `clip_stats_path` must be provided. + """ + noise_aug_config = original_config["model"]["params"]["noise_aug_config"] + noise_aug_class = noise_aug_config["target"] + noise_aug_class = noise_aug_class.split(".")[-1] + + if noise_aug_class == "CLIPEmbeddingNoiseAugmentation": + noise_aug_config = noise_aug_config.params + embedding_dim = noise_aug_config.timestep_dim + max_noise_level = noise_aug_config.noise_schedule_config.timesteps + beta_schedule = noise_aug_config.noise_schedule_config.beta_schedule + + image_normalizer = StableUnCLIPImageNormalizer(embedding_dim=embedding_dim) + image_noising_scheduler = DDPMScheduler(num_train_timesteps=max_noise_level, beta_schedule=beta_schedule) + + if "clip_stats_path" in noise_aug_config: + if clip_stats_path is None: + raise ValueError("This stable unclip config requires a `clip_stats_path`") + + clip_mean, clip_std = torch.load(clip_stats_path, map_location=device) + clip_mean = clip_mean[None, :] + clip_std = clip_std[None, :] + + clip_stats_state_dict = { + "mean": clip_mean, + "std": clip_std, + } + + image_normalizer.load_state_dict(clip_stats_state_dict) + else: + raise NotImplementedError(f"Unknown noise augmentor class: {noise_aug_class}") + + return image_normalizer, image_noising_scheduler + + +def convert_controlnet_checkpoint( + checkpoint, + original_config, + checkpoint_path, + image_size, + upcast_attention, + extract_ema, + use_linear_projection=None, + cross_attention_dim=None, +): + ctrlnet_config = create_unet_diffusers_config(original_config, image_size=image_size, controlnet=True) + ctrlnet_config["upcast_attention"] = upcast_attention + + ctrlnet_config.pop("sample_size") + + if use_linear_projection is not None: + ctrlnet_config["use_linear_projection"] = use_linear_projection + + if cross_attention_dim is not None: + ctrlnet_config["cross_attention_dim"] = cross_attention_dim + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + controlnet = ControlNetModel(**ctrlnet_config) + + # Some controlnet ckpt files are distributed independently from the rest of the + # model components i.e. https://huggingface.co/thibaud/controlnet-sd21/ + if "time_embed.0.weight" in checkpoint: + skip_extract_state_dict = True + else: + skip_extract_state_dict = False + + converted_ctrl_checkpoint = convert_ldm_unet_checkpoint( + checkpoint, + ctrlnet_config, + path=checkpoint_path, + extract_ema=extract_ema, + controlnet=True, + skip_extract_state_dict=skip_extract_state_dict, + ) + + if is_accelerate_available(): + for param_name, param in converted_ctrl_checkpoint.items(): + set_module_tensor_to_device(controlnet, param_name, "cpu", value=param) + else: + controlnet.load_state_dict(converted_ctrl_checkpoint) + + return controlnet + + +def download_from_original_stable_diffusion_ckpt( + checkpoint_path_or_dict: str | dict[str, torch.Tensor], + original_config_file: str = None, + image_size: int | None = None, + prediction_type: str = None, + model_type: str = None, + extract_ema: bool = False, + scheduler_type: str = "pndm", + num_in_channels: int | None = None, + upcast_attention: bool | None = None, + device: str = None, + from_safetensors: bool = False, + stable_unclip: str | None = None, + stable_unclip_prior: str | None = None, + clip_stats_path: str | None = None, + controlnet: bool | None = None, + adapter: bool | None = None, + load_safety_checker: bool = True, + safety_checker: StableDiffusionSafetyChecker | None = None, + feature_extractor: AutoFeatureExtractor | None = None, + pipeline_class: DiffusionPipeline = None, + local_files_only=False, + vae_path=None, + vae=None, + text_encoder=None, + text_encoder_2=None, + tokenizer=None, + tokenizer_2=None, + config_files=None, +) -> DiffusionPipeline: + """ + Load a Stable Diffusion pipeline object from a CompVis-style `.ckpt`/`.safetensors` file and (ideally) a `.yaml` + config file. + + Although many of the arguments can be automatically inferred, some of these rely on brittle checks against the + global step count, which will likely fail for models that have undergone further fine-tuning. Therefore, it is + recommended that you override the default values and/or supply an `original_config_file` wherever possible. + + Args: + checkpoint_path_or_dict (`str` or `dict`): Path to `.ckpt` file, or the state dict. + original_config_file (`str`): + Path to `.yaml` config file corresponding to the original architecture. If `None`, will be automatically + inferred by looking for a key that only exists in SD2.0 models. + image_size (`int`, *optional*, defaults to 512): + The image size that the model was trained on. Use 512 for Stable Diffusion v1.X and Stable Diffusion v2 + Base. Use 768 for Stable Diffusion v2. + prediction_type (`str`, *optional*): + The prediction type that the model was trained on. Use `'epsilon'` for Stable Diffusion v1.X and Stable + Diffusion v2 Base. Use `'v_prediction'` for Stable Diffusion v2. + num_in_channels (`int`, *optional*, defaults to None): + The number of input channels. If `None`, it will be automatically inferred. + scheduler_type (`str`, *optional*, defaults to 'pndm'): + Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm", + "ddim"]`. + model_type (`str`, *optional*, defaults to `None`): + The pipeline type. `None` to automatically infer, or one of `["FrozenOpenCLIPEmbedder", + "FrozenCLIPEmbedder", "PaintByExample"]`. + is_img2img (`bool`, *optional*, defaults to `False`): + Whether the model should be loaded as an img2img pipeline. + extract_ema (`bool`, *optional*, defaults to `False`): Only relevant for + checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights or not. Defaults to + `False`. Pass `True` to extract the EMA weights. EMA weights usually yield higher quality images for + inference. Non-EMA weights are usually better to continue fine-tuning. + upcast_attention (`bool`, *optional*, defaults to `None`): + Whether the attention computation should always be upcasted. This is necessary when running stable + diffusion 2.1. + device (`str`, *optional*, defaults to `None`): + The device to use. Pass `None` to determine automatically. + from_safetensors (`str`, *optional*, defaults to `False`): + If `checkpoint_path` is in `safetensors` format, load checkpoint with safetensors instead of PyTorch. + load_safety_checker (`bool`, *optional*, defaults to `True`): + Whether to load the safety checker or not. Defaults to `True`. + safety_checker (`StableDiffusionSafetyChecker`, *optional*, defaults to `None`): + Safety checker to use. If this parameter is `None`, the function will load a new instance of + [StableDiffusionSafetyChecker] by itself, if needed. + feature_extractor (`AutoFeatureExtractor`, *optional*, defaults to `None`): + Feature extractor to use. If this parameter is `None`, the function will load a new instance of + [AutoFeatureExtractor] by itself, if needed. + pipeline_class (`str`, *optional*, defaults to `None`): + The pipeline class to use. Pass `None` to determine automatically. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether or not to only look at local files (i.e., do not try to download the model). + vae (`AutoencoderKL`, *optional*, defaults to `None`): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If + this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed. + text_encoder (`CLIPTextModel`, *optional*, defaults to `None`): + An instance of [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel) + to use, specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) + variant. If this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed. + tokenizer (`CLIPTokenizer`, *optional*, defaults to `None`): + An instance of + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer) + to use. If this parameter is `None`, the function will load a new instance of [CLIPTokenizer] by itself, if + needed. + config_files (`dict[str, str]`, *optional*, defaults to `None`): + A dictionary mapping from config file names to their contents. If this parameter is `None`, the function + will load the config files by itself, if needed. Valid keys are: + - `v1`: Config file for Stable Diffusion v1 + - `v2`: Config file for Stable Diffusion v2 + - `xl`: Config file for Stable Diffusion XL + - `xl_refiner`: Config file for Stable Diffusion XL Refiner + return: A StableDiffusionPipeline object representing the passed-in `.ckpt`/`.safetensors` file. + """ + + # import pipelines here to avoid circular import error when using from_single_file method + from diffusers import ( + LDMTextToImagePipeline, + PaintByExamplePipeline, + StableDiffusionControlNetPipeline, + StableDiffusionInpaintPipeline, + StableDiffusionPipeline, + StableDiffusionUpscalePipeline, + StableDiffusionXLControlNetInpaintPipeline, + StableDiffusionXLImg2ImgPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLPipeline, + StableUnCLIPImg2ImgPipeline, + StableUnCLIPPipeline, + ) + + if prediction_type == "v-prediction": + prediction_type = "v_prediction" + + if isinstance(checkpoint_path_or_dict, str): + if from_safetensors: + from safetensors.torch import load_file as safe_load + + checkpoint = safe_load(checkpoint_path_or_dict, device="cpu") + else: + if device is None: + device = get_device() + checkpoint = torch.load(checkpoint_path_or_dict, map_location=device) + else: + checkpoint = torch.load(checkpoint_path_or_dict, map_location=device) + elif isinstance(checkpoint_path_or_dict, dict): + checkpoint = checkpoint_path_or_dict + + # Sometimes models don't have the global_step item + if "global_step" in checkpoint: + global_step = checkpoint["global_step"] + else: + logger.debug("global_step key not found in model") + global_step = None + + # NOTE: this while loop isn't great but this controlnet checkpoint has one additional + # "state_dict" key https://huggingface.co/thibaud/controlnet-canny-sd21 + while "state_dict" in checkpoint: + checkpoint = checkpoint["state_dict"] + + if original_config_file is None: + key_name_v2_1 = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight" + key_name_sd_xl_base = "conditioner.embedders.1.model.transformer.resblocks.9.mlp.c_proj.bias" + key_name_sd_xl_refiner = "conditioner.embedders.0.model.transformer.resblocks.9.mlp.c_proj.bias" + is_upscale = pipeline_class == StableDiffusionUpscalePipeline + + config_url = None + + # model_type = "v1" + if config_files is not None and "v1" in config_files: + original_config_file = config_files["v1"] + else: + config_url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml" + + if key_name_v2_1 in checkpoint and checkpoint[key_name_v2_1].shape[-1] == 1024: + # model_type = "v2" + if config_files is not None and "v2" in config_files: + original_config_file = config_files["v2"] + else: + config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml" + if global_step == 110000: + # v2.1 needs to upcast attention + upcast_attention = True + elif key_name_sd_xl_base in checkpoint: + # only base xl has two text embedders + if config_files is not None and "xl" in config_files: + original_config_file = config_files["xl"] + else: + config_url = "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml" + elif key_name_sd_xl_refiner in checkpoint: + # only refiner xl has embedder and one text embedders + if config_files is not None and "xl_refiner" in config_files: + original_config_file = config_files["xl_refiner"] + else: + config_url = "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_refiner.yaml" + + if is_upscale: + config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/x4-upscaling.yaml" + + if config_url is not None: + original_config_file = BytesIO(requests.get(config_url, timeout=DIFFUSERS_REQUEST_TIMEOUT).content) + else: + with open(original_config_file, "r") as f: + original_config_file = f.read() + else: + with open(original_config_file, "r") as f: + original_config_file = f.read() + + original_config = yaml.safe_load(original_config_file) + + # Convert the text model. + if ( + model_type is None + and "cond_stage_config" in original_config["model"]["params"] + and original_config["model"]["params"]["cond_stage_config"] is not None + ): + model_type = original_config["model"]["params"]["cond_stage_config"]["target"].split(".")[-1] + logger.debug(f"no `model_type` given, `model_type` inferred as: {model_type}") + elif model_type is None and original_config["model"]["params"]["network_config"] is not None: + if original_config["model"]["params"]["network_config"]["params"]["context_dim"] == 2048: + model_type = "SDXL" + else: + model_type = "SDXL-Refiner" + if image_size is None: + image_size = 1024 + + if pipeline_class is None: + # Check if we have a SDXL or SD model and initialize default pipeline + if model_type not in ["SDXL", "SDXL-Refiner"]: + pipeline_class = StableDiffusionPipeline if not controlnet else StableDiffusionControlNetPipeline + else: + pipeline_class = StableDiffusionXLPipeline if model_type == "SDXL" else StableDiffusionXLImg2ImgPipeline + + if num_in_channels is None and pipeline_class in [ + StableDiffusionInpaintPipeline, + StableDiffusionXLInpaintPipeline, + StableDiffusionXLControlNetInpaintPipeline, + ]: + num_in_channels = 9 + if num_in_channels is None and pipeline_class == StableDiffusionUpscalePipeline: + num_in_channels = 7 + elif num_in_channels is None: + num_in_channels = 4 + + if "unet_config" in original_config["model"]["params"]: + original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels + elif "network_config" in original_config["model"]["params"]: + original_config["model"]["params"]["network_config"]["params"]["in_channels"] = num_in_channels + + if ( + "parameterization" in original_config["model"]["params"] + and original_config["model"]["params"]["parameterization"] == "v" + ): + if prediction_type is None: + # NOTE: For stable diffusion 2 base it is recommended to pass `prediction_type=="epsilon"` + # as it relies on a brittle global step parameter here + prediction_type = "epsilon" if global_step == 875000 else "v_prediction" + if image_size is None: + # NOTE: For stable diffusion 2 base one has to pass `image_size==512` + # as it relies on a brittle global step parameter here + image_size = 512 if global_step == 875000 else 768 + else: + if prediction_type is None: + prediction_type = "epsilon" + if image_size is None: + image_size = 512 + + if controlnet is None and "control_stage_config" in original_config["model"]["params"]: + path = checkpoint_path_or_dict if isinstance(checkpoint_path_or_dict, str) else "" + controlnet = convert_controlnet_checkpoint( + checkpoint, original_config, path, image_size, upcast_attention, extract_ema + ) + + if "timesteps" in original_config["model"]["params"]: + num_train_timesteps = original_config["model"]["params"]["timesteps"] + else: + num_train_timesteps = 1000 + + if model_type in ["SDXL", "SDXL-Refiner"]: + scheduler_dict = { + "beta_schedule": "scaled_linear", + "beta_start": 0.00085, + "beta_end": 0.012, + "interpolation_type": "linear", + "num_train_timesteps": num_train_timesteps, + "prediction_type": "epsilon", + "sample_max_value": 1.0, + "set_alpha_to_one": False, + "skip_prk_steps": True, + "steps_offset": 1, + "timestep_spacing": "leading", + } + scheduler = EulerDiscreteScheduler.from_config(scheduler_dict) + scheduler_type = "euler" + else: + if "linear_start" in original_config["model"]["params"]: + beta_start = original_config["model"]["params"]["linear_start"] + else: + beta_start = 0.02 + + if "linear_end" in original_config["model"]["params"]: + beta_end = original_config["model"]["params"]["linear_end"] + else: + beta_end = 0.085 + scheduler = DDIMScheduler( + beta_end=beta_end, + beta_schedule="scaled_linear", + beta_start=beta_start, + num_train_timesteps=num_train_timesteps, + steps_offset=1, + clip_sample=False, + set_alpha_to_one=False, + prediction_type=prediction_type, + ) + # make sure scheduler works correctly with DDIM + scheduler.register_to_config(clip_sample=False) + + if scheduler_type == "pndm": + config = dict(scheduler.config) + config["skip_prk_steps"] = True + scheduler = PNDMScheduler.from_config(config) + elif scheduler_type == "lms": + scheduler = LMSDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "heun": + scheduler = HeunDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "euler": + scheduler = EulerDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "euler-ancestral": + scheduler = EulerAncestralDiscreteScheduler.from_config(scheduler.config) + elif scheduler_type == "dpm": + scheduler = DPMSolverMultistepScheduler.from_config(scheduler.config) + elif scheduler_type == "ddim": + scheduler = scheduler + else: + raise ValueError(f"Scheduler of type {scheduler_type} doesn't exist!") + + if pipeline_class == StableDiffusionUpscalePipeline: + image_size = original_config["model"]["params"]["unet_config"]["params"]["image_size"] + + # Convert the UNet2DConditionModel model. + unet_config = create_unet_diffusers_config(original_config, image_size=image_size) + unet_config["upcast_attention"] = upcast_attention + + path = checkpoint_path_or_dict if isinstance(checkpoint_path_or_dict, str) else "" + converted_unet_checkpoint = convert_ldm_unet_checkpoint( + checkpoint, unet_config, path=path, extract_ema=extract_ema + ) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + unet = UNet2DConditionModel(**unet_config) + + if is_accelerate_available(): + if model_type not in ["SDXL", "SDXL-Refiner"]: # SBM Delay this. + for param_name, param in converted_unet_checkpoint.items(): + set_module_tensor_to_device(unet, param_name, "cpu", value=param) + else: + unet.load_state_dict(converted_unet_checkpoint) + + # Convert the VAE model. + if vae_path is None and vae is None: + vae_config = create_vae_diffusers_config(original_config, image_size=image_size) + converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config) + + if ( + "model" in original_config + and "params" in original_config["model"] + and "scale_factor" in original_config["model"]["params"] + ): + vae_scaling_factor = original_config["model"]["params"]["scale_factor"] + else: + vae_scaling_factor = 0.18215 # default SD scaling factor + + vae_config["scaling_factor"] = vae_scaling_factor + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + vae = AutoencoderKL(**vae_config) + + if is_accelerate_available(): + for param_name, param in converted_vae_checkpoint.items(): + set_module_tensor_to_device(vae, param_name, "cpu", value=param) + else: + vae.load_state_dict(converted_vae_checkpoint) + elif vae is None: + vae = AutoencoderKL.from_pretrained(vae_path, local_files_only=local_files_only) + + if model_type == "FrozenOpenCLIPEmbedder": + config_name = "stabilityai/stable-diffusion-2" + config_kwargs = {"subfolder": "text_encoder"} + + if text_encoder is None: + text_model = convert_open_clip_checkpoint( + checkpoint, config_name, local_files_only=local_files_only, **config_kwargs + ) + else: + text_model = text_encoder + + try: + tokenizer = CLIPTokenizer.from_pretrained( + "stabilityai/stable-diffusion-2", subfolder="tokenizer", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'stabilityai/stable-diffusion-2'." + ) + + if stable_unclip is None: + if controlnet: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + controlnet=controlnet, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + if hasattr(pipe, "requires_safety_checker"): + pipe.requires_safety_checker = False + + elif pipeline_class == StableDiffusionUpscalePipeline: + scheduler = DDIMScheduler.from_pretrained( + "stabilityai/stable-diffusion-x4-upscaler", subfolder="scheduler" + ) + low_res_scheduler = DDPMScheduler.from_pretrained( + "stabilityai/stable-diffusion-x4-upscaler", subfolder="low_res_scheduler" + ) + + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + low_res_scheduler=low_res_scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + + else: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + if hasattr(pipe, "requires_safety_checker"): + pipe.requires_safety_checker = False + + else: + image_normalizer, image_noising_scheduler = stable_unclip_image_noising_components( + original_config, clip_stats_path=clip_stats_path, device=device + ) + + if stable_unclip == "img2img": + feature_extractor, image_encoder = stable_unclip_image_encoder(original_config) + + pipe = StableUnCLIPImg2ImgPipeline( + # image encoding components + feature_extractor=feature_extractor, + image_encoder=image_encoder, + # image noising components + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + # regular denoising components + tokenizer=tokenizer, + text_encoder=text_model, + unet=unet, + scheduler=scheduler, + # vae + vae=vae, + ) + elif stable_unclip == "txt2img": + if stable_unclip_prior is None or stable_unclip_prior == "karlo": + karlo_model = "kakaobrain/karlo-v1-alpha" + prior = PriorTransformer.from_pretrained( + karlo_model, subfolder="prior", local_files_only=local_files_only + ) + + try: + prior_tokenizer = CLIPTokenizer.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + prior_text_model = CLIPTextModelWithProjection.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + + prior_scheduler = UnCLIPScheduler.from_pretrained( + karlo_model, subfolder="prior_scheduler", local_files_only=local_files_only + ) + prior_scheduler = DDPMScheduler.from_config(prior_scheduler.config) + else: + raise NotImplementedError(f"unknown prior for stable unclip model: {stable_unclip_prior}") + + pipe = StableUnCLIPPipeline( + # prior components + prior_tokenizer=prior_tokenizer, + prior_text_encoder=prior_text_model, + prior=prior, + prior_scheduler=prior_scheduler, + # image noising components + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + # regular denoising components + tokenizer=tokenizer, + text_encoder=text_model, + unet=unet, + scheduler=scheduler, + # vae + vae=vae, + ) + else: + raise NotImplementedError(f"unknown `stable_unclip` type: {stable_unclip}") + elif model_type == "PaintByExample": + vision_model = convert_paint_by_example_checkpoint(checkpoint) + try: + tokenizer = CLIPTokenizer.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + try: + feature_extractor = AutoFeatureExtractor.from_pretrained( + "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the feature_extractor in the following path: 'CompVis/stable-diffusion-safety-checker'." + ) + pipe = PaintByExamplePipeline( + vae=vae, + image_encoder=vision_model, + unet=unet, + scheduler=scheduler, + safety_checker=None, + feature_extractor=feature_extractor, + ) + elif model_type == "FrozenCLIPEmbedder": + text_model = convert_ldm_clip_checkpoint( + checkpoint, local_files_only=local_files_only, text_encoder=text_encoder + ) + try: + tokenizer = ( + CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only) + if tokenizer is None + else tokenizer + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + + if load_safety_checker: + safety_checker = StableDiffusionSafetyChecker.from_pretrained( + "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only + ) + feature_extractor = AutoFeatureExtractor.from_pretrained( + "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only + ) + + if controlnet: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + else: + pipe = pipeline_class( + vae=vae, + text_encoder=text_model, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + elif model_type in ["SDXL", "SDXL-Refiner"]: + is_refiner = model_type == "SDXL-Refiner" + + if (is_refiner is False) and (tokenizer is None): + try: + tokenizer = CLIPTokenizer.from_pretrained( + "openai/clip-vit-large-patch14", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'." + ) + + if (is_refiner is False) and (text_encoder is None): + text_encoder = convert_ldm_clip_checkpoint(checkpoint, local_files_only=local_files_only) + + if tokenizer_2 is None: + try: + tokenizer_2 = CLIPTokenizer.from_pretrained( + "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!", local_files_only=local_files_only + ) + except Exception: + raise ValueError( + f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'laion/CLIP-ViT-bigG-14-laion2B-39B-b160k' with `pad_token` set to '!'." + ) + + if text_encoder_2 is None: + config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k" + config_kwargs = {"projection_dim": 1280} + prefix = "conditioner.embedders.0.model." if is_refiner else "conditioner.embedders.1.model." + + text_encoder_2 = convert_open_clip_checkpoint( + checkpoint, + config_name, + prefix=prefix, + has_projection=True, + local_files_only=local_files_only, + **config_kwargs, + ) + + if is_accelerate_available(): # SBM Now move model to cpu. + for param_name, param in converted_unet_checkpoint.items(): + set_module_tensor_to_device(unet, param_name, "cpu", value=param) + + if controlnet: + pipe = pipeline_class( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + unet=unet, + controlnet=controlnet, + scheduler=scheduler, + force_zeros_for_empty_prompt=True, + ) + elif adapter: + pipe = pipeline_class( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + text_encoder_2=text_encoder_2, + tokenizer_2=tokenizer_2, + unet=unet, + adapter=adapter, + scheduler=scheduler, + force_zeros_for_empty_prompt=True, + ) + + else: + pipeline_kwargs = { + "vae": vae, + "text_encoder": text_encoder, + "tokenizer": tokenizer, + "text_encoder_2": text_encoder_2, + "tokenizer_2": tokenizer_2, + "unet": unet, + "scheduler": scheduler, + } + + if (pipeline_class == StableDiffusionXLImg2ImgPipeline) or ( + pipeline_class == StableDiffusionXLInpaintPipeline + ): + pipeline_kwargs.update({"requires_aesthetics_score": is_refiner}) + + if is_refiner: + pipeline_kwargs.update({"force_zeros_for_empty_prompt": False}) + + pipe = pipeline_class(**pipeline_kwargs) + else: + text_config = create_ldm_bert_config(original_config) + text_model = convert_ldm_bert_checkpoint(checkpoint, text_config) + tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased", local_files_only=local_files_only) + pipe = LDMTextToImagePipeline(vqvae=vae, bert=text_model, tokenizer=tokenizer, unet=unet, scheduler=scheduler) + + return pipe + + +def download_controlnet_from_original_ckpt( + checkpoint_path: str, + original_config_file: str, + image_size: int = 512, + extract_ema: bool = False, + num_in_channels: int | None = None, + upcast_attention: bool | None = None, + device: str = None, + from_safetensors: bool = False, + use_linear_projection: bool | None = None, + cross_attention_dim: bool | None = None, +) -> DiffusionPipeline: + if from_safetensors: + from safetensors import safe_open + + checkpoint = {} + with safe_open(checkpoint_path, framework="pt", device="cpu") as f: + for key in f.keys(): + checkpoint[key] = f.get_tensor(key) + else: + if device is None: + device = get_device() + checkpoint = torch.load(checkpoint_path, map_location=device) + else: + checkpoint = torch.load(checkpoint_path, map_location=device) + + # NOTE: this while loop isn't great but this controlnet checkpoint has one additional + # "state_dict" key https://huggingface.co/thibaud/controlnet-canny-sd21 + while "state_dict" in checkpoint: + checkpoint = checkpoint["state_dict"] + + with open(original_config_file, "r") as f: + original_config_file = f.read() + original_config = yaml.safe_load(original_config_file) + + if num_in_channels is not None: + original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels + + if "control_stage_config" not in original_config["model"]["params"]: + raise ValueError("`control_stage_config` not present in original config") + + controlnet = convert_controlnet_checkpoint( + checkpoint, + original_config, + checkpoint_path, + image_size, + upcast_attention, + extract_ema, + use_linear_projection=use_linear_projection, + cross_attention_dim=cross_attention_dim, + ) + + return controlnet diff --git a/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..a7a81d2c13556291569386d45fcdcef32e6a3038 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py @@ -0,0 +1,490 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) + + +class OnnxStableDiffusionPipeline(DiffusionPipeline): + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler, + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def _encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int | None, + do_classifier_free_guidance: bool, + negative_prompt: str | None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def check_inputs( + self, + prompt: str | list[str], + height: int | None, + width: int | None, + callback_steps: int, + negative_prompt: str | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = 512, + width: int | None = 512, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: np.random.RandomState | None = None, + latents: np.ndarray | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, np.ndarray], None] | None = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` + is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + One or a list of [numpy generator(s)](TODO) to make generation deterministic. + latents (`np.ndarray`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if generator is None: + generator = np.random + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # get the initial random noise unless the user supplied it + latents_dtype = prompt_embeds.dtype + latents_shape = (batch_size * num_images_per_prompt, 4, height // 8, width // 8) + if latents is None: + latents = generator.randn(*latents_shape).astype(latents_dtype) + elif latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) + latent_model_input = latent_model_input.cpu().numpy() + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds) + noise_pred = noise_pred[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + scheduler_output = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ) + latents = scheduler_output.prev_sample.numpy() + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + +class StableDiffusionOnnxPipeline(OnnxStableDiffusionPipeline): + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler, + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + ): + deprecation_message = "Please use `OnnxStableDiffusionPipeline` instead of `StableDiffusionOnnxPipeline`." + deprecate("StableDiffusionOnnxPipeline", "1.0.0", deprecation_message) + super().__init__( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..543881ab87bf4ebaf439fa3086aa71729f84ffa3 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py @@ -0,0 +1,550 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import PIL_INTERPOLATION, deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess with 8->64 +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class OnnxStableDiffusionImg2ImgPipeline(DiffusionPipeline): + r""" + Pipeline for text-guided image to image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler, + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int | None, + do_classifier_free_guidance: bool, + negative_prompt: str | None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def check_inputs( + self, + prompt: str | list[str], + callback_steps: int, + negative_prompt: str | list[str] | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def __call__( + self, + prompt: str | list[str], + image: np.ndarray | PIL.Image.Image = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: np.random.RandomState | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, np.ndarray], None] | None = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`np.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if generator is None: + generator = np.random + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + image = preprocess(image).cpu().numpy() + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + latents_dtype = prompt_embeds.dtype + image = image.astype(latents_dtype) + # encode the init image into latents and scale the latents + init_latents = self.vae_encoder(sample=image)[0] + init_latents = 0.18215 * init_latents + + if isinstance(prompt, str): + prompt = [prompt] + if len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {len(prompt)} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = len(prompt) // init_latents.shape[0] + init_latents = np.concatenate([init_latents] * additional_image_per_prompt * num_images_per_prompt, axis=0) + elif len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {len(prompt)} text prompts." + ) + else: + init_latents = np.concatenate([init_latents] * num_images_per_prompt, axis=0) + + # get the original timestep using init_timestep + offset = self.scheduler.config.get("steps_offset", 0) + init_timestep = int(num_inference_steps * strength) + offset + init_timestep = min(init_timestep, num_inference_steps) + + timesteps = self.scheduler.timesteps.numpy()[-init_timestep] + timesteps = np.array([timesteps] * batch_size * num_images_per_prompt) + + # add noise to latents using the timesteps + noise = generator.randn(*init_latents.shape).astype(latents_dtype) + init_latents = self.scheduler.add_noise( + torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps) + ) + init_latents = init_latents.numpy() + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + latents = init_latents + + t_start = max(num_inference_steps - init_timestep + offset, 0) + timesteps = self.scheduler.timesteps[t_start:].numpy() + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) + latent_model_input = latent_model_input.cpu().numpy() + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[ + 0 + ] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + scheduler_output = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ) + latents = scheduler_output.prev_sample.numpy() + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + # safety_checker does not support batched inputs yet + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..865cdd82cd5d164447edc84ea7ef8968e2119d2d --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py @@ -0,0 +1,564 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler +from ...utils import PIL_INTERPOLATION, deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +NUM_UNET_INPUT_CHANNELS = 9 +NUM_LATENT_CHANNELS = 4 + + +def prepare_mask_and_masked_image(image, mask, latents_shape): + image = np.array(image.convert("RGB").resize((latents_shape[1] * 8, latents_shape[0] * 8))) + image = image[None].transpose(0, 3, 1, 2) + image = image.astype(np.float32) / 127.5 - 1.0 + + image_mask = np.array(mask.convert("L").resize((latents_shape[1] * 8, latents_shape[0] * 8))) + masked_image = image * (image_mask < 127.5) + + mask = mask.resize((latents_shape[1], latents_shape[0]), PIL_INTERPOLATION["nearest"]) + mask = np.array(mask.convert("L")) + mask = mask.astype(np.float32) / 255.0 + mask = mask[None, None] + mask[mask < 0.5] = 0 + mask[mask >= 0.5] = 1 + + return mask, masked_image + + +class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + vae_encoder: OnnxRuntimeModel + vae_decoder: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae_encoder: OnnxRuntimeModel, + vae_decoder: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: CLIPTokenizer, + unet: OnnxRuntimeModel, + scheduler: DDIMScheduler | PNDMScheduler | LMSDiscreteScheduler, + safety_checker: OnnxRuntimeModel, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + logger.info("`OnnxStableDiffusionInpaintPipeline` is experimental and will very likely change in the future.") + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae_encoder=vae_encoder, + vae_decoder=vae_decoder, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int | None, + do_classifier_free_guidance: bool, + negative_prompt: str | None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline.check_inputs + def check_inputs( + self, + prompt: str | list[str], + height: int | None, + width: int | None, + callback_steps: int, + negative_prompt: str | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str], + image: PIL.Image.Image, + mask_image: PIL.Image.Image, + height: int | None = 512, + width: int | None = 512, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: np.random.RandomState | None = None, + latents: np.ndarray | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, np.ndarray], None] | None = None, + callback_steps: int = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + latents (`np.ndarray`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if generator is None: + generator = np.random + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + num_channels_latents = NUM_LATENT_CHANNELS + latents_shape = (batch_size * num_images_per_prompt, num_channels_latents, height // 8, width // 8) + latents_dtype = prompt_embeds.dtype + if latents is None: + latents = generator.randn(*latents_shape).astype(latents_dtype) + else: + if latents.shape != latents_shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}") + + # prepare mask and masked_image + mask, masked_image = prepare_mask_and_masked_image(image, mask_image, latents_shape[-2:]) + mask = mask.astype(latents.dtype) + masked_image = masked_image.astype(latents.dtype) + + masked_image_latents = self.vae_encoder(sample=masked_image)[0] + masked_image_latents = 0.18215 * masked_image_latents + + # duplicate mask and masked_image_latents for each generation per prompt + mask = mask.repeat(batch_size * num_images_per_prompt, 0) + masked_image_latents = masked_image_latents.repeat(batch_size * num_images_per_prompt, 0) + + mask = np.concatenate([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + np.concatenate([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + + unet_input_channels = NUM_UNET_INPUT_CHANNELS + if num_channels_latents + num_channels_mask + num_channels_masked_image != unet_input_channels: + raise ValueError( + "Incorrect configuration settings! The config of `pipeline.unet` expects" + f" {unet_input_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + + # set timesteps + self.scheduler.set_timesteps(num_inference_steps) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + # concat latents, mask, masked_image_latnets in the channel dimension + latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t) + latent_model_input = latent_model_input.cpu().numpy() + latent_model_input = np.concatenate([latent_model_input, mask, masked_image_latents], axis=1) + + # predict the noise residual + timestep = np.array([t], dtype=timestep_dtype) + noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[ + 0 + ] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + scheduler_output = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ) + latents = scheduler_output.prev_sample.numpy() + + # call the callback, if provided + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + latents = 1 / 0.18215 * latents + # image = self.vae_decoder(latent_sample=latents)[0] + # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1 + image = np.concatenate( + [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])] + ) + + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + # safety_checker does not support batched inputs yet + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py new file mode 100644 index 0000000000000000000000000000000000000000..c36204203d73ccb6145fbb067b04e553c1e4eaef --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py @@ -0,0 +1,586 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTokenizer + +from ...configuration_utils import FrozenDict +from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers +from ...utils import deprecate, logging +from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel +from ..pipeline_utils import DiffusionPipeline +from . import StableDiffusionPipelineOutput + + +logger = logging.get_logger(__name__) + + +def preprocess(image): + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 32 + + image = [np.array(i.resize((w, h)))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + + return image + + +class OnnxStableDiffusionUpscalePipeline(DiffusionPipeline): + vae: OnnxRuntimeModel + text_encoder: OnnxRuntimeModel + tokenizer: CLIPTokenizer + unet: OnnxRuntimeModel + low_res_scheduler: DDPMScheduler + scheduler: KarrasDiffusionSchedulers + safety_checker: OnnxRuntimeModel + feature_extractor: CLIPImageProcessor + + _optional_components = ["safety_checker", "feature_extractor"] + _is_onnx = True + + def __init__( + self, + vae: OnnxRuntimeModel, + text_encoder: OnnxRuntimeModel, + tokenizer: Any, + unet: OnnxRuntimeModel, + low_res_scheduler: DDPMScheduler, + scheduler: KarrasDiffusionSchedulers, + safety_checker: OnnxRuntimeModel | None = None, + feature_extractor: CLIPImageProcessor | None = None, + max_noise_level: int = 350, + num_latent_channels=4, + num_unet_input_channels=7, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + low_res_scheduler=low_res_scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.register_to_config( + max_noise_level=max_noise_level, + num_latent_channels=num_latent_channels, + num_unet_input_channels=num_unet_input_channels, + ) + + def check_inputs( + self, + prompt: str | list[str], + image, + noise_level, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, np.ndarray) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}" + ) + + # verify batch size of prompt and image are same if image is a list or tensor or numpy array + if isinstance(image, (list, np.ndarray)): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if isinstance(image, list): + image_batch_size = len(image) + else: + image_batch_size = image.shape[0] + if batch_size != image_batch_size: + raise ValueError( + f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}." + " Please make sure that passed `prompt` matches the batch size of `image`." + ) + + # check noise level + if noise_level > self.config.max_noise_level: + raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, generator, latents=None): + shape = (batch_size, num_channels_latents, height, width) + if latents is None: + latents = generator.randn(*shape).astype(dtype) + elif latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + + return latents + + def decode_latents(self, latents): + latents = 1 / 0.08333 * latents + image = self.vae(latent_sample=latents)[0] + image = np.clip(image / 2 + 0.5, 0, 1) + image = image.transpose((0, 2, 3, 1)) + return image + + def _encode_prompt( + self, + prompt: str | list[str], + num_images_per_prompt: int | None, + do_classifier_free_guidance: bool, + negative_prompt: str | None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`): + prompt to be encoded + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # get prompt text embeddings + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="np", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids + + if not np.array_equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0] + + prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] * batch_size + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="np", + ) + negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0] + + if do_classifier_free_guidance: + negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def __call__( + self, + prompt: str | list[str], + image: np.ndarray | PIL.Image.Image | list[PIL.Image.Image], + num_inference_steps: int = 75, + guidance_scale: float = 9.0, + noise_level: int = 20, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: np.random.RandomState | list[np.random.RandomState] | None = None, + latents: np.ndarray | None = None, + prompt_embeds: np.ndarray | None = None, + negative_prompt_embeds: np.ndarray | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, np.ndarray], None] | None = None, + callback_steps: int | None = 1, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide the image generation. + image (`np.ndarray` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + noise_level (`float`, defaults to 0.2): + Deteremines the amount of noise to add to the initial image before performing upscaling. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`np.random.RandomState`, *optional*): + A np.random.RandomState to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`np.ndarray`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`np.ndarray`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + image, + noise_level, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if generator is None: + generator = np.random + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + prompt_embeds = self._encode_prompt( + prompt, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + latents_dtype = prompt_embeds.dtype + image = preprocess(image).cpu().numpy() + height, width = image.shape[2:] + + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + self.config.num_latent_channels, + height, + width, + latents_dtype, + generator, + ) + image = image.astype(latents_dtype) + + self.scheduler.set_timesteps(num_inference_steps) + timesteps = self.scheduler.timesteps + + # Scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + + # 5. Add noise to image + noise_level = np.array([noise_level]).astype(np.int64) + noise = generator.randn(*image.shape).astype(latents_dtype) + + image = self.low_res_scheduler.add_noise( + torch.from_numpy(image), torch.from_numpy(noise), torch.from_numpy(noise_level) + ) + image = image.numpy() + + batch_multiplier = 2 if do_classifier_free_guidance else 1 + image = np.concatenate([image] * batch_multiplier * num_images_per_prompt) + noise_level = np.concatenate([noise_level] * image.shape[0]) + + # 7. Check that sizes of image and latents match + num_channels_image = image.shape[1] + if self.config.num_latent_channels + num_channels_image != self.config.num_unet_input_channels: + raise ValueError( + "Incorrect configuration settings! The config of `pipeline.unet` expects" + f" {self.config.num_unet_input_channels} but received `num_channels_latents`: {self.config.num_latent_channels} +" + f" `num_channels_image`: {num_channels_image} " + f" = {self.config.num_latent_channels + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + timestep_dtype = next( + (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)" + ) + timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype] + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = np.concatenate([latent_model_input, image], axis=1) + + # timestep to tensor + timestep = np.array([t], dtype=timestep_dtype) + + # predict the noise residual + noise_pred = self.unet( + sample=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + class_labels=noise_level, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step( + torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs + ).prev_sample + latents = latents.numpy() + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + # 10. Post-processing + image = self.decode_latents(latents) + + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor( + self.numpy_to_pil(image), return_tensors="np" + ).pixel_values.astype(image.dtype) + + images, has_nsfw_concept = [], [] + for i in range(image.shape[0]): + image_i, has_nsfw_concept_i = self.safety_checker( + clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1] + ) + images.append(image_i) + has_nsfw_concept.append(has_nsfw_concept_i[0]) + image = np.concatenate(images) + else: + has_nsfw_concept = None + + if output_type == "pil": + image = self.numpy_to_pil(image) + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_output.py b/diffusers/pipelines/stable_diffusion/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..22b9433a26d0783cf0264a323996d068425fddfe --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_output.py @@ -0,0 +1,24 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class StableDiffusionPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width, + num_channels)`. + nsfw_content_detected (`list[bool]`) + list indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or + `None` if safety checking could not be performed. + """ + + images: list[PIL.Image.Image] | np.ndarray + nsfw_content_detected: list[bool] | None diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..d28bb2a9fe59a56dd3fa74c129e29478d49e4175 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py @@ -0,0 +1,1107 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import inspect +from typing import Any, Callable + +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + """ + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + self._is_unet_config_sample_size_int = unet is not None and isinstance(unet.config.sample_size, int) + is_unet_sample_size_less_64 = ( + unet is not None + and hasattr(unet.config, "sample_size") + and self._is_unet_config_sample_size_int + and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + if not height or not width: + height = ( + self.unet.config.sample_size + if self._is_unet_config_sample_size_int + else self.unet.config.sample_size[0] + ) + width = ( + self.unet.config.sample_size + if self._is_unet_config_sample_size_int + else self.unet.config.sample_size[1] + ) + height, width = height * self.vae_scale_factor, width * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare timesteps + timestep_device = device + if XLA_AVAILABLE: + timestep_device = "cpu" + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) + else None + ) + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + if hasattr(self.scheduler, "scale_model_input"): + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + if XLA_AVAILABLE: + xm.mark_step() + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py new file mode 100644 index 0000000000000000000000000000000000000000..977de5d7fb398ada76aae987c70e17afa07a0223 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py @@ -0,0 +1,897 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import contextlib +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPTextModel, CLIPTokenizer, DPTForDepthEstimation, DPTImageProcessor + +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin): + r""" + Pipeline for text-guided depth-based image-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "depth_mask"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + depth_estimator: DPTForDepthEstimation, + feature_extractor: DPTImageProcessor, + ): + super().__init__() + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + depth_estimator=depth_estimator, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + def prepare_depth_map(self, image, depth_map, batch_size, do_classifier_free_guidance, dtype, device): + if isinstance(image, PIL.Image.Image): + image = [image] + else: + image = list(image) + + if isinstance(image[0], PIL.Image.Image): + width, height = image[0].size + elif isinstance(image[0], np.ndarray): + width, height = image[0].shape[:-1] + else: + height, width = image[0].shape[-2:] + + if depth_map is None: + pixel_values = self.feature_extractor(images=image, return_tensors="pt").pixel_values + pixel_values = pixel_values.to(device=device, dtype=dtype) + # The DPT-Hybrid model uses batch-norm layers which are not compatible with fp16. + # So we use `torch.autocast` here for half precision inference. + if torch.backends.mps.is_available(): + autocast_ctx = contextlib.nullcontext() + logger.warning( + "The DPT-Hybrid model uses batch-norm layers which are not compatible with fp16, but autocast is not yet supported on MPS." + ) + else: + autocast_ctx = torch.autocast(device.type, dtype=dtype) + + with autocast_ctx: + depth_map = self.depth_estimator(pixel_values).predicted_depth + else: + depth_map = depth_map.to(device=device, dtype=dtype) + + depth_map = torch.nn.functional.interpolate( + depth_map.unsqueeze(1), + size=(height // self.vae_scale_factor, width // self.vae_scale_factor), + mode="bicubic", + align_corners=False, + ) + + depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True) + depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True) + depth_map = 2.0 * (depth_map - depth_min) / (depth_max - depth_min) - 1.0 + depth_map = depth_map.to(dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if depth_map.shape[0] < batch_size: + repeat_by = batch_size // depth_map.shape[0] + depth_map = depth_map.repeat(repeat_by, 1, 1, 1) + + depth_map = torch.cat([depth_map] * 2) if do_classifier_free_guidance else depth_map + return depth_map + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + depth_map: torch.Tensor | None = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be used as the starting point. Can accept image + latents as `image` only if `depth_map` is not `None`. + depth_map (`torch.Tensor`, *optional*): + Depth prediction to be used as additional conditioning for the image generation process. If not + defined, it automatically predicts the depth with `self.depth_estimator`. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + ```py + >>> import torch + >>> import requests + >>> from PIL import Image + + >>> from diffusers import StableDiffusionDepth2ImgPipeline + + >>> pipe = StableDiffusionDepth2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-depth", + ... torch_dtype=torch.float16, + ... ) + >>> pipe.to("cuda") + + + >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" + >>> init_image = Image.open(requests.get(url, stream=True).raw) + >>> prompt = "two tigers" + >>> n_prompt = "bad, deformed, ugly, bad anotomy" + >>> image = pipe(prompt=prompt, image=init_image, negative_prompt=n_prompt, strength=0.7).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + # 1. Check inputs + self.check_inputs( + prompt, + strength, + callback_steps, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare depth mask + depth_mask = self.prepare_depth_map( + image, + depth_map, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + prompt_embeds.dtype, + device, + ) + + # 5. Preprocess image + image = self.image_processor.preprocess(image) + + # 6. Set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 7. Prepare latent variables + latents = self.prepare_latents( + image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = torch.cat([latent_model_input, depth_mask], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=self.cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + depth_mask = callback_outputs.pop("depth_mask", depth_mask) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py new file mode 100644 index 0000000000000000000000000000000000000000..15b8daf334ed7bb5c0dd3edd2a690c3813815b29 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py @@ -0,0 +1,439 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Callable + +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...configuration_utils import FrozenDict +from ...image_processor import VaeImageProcessor +from ...models import AutoencoderKL, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class StableDiffusionImageVariationPipeline(DiffusionPipeline, StableDiffusionMixin): + r""" + Pipeline to generate image variations from an input image using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + # TODO: feature_extractor is required to encode images (if they are in PIL format), + # we should give a descriptive message if the pipeline doesn't have one. + _optional_components = ["safety_checker"] + model_cpu_offload_seq = "image_encoder->unet->vae" + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + image_encoder: CLIPVisionModelWithProjection, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + image_encoder=image_encoder, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def _encode_image(self, image, device, num_images_per_prompt, do_classifier_free_guidance): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + image_embeddings = image_embeddings.unsqueeze(1) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(image_embeddings) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings]) + + return image_embeddings + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs(self, image, height, width, callback_steps): + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `list[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + image: PIL.Image.Image | list[PIL.Image.Image] | torch.Tensor, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + guidance_scale: float = 7.5, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.Tensor`): + Image or images to guide image generation. If you provide a tensor, it needs to be compatible with + [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json). + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + + Examples: + + ```py + from diffusers import StableDiffusionImageVariationPipeline + from PIL import Image + from io import BytesIO + import requests + + pipe = StableDiffusionImageVariationPipeline.from_pretrained( + "lambdalabs/sd-image-variations-diffusers", revision="v2.0" + ) + pipe = pipe.to("cuda") + + url = "https://lh3.googleusercontent.com/y-iFOHfLTwkuQSUegpwDdgKmOjRSTvPxat63dQLB25xkTs4lhIbRUFeNBWZzYf370g=s1200" + + response = requests.get(url) + image = Image.open(BytesIO(response.content)).convert("RGB") + + out = pipe(image, num_images_per_prompt=3, guidance_scale=15) + out["images"][0].save("result.jpg") + ``` + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width, callback_steps) + + # 2. Define call parameters + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input image + image_embeddings = self._encode_image(image, device, num_images_per_prompt, do_classifier_free_guidance) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + self.maybe_free_model_hooks() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, image_embeddings.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..719be92583416881bfb534617e70ea23fab20083 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py @@ -0,0 +1,1163 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import requests + >>> import torch + >>> from PIL import Image + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionImg2ImgPipeline + + >>> device = "cuda" + >>> model_id_or_path = "stable-diffusion-v1-5/stable-diffusion-v1-5" + >>> pipe = StableDiffusionImg2ImgPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16) + >>> pipe = pipe.to(device) + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + + >>> response = requests.get(url) + >>> init_image = Image.open(BytesIO(response.content)).convert("RGB") + >>> init_image = init_image.resize((768, 512)) + + >>> prompt = "A fantasy landscape, trending on artstation" + + >>> images = pipe(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images + >>> images[0].save("fantasy_landscape.png") + ``` +""" + + +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-guided image-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + strength, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + strength: float = 0.8, + num_inference_steps: int | None = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float | None = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float | None = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.8): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + strength, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. set timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 6. Prepare latent variables + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 7.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 8. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..96794eaa297a92292f7c575d609397d2a5dcc14f --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py @@ -0,0 +1,1363 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...configuration_utils import FrozenDict +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-guided image inpainting using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`, `AsymmetricAutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "mask", "masked_image_latents"] + + def __init__( + self, + vae: AutoencoderKL | AsymmetricAutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if scheduler is not None and getattr(scheduler.config, "skip_prk_steps", True) is False: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration" + " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make" + " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to" + " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face" + " Hub, it would be very nice if you could open a Pull request for the" + " `scheduler/scheduler_config.json` file" + ) + deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["skip_prk_steps"] = True + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = ( + unet is not None + and hasattr(unet.config, "_diffusers_version") + and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0") + ) + is_unet_sample_size_less_64 = ( + unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + ) + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- stable-diffusion-v1-5/stable-diffusion-v1-5" + " \n- stable-diffusion-v1-5/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + # Check shapes, assume num_channels_latents == 4, num_channels_mask == 1, num_channels_masked == 4 + if unet is not None and unet.config.in_channels != 9: + logger.info(f"You have loaded a UNet with {unet.config.in_channels} input channels which.") + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + else: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.Tensor = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 1.0, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be inpainted (which parts of the image to + be masked out with `mask_image` and repainted according to `prompt`). For both numpy array and pytorch + tensor, the expected value range is between `[0, 1]` If it's a tensor or a list or tensors, the + expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a list of arrays, the + expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image latents as `image`, but + if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, *optional*): + Pre-encoded latent of the masked image (for inpainting). + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter is modulated by `strength`. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + Examples: + + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionInpaintPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = download_image(img_url).resize((512, 512)) + >>> mask_image = download_image(mask_url).resize((512, 512)) + + >>> pipe = StableDiffusionInpaintPipeline.from_pretrained( + ... "stable-diffusion-v1-5/stable-diffusion-inpainting", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs + self.check_inputs( + prompt, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. set timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps=num_inference_steps, strength=strength, device=device + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9.1 Add image embeds for IP-Adapter + added_cond_kwargs = ( + {"image_embeds": image_embeds} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None + else None + ) + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 10. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if num_channels_unet == 9: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + condition_kwargs = {} + if isinstance(self.vae, AsymmetricAutoencoderKL): + init_image = init_image.to(device=device, dtype=masked_image_latents.dtype) + init_image_condition = init_image.clone() + init_image = self._encode_vae_image(init_image, generator=generator) + mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype) + condition_kwargs = {"image": init_image_condition, "mask": mask_condition} + image = self.vae.decode( + latents / self.vae.config.scaling_factor, return_dict=False, generator=generator, **condition_kwargs + )[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py new file mode 100644 index 0000000000000000000000000000000000000000..7a24e60083510475ab973c3a65d924d0edff1c10 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py @@ -0,0 +1,920 @@ +# Copyright 2025 The InstructPix2Pix Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import PIL_INTERPOLATION, deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput +from .safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess +def preprocess(image): + deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead" + deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + + image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class StableDiffusionInstructPix2PixPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + r""" + Pipeline for pixel-level image editing by following text instructions (based on Stable Diffusion). + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + more details about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "image_latents"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection | None = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 100, + guidance_scale: float = 7.5, + image_guidance_scale: float = 1.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + cross_attention_kwargs: dict[str, Any] | None = None, + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor` `np.ndarray`, `PIL.Image.Image`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be repainted according to `prompt`. Can also accept + image latents as `image`, but if passing latents directly it is not encoded again. + num_inference_steps (`int`, *optional*, defaults to 100): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + image_guidance_scale (`float`, *optional*, defaults to 1.5): + Push the generated image towards the initial `image`. Image guidance scale is enabled by setting + `image_guidance_scale > 1`. Higher image guidance scale encourages generated images that are closely + linked to the source `image`, usually at the expense of lower image quality. This pipeline requires a + value of at least `1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + + Examples: + + ```py + >>> import PIL + >>> import requests + >>> import torch + >>> from io import BytesIO + + >>> from diffusers import StableDiffusionInstructPix2PixPipeline + + + >>> def download_image(url): + ... response = requests.get(url) + ... return PIL.Image.open(BytesIO(response.content)).convert("RGB") + + + >>> img_url = "https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png" + + >>> image = download_image(img_url).resize((512, 512)) + + >>> pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained( + ... "timbrooks/instruct-pix2pix", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "make the mountains snowy" + >>> image = pipe(prompt=prompt, image=image).images[0] + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Check inputs + self.check_inputs( + prompt, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + self._guidance_scale = guidance_scale + self._image_guidance_scale = image_guidance_scale + + device = self._execution_device + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 1. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 2. Encode input prompt + prompt_embeds = self._encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + # 3. Preprocess image + image = self.image_processor.preprocess(image) + + # 4. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare Image latents + image_latents = self.prepare_image_latents( + image, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + self.do_classifier_free_guidance, + ) + + height, width = image_latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 7. Check that shapes of latents and image match the UNet channels + num_channels_image = image_latents.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Expand the latents if we are doing classifier free guidance. + # The latents are expanded 3 times because for pix2pix the guidance\ + # is applied for both the text and the input image. + latent_model_input = torch.cat([latents] * 3) if self.do_classifier_free_guidance else latents + + # concat latents, image_latents in the channel dimension + scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1) + + # predict the noise residual + noise_pred = self.unet( + scaled_latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + added_cond_kwargs=added_cond_kwargs, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_text, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3) + noise_pred = ( + noise_pred_uncond + + self.guidance_scale * (noise_pred_text - noise_pred_image) + + self.image_guidance_scale * (noise_pred_image - noise_pred_uncond) + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + image_latents = callback_outputs.pop("image_latents", image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_ prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + else: + prompt_embeds_dtype = self.unet.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + # pix2pix has two negative embeddings, and unlike in other pipelines latents are ordered [prompt_embeds, negative_prompt_embeds, negative_prompt_embeds] + prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack( + [single_negative_image_embeds] * num_images_per_prompt, dim=0 + ) + + if do_classifier_free_guidance: + single_image_embeds = torch.cat( + [single_image_embeds, single_negative_image_embeds, single_negative_image_embeds] + ) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + else: + repeat_dims = [1] + image_embeds = [] + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + ( + single_image_embeds, + single_negative_image_embeds, + single_negative_image_embeds, + ) = single_image_embeds.chunk(3) + single_image_embeds = single_image_embeds.repeat( + num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:])) + ) + single_negative_image_embeds = single_negative_image_embeds.repeat( + num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:])) + ) + single_image_embeds = torch.cat( + [single_image_embeds, single_negative_image_embeds, single_negative_image_embeds] + ) + else: + single_image_embeds = single_image_embeds.repeat( + num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:])) + ) + image_embeds.append(single_image_embeds) + + return image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_latents( + self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + image_latents = image + else: + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if do_classifier_free_guidance: + uncond_image_latents = torch.zeros_like(image_latents) + image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0) + + return image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def image_guidance_scale(self): + return self._image_guidance_scale + + @property + def num_timesteps(self): + return self._num_timesteps + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self.guidance_scale > 1.0 and self.image_guidance_scale >= 1.0 diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py new file mode 100644 index 0000000000000000000000000000000000000000..1920f033c1264c03794be3d8706f792bf231cde6 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py @@ -0,0 +1,674 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import warnings +from typing import Callable + +import numpy as np +import PIL.Image +import torch +import torch.nn.functional as F +from transformers import CLIPTextModel, CLIPTokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...schedulers import EulerDiscreteScheduler +from ...utils import deprecate, is_torch_xla_available, logging +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.preprocess +def preprocess(image): + warnings.warn( + "The preprocess method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor.preprocess instead", + FutureWarning, + ) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64 + + image = [np.array(i.resize((w, h)))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin): + r""" + Pipeline for upscaling Stable Diffusion output image resolution by a factor of 2. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A [`EulerDiscreteScheduler`] to be used in combination with `unet` to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: EulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic") + + def _encode_prompt( + self, + prompt, + device, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + device=device, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + **kwargs, + ) + + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds]) + + return prompt_embeds, pooled_prompt_embeds + + def encode_prompt( + self, + prompt, + device, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list(int)`): + prompt to be encoded + device: (`torch.device`): + torch device + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + """ + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None or pooled_prompt_embeds is None: + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_length=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + text_encoder_out = self.text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + prompt_embeds = text_encoder_out.hidden_states[-1] + pooled_prompt_embeds = text_encoder_out.pooler_output + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance: + if negative_prompt_embeds is None or negative_pooled_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + max_length = text_input_ids.shape[-1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_length=True, + return_tensors="pt", + ) + + uncond_encoder_out = self.text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + negative_prompt_embeds = uncond_encoder_out.hidden_states[-1] + negative_pooled_prompt_embeds = uncond_encoder_out.pooler_output + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + image, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and not isinstance(prompt, str) and not isinstance(prompt, list): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, np.ndarray) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or `list` but is {type(image)}" + ) + + # verify batch size of prompt and image are same if image is a list or tensor + if isinstance(image, (list, torch.Tensor)): + if prompt is not None: + if isinstance(prompt, str): + batch_size = 1 + else: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if isinstance(image, list): + image_batch_size = len(image) + else: + image_batch_size = image.shape[0] if image.ndim == 4 else 1 + if batch_size != image_batch_size: + raise ValueError( + f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}." + " Please make sure that passed `prompt` matches the batch size of `image`." + ) + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height, width) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 75, + guidance_scale: float = 9.0, + negative_prompt: str | list[str] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`): + The prompt or prompts to guide image upscaling. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be upscaled. If it's a tensor, it can be either a + latent output from a Stable Diffusion model or an image tensor in the range `[-1, 1]`. It is considered + a `latent` if `image.shape[1]` is `4`; otherwise, it is considered to be an image representation and + encoded using this pipeline's `vae` encoder. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, pooled text embeddings are generated from the `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt + weighting). If not provided, pooled `negative_prompt_embeds` are generated from the `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + + Examples: + ```py + >>> from diffusers import StableDiffusionLatentUpscalePipeline, StableDiffusionPipeline + >>> import torch + + >>> pipeline = StableDiffusionPipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16 + ... ) + >>> pipeline.to("cuda") + + >>> model_id = "stabilityai/sd-x2-latent-upscaler" + >>> upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16) + >>> upscaler.to("cuda") + + >>> prompt = "a photo of an astronaut high resolution, unreal engine, ultra realistic" + >>> generator = torch.manual_seed(33) + + >>> low_res_latents = pipeline(prompt, generator=generator, output_type="latent").images + + >>> with torch.no_grad(): + ... image = pipeline.decode_latents(low_res_latents) + >>> image = pipeline.numpy_to_pil(image)[0] + + >>> image.save("../images/a1.png") + + >>> upscaled_image = upscaler( + ... prompt=prompt, + ... image=low_res_latents, + ... num_inference_steps=20, + ... guidance_scale=0, + ... generator=generator, + ... ).images[0] + + >>> upscaled_image.save("../images/a2.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + image, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None: + batch_size = 1 if isinstance(prompt, str) else len(prompt) + else: + batch_size = prompt_embeds.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + if guidance_scale == 0: + prompt = [""] * batch_size + + # 3. Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt, + device, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds]) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + image = image.to(dtype=prompt_embeds.dtype, device=device) + if image.shape[1] == 3: + # encode image if not in latent-space yet + image = retrieve_latents(self.vae.encode(image), generator=generator) * self.vae.config.scaling_factor + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + batch_multiplier = 2 if do_classifier_free_guidance else 1 + image = image[None, :] if image.ndim == 3 else image + image = torch.cat([image] * batch_multiplier) + + # 5. Add noise to image (set to be 0): + # (see below notes from the author): + # "the This step theoretically can make the model work better on out-of-distribution inputs, but mostly just seems to make it match the input less, so it's turned off by default." + noise_level = torch.tensor([0.0], dtype=torch.float32, device=device) + noise_level = torch.cat([noise_level] * image.shape[0]) + inv_noise_level = (noise_level**2 + 1) ** (-0.5) + + image_cond = F.interpolate(image, scale_factor=2, mode="nearest") * inv_noise_level[:, None, None, None] + image_cond = image_cond.to(prompt_embeds.dtype) + + noise_level_embed = torch.cat( + [ + torch.ones(pooled_prompt_embeds.shape[0], 64, dtype=pooled_prompt_embeds.dtype, device=device), + torch.zeros(pooled_prompt_embeds.shape[0], 64, dtype=pooled_prompt_embeds.dtype, device=device), + ], + dim=1, + ) + + timestep_condition = torch.cat([noise_level_embed, pooled_prompt_embeds], dim=1) + + # 6. Prepare latent variables + height, width = image.shape[2:] + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size, + num_channels_latents, + height * 2, # 2x upscale + width * 2, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 7. Check that sizes of image and latents match + num_channels_image = image.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 9. Denoising loop + num_warmup_steps = 0 + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + sigma = self.scheduler.sigmas[i] + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + scaled_model_input = torch.cat([scaled_model_input, image_cond], dim=1) + # preconditioning parameter based on Karras et al. (2022) (table 1) + timestep = torch.log(sigma) * 0.25 + + noise_pred = self.unet( + scaled_model_input, + timestep, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_condition, + ).sample + + # in original repo, the output contains a variance channel that's not used + noise_pred = noise_pred[:, :-1] + + # apply preconditioning, based on table 1 in Karras et al. (2022) + inv_sigma = 1 / (sigma**2 + 1) + noise_pred = inv_sigma * latent_model_input + self.scheduler.scale_model_input(sigma, t) * noise_pred + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py new file mode 100644 index 0000000000000000000000000000000000000000..1a0a7412e5d7ff1095e1d16faa875f7ebc7d2f56 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py @@ -0,0 +1,817 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import warnings +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from . import StableDiffusionPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def preprocess(image): + warnings.warn( + "The preprocess method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor.preprocess instead", + FutureWarning, + ) + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + w, h = image[0].size + w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64 + + image = [np.array(i.resize((w, h)))[None, :] for i in image] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = 2.0 * image - 1.0 + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + image = torch.cat(image, dim=0) + return image + + +class StableDiffusionUpscalePipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionLoraLoaderMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-guided image super-resolution using Stable Diffusion 2. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + low_res_scheduler ([`SchedulerMixin`]): + A scheduler used to add initial noise to the low resolution conditioning image. It must be an instance of + [`DDPMScheduler`]. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + model_cpu_offload_seq = "text_encoder->unet->vae" + _optional_components = ["watermarker", "safety_checker", "feature_extractor"] + _exclude_from_cpu_offload = ["safety_checker"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + low_res_scheduler: DDPMScheduler, + scheduler: KarrasDiffusionSchedulers, + safety_checker: Any | None = None, + feature_extractor: CLIPImageProcessor | None = None, + watermarker: Any | None = None, + max_noise_level: int = 350, + ): + super().__init__() + + if hasattr( + vae, "config" + ): # check if vae has a config attribute `scaling_factor` and if it is set to 0.08333, else set it to 0.08333 and deprecate + is_vae_scaling_factor_set_to_0_08333 = ( + hasattr(vae.config, "scaling_factor") and vae.config.scaling_factor == 0.08333 + ) + if not is_vae_scaling_factor_set_to_0_08333: + deprecation_message = ( + "The configuration file of the vae does not contain `scaling_factor` or it is set to" + f" {vae.config.scaling_factor}, which seems highly unlikely. If your checkpoint is a fine-tuned" + " version of `stabilityai/stable-diffusion-x4-upscaler` you should change 'scaling_factor' to" + " 0.08333 Please make sure to update the config accordingly, as not doing so might lead to" + " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging" + " Face Hub, it would be very nice if you could open a Pull Request for the `vae/config.json` file" + ) + deprecate("wrong scaling_factor", "1.0.0", deprecation_message, standard_warn=False) + vae.register_to_config(scaling_factor=0.08333) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + low_res_scheduler=low_res_scheduler, + scheduler=scheduler, + safety_checker=safety_checker, + watermarker=watermarker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic") + self.register_to_config(max_noise_level=max_noise_level) + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, nsfw_detected, watermark_detected = self.safety_checker( + images=image, + clip_input=safety_checker_input.pixel_values.to(dtype=dtype), + ) + else: + nsfw_detected = None + watermark_detected = None + + if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None: + self.unet_offload_hook.offload() + + return image, nsfw_detected, watermark_detected + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs( + self, + prompt, + image, + noise_level, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, np.ndarray) + and not isinstance(image, list) + ): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}" + ) + + # verify batch size of prompt and image are same if image is a list or tensor or numpy array + if isinstance(image, (list, np.ndarray, torch.Tensor)): + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if isinstance(image, list): + image_batch_size = len(image) + else: + image_batch_size = image.shape[0] + if batch_size != image_batch_size: + raise ValueError( + f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}." + " Please make sure that passed `prompt` matches the batch size of `image`." + ) + + # check noise level + if noise_level > self.config.max_noise_level: + raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height, width) + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + @torch.no_grad() + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + num_inference_steps: int = 75, + guidance_scale: float = 9.0, + noise_level: int = 20, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image` or tensor representing an image batch to be upscaled. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + noise_level (`int`, *optional*, defaults to 20): + The amount of noise to add to the upscaled input image. Must be in the range `[0, max_noise_level]` + where `max_noise_level` is defined by the scheduler. A higher `noise_level` adds more noise to the + input, increasing variation but reducing fidelity to the source image. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + ```py + >>> import requests + >>> from PIL import Image + >>> from io import BytesIO + >>> from diffusers import StableDiffusionUpscalePipeline + >>> import torch + + >>> # load model and scheduler + >>> model_id = "stabilityai/stable-diffusion-x4-upscaler" + >>> pipeline = StableDiffusionUpscalePipeline.from_pretrained( + ... model_id, variant="fp16", torch_dtype=torch.float16 + ... ) + >>> pipeline = pipeline.to("cuda") + + >>> # let's download an image + >>> url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale/low_res_cat.png" + >>> response = requests.get(url) + >>> low_res_img = Image.open(BytesIO(response.content)).convert("RGB") + >>> low_res_img = low_res_img.resize((128, 128)) + >>> prompt = "a white cat" + + >>> upscaled_image = pipeline(prompt=prompt, image=low_res_img).images[0] + >>> upscaled_image.save("upsampled_cat.png") + ``` + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + # 1. Check inputs + self.check_inputs( + prompt, + image, + noise_level, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + ) + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + image = image.to(dtype=prompt_embeds.dtype, device=device) + + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Add noise to image + noise_level = torch.tensor([noise_level], dtype=torch.long, device=device) + noise = randn_tensor(image.shape, generator=generator, device=device, dtype=prompt_embeds.dtype) + image = self.low_res_scheduler.add_noise(image, noise, noise_level) + + batch_multiplier = 2 if do_classifier_free_guidance else 1 + image = torch.cat([image] * batch_multiplier * num_images_per_prompt) + noise_level = torch.cat([noise_level] * image.shape[0]) + + # 6. Prepare latent variables + height, width = image.shape[2:] + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 7. Check that sizes of image and latents match + num_channels_image = image.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + latent_model_input = torch.cat([latent_model_input, image], dim=1) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=noise_level, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + + # Ensure latents are always the same type as the VAE + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + image, has_nsfw_concept, _ = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # 11. Apply watermark + if output_type == "pil" and self.watermarker is not None: + image = self.watermarker.apply_watermark(image) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py new file mode 100644 index 0000000000000000000000000000000000000000..d2c9bf0c41620ce4fa6a330842572ba4396fc071 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip.py @@ -0,0 +1,952 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer +from transformers.models.clip.modeling_clip import CLIPTextModelOutput + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, PriorTransformer, UNet2DConditionModel +from ...models.embeddings import get_timestep_embedding +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin +from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableUnCLIPPipeline + + >>> pipe = StableUnCLIPPipeline.from_pretrained( + ... "fusing/stable-unclip-2-1-l", torch_dtype=torch.float16 + ... ) # TODO update model path + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> images = pipe(prompt).images + >>> images[0].save("astronaut_horse.png") + ``` +""" + + +class StableUnCLIPPipeline( + DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin +): + """ + Pipeline for text-to-image generation using stable unCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + prior_tokenizer ([`CLIPTokenizer`]): + A [`CLIPTokenizer`]. + prior_text_encoder ([`CLIPTextModelWithProjection`]): + Frozen [`CLIPTextModelWithProjection`] text-encoder. + prior ([`PriorTransformer`]): + The canonical unCLIP prior to approximate the image embedding from the text embedding. + prior_scheduler ([`KarrasDiffusionSchedulers`]): + Scheduler used in the prior denoising process. + image_normalizer ([`StableUnCLIPImageNormalizer`]): + Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image + embeddings after the noise has been applied. + image_noising_scheduler ([`KarrasDiffusionSchedulers`]): + Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined + by the `noise_level`. + tokenizer ([`CLIPTokenizer`]): + A [`CLIPTokenizer`]. + text_encoder ([`CLIPTextModel`]): + Frozen [`CLIPTextModel`] text-encoder. + unet ([`UNet2DConditionModel`]): + A [`UNet2DConditionModel`] to denoise the encoded image latents. + scheduler ([`KarrasDiffusionSchedulers`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + """ + + _exclude_from_cpu_offload = ["prior", "image_normalizer"] + model_cpu_offload_seq = "text_encoder->prior_text_encoder->unet->vae" + + # prior components + prior_tokenizer: CLIPTokenizer + prior_text_encoder: CLIPTextModelWithProjection + prior: PriorTransformer + prior_scheduler: KarrasDiffusionSchedulers + + # image noising components + image_normalizer: StableUnCLIPImageNormalizer + image_noising_scheduler: KarrasDiffusionSchedulers + + # regular denoising components + tokenizer: CLIPTokenizer + text_encoder: CLIPTextModel + unet: UNet2DConditionModel + scheduler: KarrasDiffusionSchedulers + + vae: AutoencoderKL + + def __init__( + self, + # prior components + prior_tokenizer: CLIPTokenizer, + prior_text_encoder: CLIPTextModelWithProjection, + prior: PriorTransformer, + prior_scheduler: KarrasDiffusionSchedulers, + # image noising components + image_normalizer: StableUnCLIPImageNormalizer, + image_noising_scheduler: KarrasDiffusionSchedulers, + # regular denoising components + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + # vae + vae: AutoencoderKL, + ): + super().__init__() + + self.register_modules( + prior_tokenizer=prior_tokenizer, + prior_text_encoder=prior_text_encoder, + prior=prior, + prior_scheduler=prior_scheduler, + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + tokenizer=tokenizer, + text_encoder=text_encoder, + unet=unet, + scheduler=scheduler, + vae=vae, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline._encode_prompt with _encode_prompt->_encode_prior_prompt, tokenizer->prior_tokenizer, text_encoder->prior_text_encoder + def _encode_prior_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + text_model_output: CLIPTextModelOutput | tuple | None = None, + text_attention_mask: torch.Tensor | None = None, + ): + if text_model_output is None: + batch_size = len(prompt) if isinstance(prompt, list) else 1 + # get prompt text embeddings + text_inputs = self.prior_tokenizer( + prompt, + padding="max_length", + max_length=self.prior_tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + text_mask = text_inputs.attention_mask.bool().to(device) + + untruncated_ids = self.prior_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.prior_tokenizer.batch_decode( + untruncated_ids[:, self.prior_tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.prior_tokenizer.model_max_length} tokens: {removed_text}" + ) + text_input_ids = text_input_ids[:, : self.prior_tokenizer.model_max_length] + + prior_text_encoder_output = self.prior_text_encoder(text_input_ids.to(device)) + + prompt_embeds = prior_text_encoder_output.text_embeds + text_enc_hid_states = prior_text_encoder_output.last_hidden_state + + else: + batch_size = text_model_output[0].shape[0] + prompt_embeds, text_enc_hid_states = text_model_output[0], text_model_output[1] + text_mask = text_attention_mask + + prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0) + text_enc_hid_states = text_enc_hid_states.repeat_interleave(num_images_per_prompt, dim=0) + text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + uncond_tokens = [""] * batch_size + + uncond_input = self.prior_tokenizer( + uncond_tokens, + padding="max_length", + max_length=self.prior_tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + uncond_text_mask = uncond_input.attention_mask.bool().to(device) + negative_prompt_embeds_prior_text_encoder_output = self.prior_text_encoder( + uncond_input.input_ids.to(device) + ) + + negative_prompt_embeds = negative_prompt_embeds_prior_text_encoder_output.text_embeds + uncond_text_enc_hid_states = negative_prompt_embeds_prior_text_encoder_output.last_hidden_state + + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len) + + seq_len = uncond_text_enc_hid_states.shape[1] + uncond_text_enc_hid_states = uncond_text_enc_hid_states.repeat(1, num_images_per_prompt, 1) + uncond_text_enc_hid_states = uncond_text_enc_hid_states.view( + batch_size * num_images_per_prompt, seq_len, -1 + ) + uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0) + + # done duplicates + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + text_enc_hid_states = torch.cat([uncond_text_enc_hid_states, text_enc_hid_states]) + + text_mask = torch.cat([uncond_text_mask, text_mask]) + + return prompt_embeds, text_enc_hid_states, text_mask + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs with prepare_extra_step_kwargs->prepare_prior_extra_step_kwargs, scheduler->prior_scheduler + def prepare_prior_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the prior_scheduler step, since not all prior_schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other prior_schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.prior_scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the prior_scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.prior_scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + noise_level, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two." + ) + + if prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined." + ) + + if prompt is not None and negative_prompt is not None: + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps: + raise ValueError( + f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive." + ) + + # Copied from diffusers.pipelines.deprecated.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents + def prepare_latents(self, shape, dtype, device, generator, latents, scheduler): + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + + latents = latents * scheduler.init_noise_sigma + return latents + + def noise_image_embeddings( + self, + image_embeds: torch.Tensor, + noise_level: int, + noise: torch.Tensor | None = None, + generator: torch.Generator | None = None, + ): + """ + Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher + `noise_level` increases the variance in the final un-noised images. + + The noise is applied in two ways: + 1. A noise schedule is applied directly to the embeddings. + 2. A vector of sinusoidal time embeddings are appended to the output. + + In both cases, the amount of noise is controlled by the same `noise_level`. + + The embeddings are normalized before the noise is applied and un-normalized after the noise is applied. + """ + if noise is None: + noise = randn_tensor( + image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype + ) + + noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device) + + self.image_normalizer.to(image_embeds.device) + image_embeds = self.image_normalizer.scale(image_embeds) + + image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise) + + image_embeds = self.image_normalizer.unscale(image_embeds) + + noise_level = get_timestep_embedding( + timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0 + ) + + # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors, + # but we might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + noise_level = noise_level.to(image_embeds.dtype) + + image_embeds = torch.cat((image_embeds, noise_level), 1) + + return image_embeds + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + # regular denoising process args + prompt: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 20, + guidance_scale: float = 10.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + noise_level: int = 0, + # prior args + prior_num_inference_steps: int = 25, + prior_guidance_scale: float = 4.0, + prior_latents: torch.Tensor | None = None, + clip_skip: int | None = None, + ): + """ + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 10.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + noise_level (`int`, *optional*, defaults to `0`): + The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in + the final un-noised images. See [`StableUnCLIPPipeline.noise_image_embeddings`] for more details. + prior_num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps in the prior denoising process. More denoising steps usually lead to a + higher quality image at the expense of slower inference. + prior_guidance_scale (`float`, *optional*, defaults to 4.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + prior_latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + embedding generation in the prior denoising process. Can be used to tweak the same generation with + different prompts. If not provided, a latents tensor is generated by sampling using the supplied random + `generator`. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning + a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + height=height, + width=width, + callback_steps=callback_steps, + noise_level=noise_level, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + batch_size = batch_size * num_images_per_prompt + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + prior_do_classifier_free_guidance = prior_guidance_scale > 1.0 + + # 3. Encode input prompt + prior_prompt_embeds, prior_text_encoder_hidden_states, prior_text_mask = self._encode_prior_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=prior_do_classifier_free_guidance, + ) + + # 4. Prepare prior timesteps + self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device) + prior_timesteps_tensor = self.prior_scheduler.timesteps + + # 5. Prepare prior latent variables + embedding_dim = self.prior.config.embedding_dim + prior_latents = self.prepare_latents( + (batch_size, embedding_dim), + prior_prompt_embeds.dtype, + device, + generator, + prior_latents, + self.prior_scheduler, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + prior_extra_step_kwargs = self.prepare_prior_extra_step_kwargs(generator, eta) + + # 7. Prior denoising loop + for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([prior_latents] * 2) if prior_do_classifier_free_guidance else prior_latents + latent_model_input = self.prior_scheduler.scale_model_input(latent_model_input, t) + + predicted_image_embedding = self.prior( + latent_model_input, + timestep=t, + proj_embedding=prior_prompt_embeds, + encoder_hidden_states=prior_text_encoder_hidden_states, + attention_mask=prior_text_mask, + ).predicted_image_embedding + + if prior_do_classifier_free_guidance: + predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2) + predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * ( + predicted_image_embedding_text - predicted_image_embedding_uncond + ) + + prior_latents = self.prior_scheduler.step( + predicted_image_embedding, + timestep=t, + sample=prior_latents, + **prior_extra_step_kwargs, + return_dict=False, + )[0] + + if callback is not None and i % callback_steps == 0: + callback(i, t, prior_latents) + + prior_latents = self.prior.post_process_latents(prior_latents) + + image_embeds = prior_latents + + # done prior + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 8. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 9. Prepare image embeddings + image_embeds = self.noise_image_embeddings( + image_embeds=image_embeds, + noise_level=noise_level, + generator=generator, + ) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(image_embeds) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeds = torch.cat([negative_prompt_embeds, image_embeds]) + + # 10. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 11. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + latents = self.prepare_latents( + shape=shape, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + scheduler=self.scheduler, + ) + + # 12. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 13. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + class_labels=image_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py b/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..059ae1e6fd4d455e052a6a2cbfa9f008deab8875 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py @@ -0,0 +1,858 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from ...image_processor import VaeImageProcessor +from ...loaders import StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.embeddings import get_timestep_embedding +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin +from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import requests + >>> import torch + >>> from PIL import Image + >>> from io import BytesIO + + >>> from diffusers import StableUnCLIPImg2ImgPipeline + + >>> pipe = StableUnCLIPImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-2-1-unclip-small", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + + >>> response = requests.get(url) + >>> init_image = Image.open(BytesIO(response.content)).convert("RGB") + >>> init_image = init_image.resize((768, 512)) + + >>> prompt = "A fantasy landscape, trending on artstation" + + >>> images = pipe(init_image, prompt).images + >>> images[0].save("fantasy_landscape.png") + ``` +""" + + +class StableUnCLIPImg2ImgPipeline( + DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin +): + """ + Pipeline for text-guided image-to-image generation using stable unCLIP. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + feature_extractor ([`CLIPImageProcessor`]): + Feature extractor for image pre-processing before being encoded. + image_encoder ([`CLIPVisionModelWithProjection`]): + CLIP vision model for encoding images. + image_normalizer ([`StableUnCLIPImageNormalizer`]): + Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image + embeddings after the noise has been applied. + image_noising_scheduler ([`KarrasDiffusionSchedulers`]): + Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined + by the `noise_level`. + tokenizer (`~transformers.CLIPTokenizer`): + A [`~transformers.CLIPTokenizer`)]. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen [`~transformers.CLIPTextModel`] text-encoder. + unet ([`UNet2DConditionModel`]): + A [`UNet2DConditionModel`] to denoise the encoded image latents. + scheduler ([`KarrasDiffusionSchedulers`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _exclude_from_cpu_offload = ["image_normalizer"] + + # image encoding components + feature_extractor: CLIPImageProcessor + image_encoder: CLIPVisionModelWithProjection + + # image noising components + image_normalizer: StableUnCLIPImageNormalizer + image_noising_scheduler: KarrasDiffusionSchedulers + + # regular denoising components + tokenizer: CLIPTokenizer + text_encoder: CLIPTextModel + unet: UNet2DConditionModel + scheduler: KarrasDiffusionSchedulers + + vae: AutoencoderKL + + def __init__( + self, + # image encoding components + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection, + # image noising components + image_normalizer: StableUnCLIPImageNormalizer, + image_noising_scheduler: KarrasDiffusionSchedulers, + # regular denoising components + tokenizer: CLIPTokenizer, + text_encoder: CLIPTextModel, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + # vae + vae: AutoencoderKL, + ): + super().__init__() + + self.register_modules( + feature_extractor=feature_extractor, + image_encoder=image_encoder, + image_normalizer=image_normalizer, + image_noising_scheduler=image_noising_scheduler, + tokenizer=tokenizer, + text_encoder=text_encoder, + unet=unet, + scheduler=scheduler, + vae=vae, + ) + + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def _encode_image( + self, + image, + device, + batch_size, + num_images_per_prompt, + do_classifier_free_guidance, + noise_level, + generator, + image_embeds, + ): + dtype = next(self.image_encoder.parameters()).dtype + + if isinstance(image, PIL.Image.Image): + # the image embedding should repeated so it matches the total batch size of the prompt + repeat_by = batch_size + else: + # assume the image input is already properly batched and just needs to be repeated so + # it matches the num_images_per_prompt. + # + # NOTE(will) this is probably missing a few number of side cases. I.e. batched/non-batched + # `image_embeds`. If those happen to be common use cases, let's think harder about + # what the expected dimensions of inputs should be and how we handle the encoding. + repeat_by = num_images_per_prompt + + if image_embeds is None: + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(images=image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeds = self.image_encoder(image).image_embeds + + image_embeds = self.noise_image_embeddings( + image_embeds=image_embeds, + noise_level=noise_level, + generator=generator, + ) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + image_embeds = image_embeds.unsqueeze(1) + bs_embed, seq_len, _ = image_embeds.shape + image_embeds = image_embeds.repeat(1, repeat_by, 1) + image_embeds = image_embeds.view(bs_embed * repeat_by, seq_len, -1) + image_embeds = image_embeds.squeeze(1) + + if do_classifier_free_guidance: + negative_prompt_embeds = torch.zeros_like(image_embeds) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeds = torch.cat([negative_prompt_embeds, image_embeds]) + + return image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + image, + height, + width, + callback_steps, + noise_level, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two." + ) + + if prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + + if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined." + ) + + if prompt is not None and negative_prompt is not None: + if type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps: + raise ValueError( + f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive." + ) + + if image is not None and image_embeds is not None: + raise ValueError( + "Provide either `image` or `image_embeds`. Please make sure to define only one of the two." + ) + + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `image_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + + if image is not None: + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `list[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_unclip.StableUnCLIPPipeline.noise_image_embeddings + def noise_image_embeddings( + self, + image_embeds: torch.Tensor, + noise_level: int, + noise: torch.Tensor | None = None, + generator: torch.Generator | None = None, + ): + """ + Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher + `noise_level` increases the variance in the final un-noised images. + + The noise is applied in two ways: + 1. A noise schedule is applied directly to the embeddings. + 2. A vector of sinusoidal time embeddings are appended to the output. + + In both cases, the amount of noise is controlled by the same `noise_level`. + + The embeddings are normalized before the noise is applied and un-normalized after the noise is applied. + """ + if noise is None: + noise = randn_tensor( + image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype + ) + + noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device) + + self.image_normalizer.to(image_embeds.device) + image_embeds = self.image_normalizer.scale(image_embeds) + + image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise) + + image_embeds = self.image_normalizer.unscale(image_embeds) + + noise_level = get_timestep_embedding( + timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0 + ) + + # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors, + # but we might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + noise_level = noise_level.to(image_embeds.dtype) + + image_embeds = torch.cat((image_embeds, noise_level), 1) + + return image_embeds + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: torch.Tensor | PIL.Image.Image = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 20, + guidance_scale: float = 10, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + noise_level: int = 0, + image_embeds: torch.Tensor | None = None, + clip_skip: int | None = None, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, either `prompt_embeds` will be + used or prompt is initialized to `""`. + image (`torch.Tensor` or `PIL.Image.Image`): + `Image` or tensor representing an image batch. The image is encoded to its CLIP embedding which the + `unet` is conditioned on. The image is _not_ encoded by the `vae` and then used as the latents in the + denoising process like it is in the standard Stable Diffusion text-guided image variation process. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 20): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 10.0): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://huggingface.co/papers/2010.02502) paper. Only + applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple. + callback (`Callable`, *optional*): + A function that calls every `callback_steps` steps during inference. The function is called with the + following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function is called. If not specified, the callback is called at + every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + noise_level (`int`, *optional*, defaults to `0`): + The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in + the final un-noised images. See [`StableUnCLIPPipeline.noise_image_embeddings`] for more details. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated CLIP embeddings to condition the `unet` on. These latents are not used in the denoising + process. If you want to provide pre-generated latents, pass them to `__call__` as `latents`. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.ImagePipelineOutput`] or `tuple`: + [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning + a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + if prompt is None and prompt_embeds is None: + prompt = len(image) * [""] if isinstance(image, list) else "" + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt=prompt, + image=image, + height=height, + width=width, + callback_steps=callback_steps, + noise_level=noise_level, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + image_embeds=image_embeds, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + batch_size = batch_size * num_images_per_prompt + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Encoder input image + noise_level = torch.tensor([noise_level], device=device) + image_embeds = self._encode_image( + image=image, + device=device, + batch_size=batch_size, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + noise_level=noise_level, + generator=generator, + image_embeds=image_embeds, + ) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 6. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + if latents is None: + latents = self.prepare_latents( + batch_size=batch_size, + num_channels_latents=num_channels_latents, + height=height, + width=width, + dtype=prompt_embeds.dtype, + device=device, + generator=generator, + latents=latents, + ) + + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 8. Denoising loop + for i, t in enumerate(self.progress_bar(timesteps)): + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + class_labels=image_embeds, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + # 9. Post-processing + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion/safety_checker.py b/diffusers/pipelines/stable_diffusion/safety_checker.py new file mode 100644 index 0000000000000000000000000000000000000000..c9fa98f13530ec7916f9b713f3df102bee3bac3f --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/safety_checker.py @@ -0,0 +1,129 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numpy as np +import torch +import torch.nn as nn +from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel + +from ...utils import is_transformers_version, logging + + +logger = logging.get_logger(__name__) + + +def cosine_distance(image_embeds, text_embeds): + normalized_image_embeds = nn.functional.normalize(image_embeds) + normalized_text_embeds = nn.functional.normalize(text_embeds) + return torch.mm(normalized_image_embeds, normalized_text_embeds.t()) + + +class StableDiffusionSafetyChecker(PreTrainedModel): + config_class = CLIPConfig + main_input_name = "clip_input" + + _no_split_modules = ["CLIPEncoderLayer"] + + def __init__(self, config: CLIPConfig): + super().__init__(config) + + self.vision_model = CLIPVisionModel(config.vision_config) + self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False) + + self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False) + self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False) + + self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False) + self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False) + # Model requires post_init after transformers v4.57.3 + if is_transformers_version(">", "4.57.3"): + self.post_init() + + @torch.no_grad() + def forward(self, clip_input, images): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy() + cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy() + + result = [] + batch_size = image_embeds.shape[0] + for i in range(batch_size): + result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []} + + # increase this value to create a stronger `nfsw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + for concept_idx in range(len(special_cos_dist[0])): + concept_cos = special_cos_dist[i][concept_idx] + concept_threshold = self.special_care_embeds_weights[concept_idx].item() + result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["special_scores"][concept_idx] > 0: + result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]}) + adjustment = 0.01 + + for concept_idx in range(len(cos_dist[0])): + concept_cos = cos_dist[i][concept_idx] + concept_threshold = self.concept_embeds_weights[concept_idx].item() + result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3) + if result_img["concept_scores"][concept_idx] > 0: + result_img["bad_concepts"].append(concept_idx) + + result.append(result_img) + + has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result] + + for idx, has_nsfw_concept in enumerate(has_nsfw_concepts): + if has_nsfw_concept: + if torch.is_tensor(images) or torch.is_tensor(images[0]): + images[idx] = torch.zeros_like(images[idx]) # black image + else: + images[idx] = np.zeros(images[idx].shape) # black image + + if any(has_nsfw_concepts): + logger.warning( + "Potential NSFW content was detected in one or more images. A black image will be returned instead." + " Try again with a different prompt and/or seed." + ) + + return images, has_nsfw_concepts + + @torch.no_grad() + def forward_onnx(self, clip_input: torch.Tensor, images: torch.Tensor): + pooled_output = self.vision_model(clip_input)[1] # pooled_output + image_embeds = self.visual_projection(pooled_output) + + special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds) + cos_dist = cosine_distance(image_embeds, self.concept_embeds) + + # increase this value to create a stronger `nsfw` filter + # at the cost of increasing the possibility of filtering benign images + adjustment = 0.0 + + special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment + # special_scores = special_scores.round(decimals=3) + special_care = torch.any(special_scores > 0, dim=1) + special_adjustment = special_care * 0.01 + special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1]) + + concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment + # concept_scores = concept_scores.round(decimals=3) + has_nsfw_concepts = torch.any(concept_scores > 0, dim=1) + + images[has_nsfw_concepts] = 0.0 # black image + + return images, has_nsfw_concepts diff --git a/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py b/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py new file mode 100644 index 0000000000000000000000000000000000000000..1a29bc9bb63b8d938f6cf09568e4a1519f02709f --- /dev/null +++ b/diffusers/pipelines/stable_diffusion/stable_unclip_image_normalizer.py @@ -0,0 +1,55 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +from torch import nn + +from ...configuration_utils import ConfigMixin, register_to_config +from ...models.modeling_utils import ModelMixin + + +class StableUnCLIPImageNormalizer(ModelMixin, ConfigMixin): + """ + This class is used to hold the mean and standard deviation of the CLIP embedder used in stable unCLIP. + + It is used to normalize the image embeddings before the noise is applied and un-normalize the noised image + embeddings. + """ + + @register_to_config + def __init__( + self, + embedding_dim: int = 768, + ): + super().__init__() + + self.mean = nn.Parameter(torch.zeros(1, embedding_dim)) + self.std = nn.Parameter(torch.ones(1, embedding_dim)) + + def to( + self, + torch_device: str | torch.device | None = None, + torch_dtype: torch.dtype | None = None, + ): + self.mean = nn.Parameter(self.mean.to(torch_device).to(torch_dtype)) + self.std = nn.Parameter(self.std.to(torch_device).to(torch_dtype)) + return self + + def scale(self, embeds): + embeds = (embeds - self.mean) * 1.0 / self.std + return embeds + + def unscale(self, embeds): + embeds = (embeds * self.std) + self.mean + return embeds diff --git a/diffusers/pipelines/stable_diffusion_3/__init__.py b/diffusers/pipelines/stable_diffusion_3/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2907dc5f26eee784edc0b209d5666b32502a141a --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_3/__init__.py @@ -0,0 +1,53 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["StableDiffusion3PipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_3"] = ["StableDiffusion3Pipeline"] + _import_structure["pipeline_stable_diffusion_3_img2img"] = ["StableDiffusion3Img2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_3_inpaint"] = ["StableDiffusion3InpaintPipeline"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_diffusion_3 import StableDiffusion3Pipeline + from .pipeline_stable_diffusion_3_img2img import StableDiffusion3Img2ImgPipeline + from .pipeline_stable_diffusion_3_inpaint import StableDiffusion3InpaintPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/stable_diffusion_3/pipeline_output.py b/diffusers/pipelines/stable_diffusion_3/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..43f801d81e2e00ae6563eb4f6aa8d2e614d813bc --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_3/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class StableDiffusion3PipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py b/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py new file mode 100644 index 0000000000000000000000000000000000000000..5c05b469660fb5276e2fbf6150b1d2f6c397fbb2 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3.py @@ -0,0 +1,1147 @@ +# Copyright 2025 Stability AI, The HuggingFace Team and The InstantX Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3Pipeline + + >>> pipe = StableDiffusion3Pipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "A cat holding a sign that says hello world" + >>> image = pipe(prompt).images[0] + >>> image.save("sd3.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3Pipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + image_encoder (`SiglipVisionModel`, *optional*): + Pre-trained Vision Model for IP Adapter. + feature_extractor (`SiglipImageProcessor`, *optional*): + Image processor for IP Adapter. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + image_encoder: SiglipVisionModel = None, + feature_extractor: SiglipImageProcessor = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + max_sequence_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + clip_skip: int | None = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str], + prompt_3: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + clip_skip: int | None = None, + max_sequence_length: int = 256, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + if latents is not None: + return latents.to(device=device, dtype=dtype) + + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def skip_guidance_layers(self): + return self._skip_guidance_layers + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_image + def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: + """Encodes the given image into a feature representation using a pre-trained image encoder. + + Args: + image (`PipelineImageInput`): + Input image to be encoded. + device: (`torch.device`): + Torch device. + + Returns: + `torch.Tensor`: The encoded image feature representation. + """ + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=self.dtype) + + return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + + # Adapted from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: torch.Tensor | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + ) -> torch.Tensor: + """Prepares image embeddings for use in the IP-Adapter. + + Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. + + Args: + ip_adapter_image (`PipelineImageInput`, *optional*): + The input image to extract features from for IP-Adapter. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Precomputed image embeddings. + device: (`torch.device`, *optional*): + Torch device. + num_images_per_prompt (`int`, defaults to 1): + Number of images that should be generated per prompt. + do_classifier_free_guidance (`bool`, defaults to True): + Whether to use classifier free guidance or not. + """ + device = device or self._execution_device + + if ip_adapter_image_embeds is not None: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) + else: + single_image_embeds = ip_adapter_image_embeds + elif ip_adapter_image is not None: + single_image_embeds = self.encode_image(ip_adapter_image, device) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.zeros_like(single_image_embeds) + else: + raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") + + image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + + return image_embeds.to(device=device) + + def enable_sequential_cpu_offload(self, *args, **kwargs): + if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: + logger.warning( + "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " + "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " + "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." + ) + + super().enable_sequential_cpu_offload(*args, **kwargs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 28, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + skip_guidance_layers: list[int] = None, + skip_layer_guidance_scale: float = 2.8, + skip_layer_guidance_stop: float = 0.2, + skip_layer_guidance_start: float = 0.01, + mu: float | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, + emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to + `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of + a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + skip_guidance_layers (`list[int]`, *optional*): + A list of integers that specify layers to skip during guidance. If not provided, all layers will be + used for guidance. If provided, the guidance will only be applied to the layers specified in the list. + Recommended value by StabiltyAI for Stable Diffusion 3.5 Medium is [7, 8, 9]. + skip_layer_guidance_scale (`int`, *optional*): The scale of the guidance for the layers specified in + `skip_guidance_layers`. The guidance will be applied to the layers specified in `skip_guidance_layers` + with a scale of `skip_layer_guidance_scale`. The guidance will be applied to the rest of the layers + with a scale of `1`. + skip_layer_guidance_stop (`int`, *optional*): The step at which the guidance for the layers specified in + `skip_guidance_layers` will stop. The guidance will be applied to the layers specified in + `skip_guidance_layers` until the fraction specified in `skip_layer_guidance_stop`. Recommended value by + StabiltyAI for Stable Diffusion 3.5 Medium is 0.2. + skip_layer_guidance_start (`int`, *optional*): The step at which the guidance for the layers specified in + `skip_guidance_layers` will start. The guidance will be applied to the layers specified in + `skip_guidance_layers` from the fraction specified in `skip_layer_guidance_start`. Recommended value by + StabiltyAI for Stable Diffusion 3.5 Medium is 0.01. + mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._skip_layer_guidance_scale = skip_layer_guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_classifier_free_guidance: + if skip_guidance_layers is not None: + original_prompt_embeds = prompt_embeds + original_pooled_prompt_embeds = pooled_prompt_embeds + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 5. Prepare timesteps + scheduler_kwargs = {} + if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: + _, _, height, width = latents.shape + image_seq_len = (height // self.transformer.config.patch_size) * ( + width // self.transformer.config.patch_size + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + scheduler_kwargs["mu"] = mu + elif mu is not None: + scheduler_kwargs["mu"] = mu + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + timestep_device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Prepare image embeddings + if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} + else: + self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) + + # 7. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + should_skip_layers = ( + True + if i > num_inference_steps * skip_layer_guidance_start + and i < num_inference_steps * skip_layer_guidance_stop + else False + ) + if skip_guidance_layers is not None and should_skip_layers: + timestep = t.expand(latents.shape[0]) + latent_model_input = latents + noise_pred_skip_layers = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=original_prompt_embeds, + pooled_projections=original_pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + skip_layers=skip_guidance_layers, + )[0] + noise_pred = ( + noise_pred + (noise_pred_text - noise_pred_skip_layers) * self._skip_layer_guidance_scale + ) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + pooled_prompt_embeds = callback_outputs.pop("pooled_prompt_embeds", pooled_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py b/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..c0ab805a4ef4aeaca5ce2b0f6776722c2028ebc7 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_img2img.py @@ -0,0 +1,1173 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + + >>> from diffusers import AutoPipelineForImage2Image + >>> from diffusers.utils import load_image + + >>> device = "cuda" + >>> model_id_or_path = "stabilityai/stable-diffusion-3-medium-diffusers" + >>> pipe = AutoPipelineForImage2Image.from_pretrained(model_id_or_path, torch_dtype=torch.float16) + >>> pipe = pipe.to(device) + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + + >>> prompt = "cat wizard, gandalf, lord of the rings, detailed, fantasy, cute, adorable, Pixar, Disney, 8k" + + >>> images = pipe(prompt=prompt, image=init_image, strength=0.95, guidance_scale=7.5).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3Img2ImgPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + image_encoder (`SiglipVisionModel`, *optional*): + Pre-trained Vision Model for IP Adapter. + feature_extractor (`SiglipImageProcessor`, *optional*): + Image processor for IP Adapter. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + image_encoder: SiglipVisionModel | None = None, + feature_extractor: SiglipImageProcessor | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, vae_latent_channels=latent_channels + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + max_sequence_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + clip_skip: int | None = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str], + prompt_3: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + clip_skip: int | None = None, + max_sequence_length: int = 256, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + if image.shape[1] == self.vae.config.latent_channels: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + init_latents = (init_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler.scale_noise(init_latents, timestep, noise) + latents = init_latents.to(device=device, dtype=dtype) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_image + def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: + """Encodes the given image into a feature representation using a pre-trained image encoder. + + Args: + image (`PipelineImageInput`): + Input image to be encoded. + device: (`torch.device`): + Torch device. + + Returns: + `torch.Tensor`: The encoded image feature representation. + """ + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=self.dtype) + + return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: torch.Tensor | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + ) -> torch.Tensor: + """Prepares image embeddings for use in the IP-Adapter. + + Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. + + Args: + ip_adapter_image (`PipelineImageInput`, *optional*): + The input image to extract features from for IP-Adapter. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Precomputed image embeddings. + device: (`torch.device`, *optional*): + Torch device. + num_images_per_prompt (`int`, defaults to 1): + Number of images that should be generated per prompt. + do_classifier_free_guidance (`bool`, defaults to True): + Whether to use classifier free guidance or not. + """ + device = device or self._execution_device + + if ip_adapter_image_embeds is not None: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) + else: + single_image_embeds = ip_adapter_image_embeds + elif ip_adapter_image is not None: + single_image_embeds = self.encode_image(ip_adapter_image, device) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.zeros_like(single_image_embeds) + else: + raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") + + image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + + return image_embeds.to(device=device) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.enable_sequential_cpu_offload + def enable_sequential_cpu_offload(self, *args, **kwargs): + if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: + logger.warning( + "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " + "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " + "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." + ) + + super().enable_sequential_cpu_offload(*args, **kwargs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + image: PipelineImageInput = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: torch.Tensor | None = None, + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + mu: float | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a + list of tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or + a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + strength (`float`, *optional*, defaults to 0.6): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, + emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to + `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of + a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Preprocess image + image = self.image_processor.preprocess(image, height=height, width=width) + + # 4. Prepare timesteps + scheduler_kwargs = {} + if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: + image_seq_len = (int(height) // self.vae_scale_factor // self.transformer.config.patch_size) * ( + int(width) // self.vae_scale_factor // self.transformer.config.patch_size + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + scheduler_kwargs["mu"] = mu + elif mu is not None: + scheduler_kwargs["mu"] = mu + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, sigmas=sigmas, **scheduler_kwargs + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + ) + + # 6. Prepare image embeddings + if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} + else: + self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) + + # 7. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py b/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..321e9f8dd80ecde01041048ec4dd4b22776b7cca --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_3/pipeline_stable_diffusion_3_inpaint.py @@ -0,0 +1,1387 @@ +# Copyright 2025 Stability AI and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import ( + CLIPTextModelWithProjection, + CLIPTokenizer, + SiglipImageProcessor, + SiglipVisionModel, + T5EncoderModel, + T5TokenizerFast, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, SD3IPAdapterMixin, SD3LoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import SD3Transformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import StableDiffusion3PipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusion3InpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusion3InpaintPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16 + ... ) + >>> pipe.to("cuda") + >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench" + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + >>> source = load_image(img_url) + >>> mask = load_image(mask_url) + >>> image = pipe(prompt=prompt, image=source, mask_image=mask).images[0] + >>> image.save("sd3_inpainting.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusion3InpaintPipeline(DiffusionPipeline, SD3LoraLoaderMixin, FromSingleFileMixin, SD3IPAdapterMixin): + r""" + Args: + transformer ([`SD3Transformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant, + with an additional added projection layer that is initialized with a diagonal matrix with the `hidden_size` + as its dimension. + text_encoder_2 ([`CLIPTextModelWithProjection`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + text_encoder_3 ([`T5EncoderModel`]): + Frozen text-encoder. Stable Diffusion 3 uses + [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the + [t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_3 (`T5TokenizerFast`): + Tokenizer of class + [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer). + image_encoder (`SiglipVisionModel`, *optional*): + Pre-trained Vision Model for IP Adapter. + feature_extractor (`SiglipImageProcessor`, *optional*): + Image processor for IP Adapter. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->text_encoder_3->image_encoder->transformer->vae" + _optional_components = ["image_encoder", "feature_extractor"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "negative_pooled_prompt_embeds"] + + def __init__( + self, + transformer: SD3Transformer2DModel, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer_2: CLIPTokenizer, + text_encoder_3: T5EncoderModel, + tokenizer_3: T5TokenizerFast, + image_encoder: SiglipVisionModel | None = None, + feature_extractor: SiglipImageProcessor | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + text_encoder_3=text_encoder_3, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + tokenizer_3=tokenizer_3, + transformer=transformer, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, vae_latent_channels=latent_channels + ) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, + vae_latent_channels=latent_channels, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = ( + self.transformer.config.sample_size + if hasattr(self, "transformer") and self.transformer is not None + else 128 + ) + self.patch_size = ( + self.transformer.config.patch_size if hasattr(self, "transformer") and self.transformer is not None else 2 + ) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 256, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if self.text_encoder_3 is None: + return torch.zeros( + ( + batch_size * num_images_per_prompt, + max_sequence_length, + self.transformer.config.joint_attention_dim, + ), + device=device, + dtype=dtype, + ) + + text_inputs = self.tokenizer_3( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_3(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_3.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_3(text_input_ids.to(device))[0] + + dtype = self.text_encoder_3.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + clip_skip: int | None = None, + clip_model_index: int = 0, + ): + device = device or self._execution_device + + clip_tokenizers = [self.tokenizer, self.tokenizer_2] + clip_text_encoders = [self.text_encoder, self.text_encoder_2] + + tokenizer = clip_tokenizers[clip_model_index] + text_encoder = clip_text_encoders[clip_model_index] + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt) + pooled_prompt_embeds = pooled_prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds, pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + prompt_2: str | list[str], + prompt_3: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + prompt_embeds: torch.FloatTensor | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + negative_pooled_prompt_embeds: torch.FloatTensor | None = None, + clip_skip: int | None = None, + max_sequence_length: int = 256, + lora_scale: float | None = None, + ): + r""" + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in all text-encoders + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + used in all text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders. + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used in all the text-encoders. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, SD3LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + prompt_3 = prompt_3 or prompt + prompt_3 = [prompt_3] if isinstance(prompt_3, str) else prompt_3 + + prompt_embed, pooled_prompt_embed = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=0, + ) + prompt_2_embed, pooled_prompt_2_embed = self._get_clip_prompt_embeds( + prompt=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=clip_skip, + clip_model_index=1, + ) + clip_prompt_embeds = torch.cat([prompt_embed, prompt_2_embed], dim=-1) + + t5_prompt_embed = self._get_t5_prompt_embeds( + prompt=prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + clip_prompt_embeds = torch.nn.functional.pad( + clip_prompt_embeds, (0, t5_prompt_embed.shape[-1] - clip_prompt_embeds.shape[-1]) + ) + + prompt_embeds = torch.cat([clip_prompt_embeds, t5_prompt_embed], dim=-2) + pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + negative_prompt_3 = negative_prompt_3 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + negative_prompt_3 = ( + batch_size * [negative_prompt_3] if isinstance(negative_prompt_3, str) else negative_prompt_3 + ) + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embed, negative_pooled_prompt_embed = self._get_clip_prompt_embeds( + negative_prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=0, + ) + negative_prompt_2_embed, negative_pooled_prompt_2_embed = self._get_clip_prompt_embeds( + negative_prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + clip_skip=None, + clip_model_index=1, + ) + negative_clip_prompt_embeds = torch.cat([negative_prompt_embed, negative_prompt_2_embed], dim=-1) + + t5_negative_prompt_embed = self._get_t5_prompt_embeds( + prompt=negative_prompt_3, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + negative_clip_prompt_embeds = torch.nn.functional.pad( + negative_clip_prompt_embeds, + (0, t5_negative_prompt_embed.shape[-1] - negative_clip_prompt_embeds.shape[-1]), + ) + + negative_prompt_embeds = torch.cat([negative_clip_prompt_embeds, t5_negative_prompt_embed], dim=-2) + negative_pooled_prompt_embeds = torch.cat( + [negative_pooled_prompt_embed, negative_pooled_prompt_2_embed], dim=-1 + ) + + if self.text_encoder is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, SD3LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=None, + negative_prompt_2=None, + negative_prompt_3=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if ( + height % (self.vae_scale_factor * self.patch_size) != 0 + or width % (self.vae_scale_factor * self.patch_size) != 0 + ): + raise ValueError( + f"`height` and `width` have to be divisible by {self.vae_scale_factor * self.patch_size} but are {height} and {width}." + f"You can use height {height - height % (self.vae_scale_factor * self.patch_size)} and width {width - width % (self.vae_scale_factor * self.patch_size)}." + ) + + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_3 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_3`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + elif prompt_3 is not None and (not isinstance(prompt_3, str) and not isinstance(prompt_3, list)): + raise ValueError(f"`prompt_3` has to be of type `str` or `list` but is {type(prompt_3)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_3 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_3`: {negative_prompt_3} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}") + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + + if image.shape[1] == 16: + image_latents = image + else: + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.scale_noise(image_latents, timestep, noise) + else: + noise = latents.to(device) + latents = noise + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + num_images_per_prompt, + height, + width, + dtype, + device, + generator, + do_classifier_free_guidance, + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + masked_image = masked_image.to(device=device, dtype=dtype) + + if masked_image.shape[1] == 16: + masked_image_latents = masked_image + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + return mask, masked_image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.encode_image + def encode_image(self, image: PipelineImageInput, device: torch.device) -> torch.Tensor: + """Encodes the given image into a feature representation using a pre-trained image encoder. + + Args: + image (`PipelineImageInput`): + Input image to be encoded. + device: (`torch.device`): + Torch device. + + Returns: + `torch.Tensor`: The encoded image feature representation. + """ + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=self.dtype) + + return self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: torch.Tensor | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + ) -> torch.Tensor: + """Prepares image embeddings for use in the IP-Adapter. + + Either `ip_adapter_image` or `ip_adapter_image_embeds` must be passed. + + Args: + ip_adapter_image (`PipelineImageInput`, *optional*): + The input image to extract features from for IP-Adapter. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Precomputed image embeddings. + device: (`torch.device`, *optional*): + Torch device. + num_images_per_prompt (`int`, defaults to 1): + Number of images that should be generated per prompt. + do_classifier_free_guidance (`bool`, defaults to True): + Whether to use classifier free guidance or not. + """ + device = device or self._execution_device + + if ip_adapter_image_embeds is not None: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = ip_adapter_image_embeds.chunk(2) + else: + single_image_embeds = ip_adapter_image_embeds + elif ip_adapter_image is not None: + single_image_embeds = self.encode_image(ip_adapter_image, device) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.zeros_like(single_image_embeds) + else: + raise ValueError("Neither `ip_adapter_image_embeds` or `ip_adapter_image_embeds` were provided.") + + image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + + if do_classifier_free_guidance: + negative_image_embeds = torch.cat([single_negative_image_embeds] * num_images_per_prompt, dim=0) + image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0) + + return image_embeds.to(device=device) + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3.StableDiffusion3Pipeline.enable_sequential_cpu_offload + def enable_sequential_cpu_offload(self, *args, **kwargs): + if self.image_encoder is not None and "image_encoder" not in self._exclude_from_cpu_offload: + logger.warning( + "`pipe.enable_sequential_cpu_offload()` might fail for `image_encoder` if it uses " + "`torch.nn.MultiheadAttention`. You can exclude `image_encoder` from CPU offloading by calling " + "`pipe._exclude_from_cpu_offload.append('image_encoder')` before `pipe.enable_sequential_cpu_offload()`." + ) + + super().enable_sequential_cpu_offload(*args, **kwargs) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + prompt_3: str | list[str] | None = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: PipelineImageInput = None, + height: int = None, + width: int = None, + padding_mask_crop: int | None = None, + strength: float = 0.6, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 7.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + negative_prompt_3: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 256, + mu: float | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + will be used instead + prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to `tokenizer_3` and `text_encoder_3`. If not defined, `prompt` is + will be used instead + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a + single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one + color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B, + H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W, + 1)`, or `(H, W)`. + masked_image_latents (`torch.Tensor`, *optional*): + Pre-encoded latent of the masked image (for inpainting). If not provided, the masked image latents are + generated from `mask_image` and `image`. + height (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + width (`int`, *optional*, defaults to self.transformer.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used instead + negative_prompt_3 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_3` and + `text_encoder_3`. If not defined, `negative_prompt` is used instead + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image (`PipelineImageInput`, *optional*): + Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings for IP-Adapter. Should be a tensor of shape `(batch_size, num_images, + emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to + `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] instead of + a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 256): Maximum sequence length to use with the `prompt`. + mu (`float`, *optional*): `mu` value used for `dynamic_shifting`. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_3.StableDiffusion3PipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + height = height or self.transformer.config.sample_size * self.vae_scale_factor + width = width or self.transformer.config.sample_size * self.vae_scale_factor + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + prompt_3, + height, + width, + strength, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._clip_skip = clip_skip + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + prompt_3=prompt_3, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + negative_prompt_3=negative_prompt_3, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + device=device, + clip_skip=self.clip_skip, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + ) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) + + # 3. Prepare timesteps + scheduler_kwargs = {} + if self.scheduler.config.get("use_dynamic_shifting", None) and mu is None: + image_seq_len = (int(height) // self.vae_scale_factor // self.transformer.config.patch_size) * ( + int(width) // self.vae_scale_factor // self.transformer.config.patch_size + ) + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.16), + ) + scheduler_kwargs["mu"] = mu + elif mu is not None: + scheduler_kwargs["mu"] = mu + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, sigmas=sigmas, **scheduler_kwargs + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 4. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_transformer = self.transformer.config.in_channels + return_image_latents = num_channels_transformer == 16 + + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 6. Prepare mask latent variables + mask_condition = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is None: + masked_image = init_image * (mask_condition < 0.5) + else: + masked_image = masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask_condition, + masked_image, + batch_size, + num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # match the inpainting pipeline and will be updated with input + mask inpainting model later + if num_channels_transformer == 33: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if ( + num_channels_latents + num_channels_mask + num_channels_masked_image + != self.transformer.config.in_channels + ): + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.transformer`: {self.transformer.config} expects" + f" {self.transformer.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.transformer` or your `mask_image` or `image` input." + ) + elif num_channels_transformer != 16: + raise ValueError( + f"The transformer {self.transformer.__class__} should have 16 input channels or 33 input channels, not {self.transformer.config.in_channels}." + ) + + # 7. Prepare image embeddings + if (ip_adapter_image is not None and self.is_ip_adapter_active) or ip_adapter_image_embeds is not None: + ip_adapter_image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + if self.joint_attention_kwargs is None: + self._joint_attention_kwargs = {"ip_adapter_image_embeds": ip_adapter_image_embeds} + else: + self._joint_attention_kwargs.update(ip_adapter_image_embeds=ip_adapter_image_embeds) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latent_model_input.shape[0]) + + if num_channels_transformer == 33: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + pooled_projections=pooled_prompt_embeds, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + if num_channels_transformer == 16: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + negative_pooled_prompt_embeds = callback_outputs.pop( + "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds + ) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + else: + image = latents + + do_denormalize = [True] * image.shape[0] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusion3PipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_xl/__init__.py b/diffusers/pipelines/stable_diffusion_xl/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ab0b7183019b79bab7a612463f2b6cb9fd079647 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/__init__.py @@ -0,0 +1,56 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_additional_imports = {} +_import_structure = {"pipeline_output": ["StableDiffusionXLPipelineOutput"]} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_xl"] = ["StableDiffusionXLPipeline"] + _import_structure["pipeline_stable_diffusion_xl_img2img"] = ["StableDiffusionXLImg2ImgPipeline"] + _import_structure["pipeline_stable_diffusion_xl_inpaint"] = ["StableDiffusionXLInpaintPipeline"] + _import_structure["pipeline_stable_diffusion_xl_instruct_pix2pix"] = ["StableDiffusionXLInstructPix2PixPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_diffusion_xl import StableDiffusionXLPipeline + from .pipeline_stable_diffusion_xl_img2img import StableDiffusionXLImg2ImgPipeline + from .pipeline_stable_diffusion_xl_inpaint import StableDiffusionXLInpaintPipeline + from .pipeline_stable_diffusion_xl_instruct_pix2pix import StableDiffusionXLInstructPix2PixPipeline + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) + for name, value in _additional_imports.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/stable_diffusion_xl/pipeline_output.py b/diffusers/pipelines/stable_diffusion_xl/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..805a2f7d379d232b4dc9e6d98a612227fd0d3354 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from ...utils import BaseOutput + + +@dataclass +class StableDiffusionXLPipelineOutput(BaseOutput): + """ + Output class for Stable Diffusion pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image] | np.ndarray diff --git a/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py new file mode 100644 index 0000000000000000000000000000000000000000..ba94e3051fd35f941e062c0d7166306094173b29 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py @@ -0,0 +1,1308 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLPipeline + + >>> pipe = StableDiffusionXLPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLPipeline( + DiffusionPipeline, + StableDiffusionMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Prepare timesteps + is_neuron_device = device.type == "neuron" + if XLA_AVAILABLE or is_neuron_device: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 8.1 Apply denoising_end + if ( + self.denoising_end is not None + and isinstance(self.denoising_end, float) + and self.denoising_end > 0 + and self.denoising_end < 1 + ): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9. Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(0) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + # [Neuron] pre-cast timestep to float32 on device. Neuron XLA does not support + # int64 ops; the compiled UNet graph requires a float32 timestep input on-device. + t_unet = t.to(torch.float32).to(device) if is_neuron_device else t + noise_pred = self.unet( + latent_model_input, + t_unet, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + if not output_type == "latent": + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..c7a13ca025244d577469e641546b1dbad7e64aee --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py @@ -0,0 +1,1492 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import empty_device_cache, randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-refiner-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png" + + >>> init_image = load_image(url).convert("RGB") + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt, image=init_image).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLImg2ImgPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the + config of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + strength, + num_inference_steps, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + if num_inference_steps is None: + raise ValueError("`num_inference_steps` cannot be None.") + elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0: + raise ValueError( + f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type" + f" {type(num_inference_steps)}." + ) + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + def prepare_latents( + self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + latents_mean = latents_std = None + if hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None: + latents_mean = torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1) + if hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None: + latents_std = torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + empty_device_cache() + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + if image.shape[0] < batch_size and batch_size % image.shape[0] == 0: + image = torch.cat([image] * (batch_size // image.shape[0]), dim=0) + elif image.shape[0] < batch_size and batch_size % image.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image.shape[0]} to effective batch_size {batch_size} " + ) + + init_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + if latents_mean is not None and latents_std is not None: + latents_mean = latents_mean.to(device=device, dtype=dtype) + latents_std = latents_std.to(device=device, dtype=dtype) + init_latents = (init_latents - latents_mean) * self.vae.config.scaling_factor / latents_std + else: + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + strength: float = 0.3, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_start: float | None = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `list[torch.Tensor]` or `list[PIL.Image.Image]` or `list[np.ndarray]`): + The image(s) to modify with the pipeline. + strength (`float`, *optional*, defaults to 0.3): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image` + will be used as a starting point, adding more noise to it the larger the `strength`. The number of + denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will + be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of + `denoising_start` being declared as an integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refine Image + Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + strength, + num_inference_steps, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image) + + # 5. Prepare timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + add_noise = True if self.denoising_start is None else False + + # 6. Prepare latent variables + if latents is None: + latents = self.prepare_latents( + image, + latent_timestep, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + add_noise, + ) + # 7. Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 8. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 9. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 9.1 Apply denoising_end + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 9.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..3f18cbe21d0fe12f89859d188dba2a487f3d87c5 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py @@ -0,0 +1,1734 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = StableDiffusionXLInpaintPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", + ... torch_dtype=torch.float16, + ... variant="fp16", + ... use_safetensors=True, + ... ) + >>> pipe.to("cuda") + + >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" + >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" + + >>> init_image = load_image(img_url).convert("RGB") + >>> mask_image = load_image(mask_url).convert("RGB") + + >>> prompt = "A majestic tiger sitting on a bench" + >>> image = pipe( + ... prompt=prompt, image=init_image, mask_image=mask_image, num_inference_steps=50, strength=0.80 + ... ).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def mask_pil_to_torch(mask, height, width): + # preprocess mask + if isinstance(mask, (PIL.Image.Image, np.ndarray)): + mask = [mask] + + if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): + mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] + mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) + mask = mask.astype(np.float32) / 255.0 + elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): + mask = np.concatenate([m[None, None, :] for m in mask], axis=0) + + mask = torch.from_numpy(mask) + return mask + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLInpaintPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + FromSingleFileMixin, + IPAdapterMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config + of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "image_encoder", + "feature_extractor", + ] + _callback_tensor_inputs = [ + "latents", + "prompt_embeds", + "add_text_embeds", + "add_time_ids", + "mask", + "masked_image_latents", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + image_encoder: CLIPVisionModelWithProjection = None, + feature_extractor: CLIPImageProcessor = None, + requires_aesthetics_score: bool = False, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + image_encoder=image_encoder, + feature_extractor=feature_extractor, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.register_to_config(requires_aesthetics_score=requires_aesthetics_score) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + padding_mask_crop=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + if padding_mask_crop is not None: + if not isinstance(image, PIL.Image.Image): + raise ValueError( + f"The image should be a PIL image when inpainting mask crop, but is of type {type(image)}." + ) + if not isinstance(mask_image, PIL.Image.Image): + raise ValueError( + f"The mask image should be a PIL image when inpainting mask crop, but is of type" + f" {type(mask_image)}." + ) + if output_type != "pil": + raise ValueError(f"The output type should be PIL when inpainting mask crop, but is {output_type}.") + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + image=None, + timestep=None, + is_strength_max=True, + add_noise=True, + return_noise=False, + return_image_latents=False, + ): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if (image is None or timestep is None) and not is_strength_max: + raise ValueError( + "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." + "However, either the image or the noise timestep has not been provided." + ) + + if image.shape[1] == 4: + image_latents = image.to(device=device, dtype=dtype) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + elif return_image_latents or (latents is None and not is_strength_max): + image = image.to(device=device, dtype=dtype) + image_latents = self._encode_vae_image(image=image, generator=generator) + image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) + + if latents is None and add_noise: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # if strength is 1. then initialise the latents to noise, else initial to image + noise + latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) + # if pure noise then scale the initial latents by the Scheduler's init sigma + latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents + elif add_noise: + noise = latents.to(device) + latents = noise * self.scheduler.init_noise_sigma + else: + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = image_latents.to(device) + + outputs = (latents,) + + if return_noise: + outputs += (noise,) + + if return_image_latents: + outputs += (image_latents,) + + return outputs + + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + dtype = image.dtype + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + image_latents = image_latents.to(dtype) + image_latents = self.vae.config.scaling_factor * image_latents + + return image_latents + + def prepare_mask_latents( + self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance + ): + # resize the mask to latents shape as we concatenate the mask to the latents + # we do that before converting to dtype to avoid breaking in case we're using cpu_offload + # and half precision + mask = torch.nn.functional.interpolate( + mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) + ) + mask = mask.to(device=device, dtype=dtype) + + # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + + mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask + + if masked_image is not None and masked_image.shape[1] == 4: + masked_image_latents = masked_image + else: + masked_image_latents = None + + if masked_image is not None: + if masked_image_latents is None: + masked_image = masked_image.to(device=device, dtype=dtype) + masked_image_latents = self._encode_vae_image(masked_image, generator=generator) + + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat( + batch_size // masked_image_latents.shape[0], 1, 1, 1 + ) + + masked_image_latents = ( + torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents + ) + + # aligning device to prevent device errors when concating it with the latent model input + masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) + + return mask, masked_image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + else: + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + + num_inference_steps = (self.scheduler.timesteps < discrete_timestep_cutoff).sum().item() + if self.scheduler.order == 2 and num_inference_steps % 2 == 0: + # if the scheduler is a 2nd order scheduler we might have to do +1 + # because `num_inference_steps` might be even given that every timestep + # (except the highest one) is duplicated. If `num_inference_steps` is even it would + # mean that we cut the timesteps in the middle of the denoising step + # (between 1st and 2nd derivative) which leads to incorrect results. By adding 1 + # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler + num_inference_steps = num_inference_steps + 1 + + # because t_n+1 >= t_n, we slice the timesteps starting from the end + t_start = len(self.scheduler.timesteps) - num_inference_steps + timesteps = self.scheduler.timesteps[t_start:] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start) + return timesteps, num_inference_steps + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids + def _get_add_time_ids( + self, + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype, + text_encoder_projection_dim=None, + ): + if self.config.requires_aesthetics_score: + add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,)) + add_neg_time_ids = list( + negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,) + ) + else: + add_time_ids = list(original_size + crops_coords_top_left + target_size) + add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if ( + expected_add_embed_dim > passed_add_embed_dim + and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model." + ) + elif ( + expected_add_embed_dim < passed_add_embed_dim + and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim + ): + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model." + ) + elif expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype) + + return add_time_ids, add_neg_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def denoising_end(self): + return self._denoising_end + + @property + def denoising_start(self): + return self._denoising_start + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: torch.Tensor = None, + height: int | None = None, + width: int | None = None, + padding_mask_crop: int | None = None, + strength: float = 0.9999, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_start: float | None = None, + denoising_end: float | None = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + aesthetic_score: float = 6.0, + negative_aesthetic_score: float = 2.5, + clip_skip: int | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + **kwargs, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will + be masked out with `mask_image` and repainted according to `prompt`. + mask_image (`PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted + to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) + instead of 3, so the expected shape would be `(B, H, W, 1)`. + masked_image_latents (`torch.Tensor`, *optional*): + Pre-encoded latent of the masked image (for inpainting). + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. This is set to 1024 by default for the best results. + Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + padding_mask_crop (`int`, *optional*, defaults to `None`): + The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to + image and mask_image. If `padding_mask_crop` is not `None`, it will first find a rectangular region + with the same aspect ration of the image and contains all masked area, and then expand that area based + on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before + resizing to the original image size for inpainting. This is useful when the masked area is small while + the image is large and contain information irrelevant for inpainting, such as background. + strength (`float`, *optional*, defaults to 0.9999): + Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be + between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the + `strength`. The number of denoising steps depends on the amount of noise initially added. When + `strength` is 1, added noise will be maximum and the denoising process will run for the full number of + iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked + portion of the reference `image`. Note that in the case of `denoising_start` being declared as an + integer, the value of `strength` will be ignored. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_start (`float`, *optional*): + When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be + bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and + it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, + strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline + is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be + denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the + final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline + forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output). + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + aesthetic_score (`float`, *optional*, defaults to 6.0): + Used to simulate an aesthetic score of the generated image by influencing the positive text condition. + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_aesthetic_score (`float`, *optional*, defaults to 2.5): + Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to + simulate an aesthetic score of the generated image by influencing the negative text condition. + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf). + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. `tuple. When returning a tuple, the first element is a list with the generated images. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`", + ) + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + # 1. Check inputs + self.check_inputs( + prompt, + prompt_2, + image, + mask_image, + height, + width, + strength, + callback_steps, + output_type, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + callback_on_step_end_tensor_inputs, + padding_mask_crop, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._denoising_end = denoising_end + self._denoising_start = denoising_start + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + text_encoder_lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + clip_skip=self.clip_skip, + ) + + # 4. set timesteps + def denoising_value_valid(dnv): + return isinstance(dnv, float) and 0 < dnv < 1 + + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, + strength, + device, + denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None, + ) + # check that number of inference steps is not < 1 - as this doesn't make sense + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + # at which timestep to set the initial noise (n.b. 50% if strength is 0.5) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise + is_strength_max = strength == 1.0 + + # 5. Preprocess mask and image + if padding_mask_crop is not None: + crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop) + resize_mode = "fill" + else: + crops_coords = None + resize_mode = "default" + + original_image = image + init_image = self.image_processor.preprocess( + image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode + ) + init_image = init_image.to(dtype=torch.float32) + + mask = self.mask_processor.preprocess( + mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords + ) + + if masked_image_latents is not None: + masked_image = masked_image_latents + elif init_image.shape[1] == 4: + # if images are in latent space, we can't mask it + masked_image = None + else: + masked_image = init_image * (mask < 0.5) + + # 6. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + num_channels_unet = self.unet.config.in_channels + return_image_latents = num_channels_unet == 4 + + add_noise = True if self.denoising_start is None else False + latents_outputs = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + image=init_image, + timestep=latent_timestep, + is_strength_max=is_strength_max, + add_noise=add_noise, + return_noise=True, + return_image_latents=return_image_latents, + ) + + if return_image_latents: + latents, noise, image_latents = latents_outputs + else: + latents, noise = latents_outputs + + # 7. Prepare mask latent variables + mask, masked_image_latents = self.prepare_mask_latents( + mask, + masked_image, + batch_size * num_images_per_prompt, + height, + width, + prompt_embeds.dtype, + device, + generator, + self.do_classifier_free_guidance, + ) + + # 8. Check that sizes of mask, masked image and latents match + if num_channels_unet == 9: + # default case for stable-diffusion-v1-5/stable-diffusion-inpainting + num_channels_mask = mask.shape[1] + num_channels_masked_image = masked_image_latents.shape[1] + if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" + f" = {num_channels_latents + num_channels_masked_image + num_channels_mask}. Please verify the config of" + " `pipeline.unet` or your `mask_image` or `image` input." + ) + elif num_channels_unet != 4: + raise ValueError( + f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." + ) + # 8.1 Prepare extra step kwargs. + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + height, width = latents.shape[-2:] + height = height * self.vae_scale_factor + width = width * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 10. Prepare added time ids & embeddings + if negative_original_size is None: + negative_original_size = original_size + if negative_target_size is None: + negative_target_size = target_size + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids, add_neg_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + aesthetic_score, + negative_aesthetic_score, + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1) + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1) + add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device) + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + if ( + self.denoising_end is not None + and self.denoising_start is not None + and denoising_value_valid(self.denoising_end) + and denoising_value_valid(self.denoising_start) + and self.denoising_start >= self.denoising_end + ): + raise ValueError( + f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " + + f" {self.denoising_end} when using type float." + ) + elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (self.denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + # 11.1 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + # concat latents, mask, masked_image_latents in the channel dimension + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + if num_channels_unet == 9: + latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if num_channels_unet == 4: + init_latents_proper = image_latents + if self.do_classifier_free_guidance: + init_mask, _ = mask.chunk(2) + else: + init_mask = mask + + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.add_noise( + init_latents_proper, noise, torch.tensor([noise_timestep]) + ) + + latents = (1 - init_mask) * init_latents_proper + init_mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) + add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + return StableDiffusionXLPipelineOutput(images=latents) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + if padding_mask_crop is not None: + image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image] + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py new file mode 100644 index 0000000000000000000000000000000000000000..bcd337414bac7bc3e6a9eb3fd1e9a8876da2508a --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py @@ -0,0 +1,977 @@ +# Copyright 2025 Harutatsu Akiyama and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL.Image +import torch +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + USE_PEFT_BACKEND, + deprecate, + is_invisible_watermark_available, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from .pipeline_output import StableDiffusionXLPipelineOutput + + +if is_invisible_watermark_available(): + from .watermark import StableDiffusionXLWatermarker + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLInstructPix2PixPipeline + >>> from diffusers.utils import load_image + + >>> resolution = 768 + >>> image = load_image( + ... "https://hf.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png" + ... ).resize((resolution, resolution)) + >>> edit_instruction = "Turn sky into a cloudy one" + + >>> pipe = StableDiffusionXLInstructPix2PixPipeline.from_pretrained( + ... "diffusers/sdxl-instructpix2pix-768", torch_dtype=torch.float16 + ... ).to("cuda") + + >>> edited_image = pipe( + ... prompt=edit_instruction, + ... image=image, + ... height=resolution, + ... width=resolution, + ... guidance_scale=3.0, + ... image_guidance_scale=1.5, + ... num_inference_steps=30, + ... ).images[0] + >>> edited_image + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class StableDiffusionXLInstructPix2PixPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + FromSingleFileMixin, + StableDiffusionXLLoraLoaderMixin, +): + r""" + Pipeline for pixel-level image editing by following text instructions. Based on Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`): + Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config + of `stabilityai/stable-diffusion-xl-refiner-1-0`. + force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): + Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of + `stabilityai/stable-diffusion-xl-base-1-0`. + add_watermarker (`bool`, *optional*): + Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to + watermark output images. If not defined, it will default to True if the package is installed, otherwise no + watermarker will be used. + is_cosxl_edit (`bool`, *optional*): + When set the image latents are scaled. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae" + _optional_components = ["tokenizer", "tokenizer_2", "text_encoder", "text_encoder_2"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: bool | None = None, + is_cosxl_edit: bool | None = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + self.is_cosxl_edit = is_cosxl_edit + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + prompt_embeds = prompt_embeds.hidden_states[-2] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt, negative_prompt_2] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + prompt_embeds_dtype = self.text_encoder_2.dtype if self.text_encoder_2 is not None else self.unet.dtype + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def prepare_image_latents( + self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None + ): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + image_latents = image + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + if needs_upcasting: + image = image.float() + self.upcast_vae() + + image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax") + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + # expand image_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + else: + image_latents = torch.cat([image_latents], dim=0) + + if do_classifier_free_guidance: + uncond_image_latents = torch.zeros_like(image_latents) + image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0) + + if image_latents.dtype != self.vae.dtype: + image_latents = image_latents.to(dtype=self.vae.dtype) + + if self.is_cosxl_edit: + image_latents = image_latents * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 100, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + image_guidance_scale: float = 1.5, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor` or `PIL.Image.Image` or `np.ndarray` or `list[torch.Tensor]` or `list[PIL.Image.Image]` or `list[np.ndarray]`): + The image(s) to modify with the pipeline. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + image_guidance_scale (`float`, *optional*, defaults to 1.5): + Image guidance scale is to push the generated image towards the initial image `image`. Image guidance + scale is enabled by setting `image_guidance_scale > 1`. Higher image guidance scale encourages to + generate images that are closely linked to the source image `image`, usually at the expense of lower + image quality. This pipeline requires a value of at least `1`. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds) + + if image is None: + raise ValueError("`image` input cannot be undefined.") + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 and image_guidance_scale >= 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + # 4. Preprocess image + image = self.image_processor.preprocess(image, height=height, width=width).to(device) + + # 5. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 6. Prepare Image latents + image_latents = self.prepare_image_latents( + image, + batch_size, + num_images_per_prompt, + prompt_embeds.dtype, + device, + do_classifier_free_guidance, + ) + + # 7. Prepare latent variables + num_channels_latents = self.vae.config.latent_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 8. Check that shapes of latents and image match the UNet channels + num_channels_image = image_latents.shape[1] + if num_channels_latents + num_channels_image != self.unet.config.in_channels: + raise ValueError( + f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" + f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" + f" `num_channels_image`: {num_channels_image} " + f" = {num_channels_latents + num_channels_image}. Please verify the config of" + " `pipeline.unet` or your `image` input." + ) + + # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 10. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + + if do_classifier_free_guidance: + # The extra concat similar to how it's done in SD InstructPix2Pix. + prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds], dim=0) + add_text_embeds = torch.cat( + [add_text_embeds, negative_pooled_prompt_embeds, negative_pooled_prompt_embeds], dim=0 + ) + add_time_ids = torch.cat([add_time_ids, add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 11. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # Expand the latents if we are doing classifier free guidance. + # The latents are expanded 3 times because for pix2pix the guidance + # is applied for both the text and the input image. + latent_model_input = torch.cat([latents] * 3) if do_classifier_free_guidance else latents + + # concat latents, image_latents in the channel dimension + scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + noise_pred = self.unet( + scaled_latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_text, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3) + noise_pred = ( + noise_pred_uncond + + guidance_scale * (noise_pred_text - noise_pred_image) + + image_guidance_scale * (noise_pred_image - noise_pred_uncond) + ) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + elif latents.dtype != self.vae.dtype: + if torch.backends.mps.is_available(): + # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + self.vae = self.vae.to(latents.dtype) + + # unscale/denormalize the latents + # denormalize with the mean and std if available and not None + has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None + has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None + if has_latents_mean and has_latents_std: + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents_std = ( + torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) + ) + latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean + else: + latents = latents / self.vae.config.scaling_factor + + image = self.vae.decode(latents, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + return StableDiffusionXLPipelineOutput(images=latents) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/stable_diffusion_xl/watermark.py b/diffusers/pipelines/stable_diffusion_xl/watermark.py new file mode 100644 index 0000000000000000000000000000000000000000..70d06bb6320d890ea90542cc7551bc42d7c59451 --- /dev/null +++ b/diffusers/pipelines/stable_diffusion_xl/watermark.py @@ -0,0 +1,42 @@ +import numpy as np +import torch + +from ...utils import is_invisible_watermark_available + + +if is_invisible_watermark_available(): + from imwatermark import WatermarkEncoder + + +# Copied from https://github.com/Stability-AI/generative-models/blob/613af104c6b85184091d42d374fef420eddb356d/scripts/demo/streamlit_helpers.py#L66 +WATERMARK_MESSAGE = 0b101100111110110010010000011110111011000110011110 +# bin(x)[2:] gives bits of x as str, use int to convert them to 0/1 +WATERMARK_BITS = [int(bit) for bit in bin(WATERMARK_MESSAGE)[2:]] + + +class StableDiffusionXLWatermarker: + def __init__(self): + self.watermark = WATERMARK_BITS + self.encoder = WatermarkEncoder() + + self.encoder.set_watermark("bits", self.watermark) + + def apply_watermark(self, images: torch.Tensor): + # can't encode images that are smaller than 256 + if images.shape[-1] < 256: + return images + + images = (255 * (images / 2 + 0.5)).cpu().permute(0, 2, 3, 1).float().numpy() + + # Convert RGB to BGR, which is the channel order expected by the watermark encoder. + images = images[:, :, :, ::-1] + + # Add watermark and convert BGR back to RGB + images = [self.encoder.encode(image, "dwtDct")[:, :, ::-1] for image in images] + + images = np.array(images) + + images = torch.from_numpy(images).permute(0, 3, 1, 2) + + images = torch.clamp(2 * (images / 255 - 0.5), min=-1.0, max=1.0) + return images diff --git a/diffusers/pipelines/stable_video_diffusion/__init__.py b/diffusers/pipelines/stable_video_diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3bd4dc78966e217d85769691b98ed8fb0b6ac05c --- /dev/null +++ b/diffusers/pipelines/stable_video_diffusion/__init__.py @@ -0,0 +1,58 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + BaseOutput, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure.update( + { + "pipeline_stable_video_diffusion": [ + "StableVideoDiffusionPipeline", + "StableVideoDiffusionPipelineOutput", + ], + } + ) + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_stable_video_diffusion import ( + StableVideoDiffusionPipeline, + StableVideoDiffusionPipelineOutput, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py b/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..fd46ddd1b60263fd1a1413cddb8d0ea35f090016 --- /dev/null +++ b/diffusers/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py @@ -0,0 +1,743 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from dataclasses import dataclass +from typing import Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection + +from ...image_processor import PipelineImageInput +from ...models import AutoencoderKLTemporalDecoder, UNetSpatioTemporalConditionModel +from ...schedulers import EulerDiscreteScheduler +from ...utils import BaseOutput, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import is_compiled_module, randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from diffusers import StableVideoDiffusionPipeline + >>> from diffusers.utils import load_image, export_to_video + + >>> pipe = StableVideoDiffusionPipeline.from_pretrained( + ... "stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16" + ... ) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd-docstring-example.jpeg" + ... ) + >>> image = image.resize((1024, 576)) + + >>> frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0] + >>> export_to_video(frames, "generated.mp4", fps=7) + ``` +""" + + +def _append_dims(x, target_dims): + """Appends dimensions to the end of a tensor until it has target_dims dimensions.""" + dims_to_append = target_dims - x.ndim + if dims_to_append < 0: + raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less") + return x[(...,) + (None,) * dims_to_append] + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +@dataclass +class StableVideoDiffusionPipelineOutput(BaseOutput): + r""" + Output class for Stable Video Diffusion pipeline. + + Args: + frames (`[list[list[PIL.Image.Image]]`, `np.ndarray`, `torch.Tensor`]): + list of denoised PIL images of length `batch_size` or numpy array or torch tensor of shape `(batch_size, + num_frames, height, width, num_channels)`. + """ + + frames: list[list[PIL.Image.Image]] | np.ndarray | torch.Tensor + + +class StableVideoDiffusionPipeline(DiffusionPipeline): + r""" + Pipeline to generate video from an input image using Stable Video Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + vae ([`AutoencoderKLTemporalDecoder`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + image_encoder ([`~transformers.CLIPVisionModelWithProjection`]): + Frozen CLIP image-encoder + ([laion/CLIP-ViT-H-14-laion2B-s32B-b79K](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K)). + unet ([`UNetSpatioTemporalConditionModel`]): + A `UNetSpatioTemporalConditionModel` to denoise the encoded image latents. + scheduler ([`EulerDiscreteScheduler`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images. + """ + + model_cpu_offload_seq = "image_encoder->unet->vae" + _callback_tensor_inputs = ["latents"] + + def __init__( + self, + vae: AutoencoderKLTemporalDecoder, + image_encoder: CLIPVisionModelWithProjection, + unet: UNetSpatioTemporalConditionModel, + scheduler: EulerDiscreteScheduler, + feature_extractor: CLIPImageProcessor, + ): + super().__init__() + + self.register_modules( + vae=vae, + image_encoder=image_encoder, + unet=unet, + scheduler=scheduler, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(do_resize=True, vae_scale_factor=self.vae_scale_factor) + + def _encode_image( + self, + image: PipelineImageInput, + device: str | torch.device, + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + ) -> torch.Tensor: + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.video_processor.pil_to_numpy(image) + image = self.video_processor.numpy_to_pt(image) + + # We normalize the image before resizing to match with the original implementation. + # Then we unnormalize it after resizing. + image = image * 2.0 - 1.0 + image = _resize_with_antialiasing(image, (224, 224)) + image = (image + 1.0) / 2.0 + + # Normalize the image with for CLIP input + image = self.feature_extractor( + images=image, + do_normalize=True, + do_center_crop=False, + do_resize=False, + do_rescale=False, + return_tensors="pt", + ).pixel_values + + image = image.to(device=device, dtype=dtype) + image_embeddings = self.image_encoder(image).image_embeds + image_embeddings = image_embeddings.unsqueeze(1) + + # duplicate image embeddings for each generation per prompt, using mps friendly method + bs_embed, seq_len, _ = image_embeddings.shape + image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1) + image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + negative_image_embeddings = torch.zeros_like(image_embeddings) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_embeddings = torch.cat([negative_image_embeddings, image_embeddings]) + + return image_embeddings + + def _encode_vae_image( + self, + image: torch.Tensor, + device: str | torch.device, + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + ): + image = image.to(device=device) + image_latents = self.vae.encode(image).latent_dist.mode() + + # duplicate image_latents for each generation per prompt, using mps friendly method + image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1) + + if do_classifier_free_guidance: + negative_image_latents = torch.zeros_like(image_latents) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + image_latents = torch.cat([negative_image_latents, image_latents]) + + return image_latents + + def _get_add_time_ids( + self, + fps: int, + motion_bucket_id: int, + noise_aug_strength: float, + dtype: torch.dtype, + batch_size: int, + num_videos_per_prompt: int, + do_classifier_free_guidance: bool, + ): + add_time_ids = [fps, motion_bucket_id, noise_aug_strength] + + passed_add_embed_dim = self.unet.config.addition_time_embed_dim * len(add_time_ids) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + add_time_ids = add_time_ids.repeat(batch_size * num_videos_per_prompt, 1) + + if do_classifier_free_guidance: + add_time_ids = torch.cat([add_time_ids, add_time_ids]) + + return add_time_ids + + def decode_latents(self, latents: torch.Tensor, num_frames: int, decode_chunk_size: int = 14): + # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width] + latents = latents.flatten(0, 1) + + latents = 1 / self.vae.config.scaling_factor * latents + + forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward + accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys()) + + # decode decode_chunk_size frames at a time to avoid OOM + frames = [] + for i in range(0, latents.shape[0], decode_chunk_size): + num_frames_in = latents[i : i + decode_chunk_size].shape[0] + decode_kwargs = {} + if accepts_num_frames: + # we only pass num_frames_in if it's expected + decode_kwargs["num_frames"] = num_frames_in + + frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample + frames.append(frame) + frames = torch.cat(frames, dim=0) + + # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width] + frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4) + + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + frames = frames.float() + return frames + + def check_inputs(self, image, height, width): + if ( + not isinstance(image, torch.Tensor) + and not isinstance(image, PIL.Image.Image) + and not isinstance(image, list) + ): + raise ValueError( + "`image` has to be of type `torch.Tensor` or `PIL.Image.Image` or `list[PIL.Image.Image]` but is" + f" {type(image)}" + ) + + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + def prepare_latents( + self, + batch_size: int, + num_frames: int, + num_channels_latents: int, + height: int, + width: int, + dtype: torch.dtype, + device: str | torch.device, + generator: torch.Generator, + latents: torch.Tensor | None = None, + ): + shape = ( + batch_size, + num_frames, + num_channels_latents // 2, + height // self.vae_scale_factor, + width // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + if isinstance(self.guidance_scale, (int, float)): + return self.guidance_scale > 1 + return self.guidance_scale.max() > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PIL.Image.Image | list[PIL.Image.Image] | torch.Tensor, + height: int = 576, + width: int = 1024, + num_frames: int | None = None, + num_inference_steps: int = 25, + sigmas: list[float] | None = None, + min_guidance_scale: float = 1.0, + max_guidance_scale: float = 3.0, + fps: int = 7, + motion_bucket_id: int = 127, + noise_aug_strength: float = 0.02, + decode_chunk_size: int | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + output_type: str | None = "pil", + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + return_dict: bool = True, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PIL.Image.Image` or `list[PIL.Image.Image]` or `torch.Tensor`): + Image(s) to guide image generation. If you provide a tensor, the expected value range is between `[0, + 1]`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_frames (`int`, *optional*): + The number of video frames to generate. Defaults to `self.unet.config.num_frames` (14 for + `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`). + num_inference_steps (`int`, *optional*, defaults to 25): + The number of denoising steps. More denoising steps usually lead to a higher quality video at the + expense of slower inference. This parameter is modulated by `strength`. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + min_guidance_scale (`float`, *optional*, defaults to 1.0): + The minimum guidance scale. Used for the classifier free guidance with first frame. + max_guidance_scale (`float`, *optional*, defaults to 3.0): + The maximum guidance scale. Used for the classifier free guidance with last frame. + fps (`int`, *optional*, defaults to 7): + Frames per second. The rate at which the generated images shall be exported to a video after + generation. Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training. + motion_bucket_id (`int`, *optional*, defaults to 127): + Used for conditioning the amount of motion for the generation. The higher the number the more motion + will be in the video. + noise_aug_strength (`float`, *optional*, defaults to 0.02): + The amount of noise added to the init image, the higher it is the less the video will look like the + init image. Increase it for more motion. + decode_chunk_size (`int`, *optional*): + The number of frames to decode at a time. Higher chunk size leads to better temporal consistency at the + expense of more memory usage. By default, the decoder decodes all frames at once for maximal quality. + For lower memory usage, reduce `decode_chunk_size`. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of videos to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `pil`, `np` or `pt`. + callback_on_step_end (`Callable`, *optional*): + A function that is called at the end of each denoising step during inference. The function is called + with the following arguments: + `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. + `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is + returned, otherwise a `tuple` of (`list[list[PIL.Image.Image]]` or `np.ndarray` or `torch.Tensor`) is + returned. + """ + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + + num_frames = num_frames if num_frames is not None else self.unet.config.num_frames + decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames + + # 1. Check inputs. Raise error if not correct + self.check_inputs(image, height, width) + + # 2. Define call parameters + if isinstance(image, PIL.Image.Image): + batch_size = 1 + elif isinstance(image, list): + batch_size = len(image) + else: + batch_size = image.shape[0] + device = self._execution_device + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + self._guidance_scale = max_guidance_scale + + # 3. Encode input image + image_embeddings = self._encode_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance) + + # NOTE: Stable Video Diffusion was conditioned on fps - 1, which is why it is reduced here. + # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188 + fps = fps - 1 + + # 4. Encode input image using VAE + image = self.video_processor.preprocess(image, height=height, width=width).to(device) + noise = randn_tensor(image.shape, generator=generator, device=device, dtype=image.dtype) + image = image + noise_aug_strength * noise + + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + if needs_upcasting: + self.vae.to(dtype=torch.float32) + + image_latents = self._encode_vae_image( + image, + device=device, + num_videos_per_prompt=num_videos_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + ) + image_latents = image_latents.to(image_embeddings.dtype) + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + + # Repeat the image latents for each frame so we can concatenate them with the noise + # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width] + image_latents = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1) + + # 5. Get Added Time IDs + added_time_ids = self._get_add_time_ids( + fps, + motion_bucket_id, + noise_aug_strength, + image_embeddings.dtype, + batch_size, + num_videos_per_prompt, + self.do_classifier_free_guidance, + ) + added_time_ids = added_time_ids.to(device) + + # 6. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, None, sigmas + ) + + # 7. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_frames, + num_channels_latents, + height, + width, + image_embeddings.dtype, + device, + generator, + latents, + ) + + # 8. Prepare guidance scale + guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0) + guidance_scale = guidance_scale.to(device, latents.dtype) + guidance_scale = guidance_scale.repeat(batch_size * num_videos_per_prompt, 1) + guidance_scale = _append_dims(guidance_scale, latents.ndim) + + self._guidance_scale = guidance_scale + + # 9. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # Concatenate image_latents over channels dimension + latent_model_input = torch.cat([latent_model_input, image_latents], dim=2) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=image_embeddings, + added_time_ids=added_time_ids, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents).prev_sample + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + frames = self.decode_latents(latents, num_frames, decode_chunk_size) + frames = self.video_processor.postprocess_video(video=frames, output_type=output_type) + else: + frames = latents + + self.maybe_free_model_hooks() + + if not return_dict: + return frames + + return StableVideoDiffusionPipelineOutput(frames=frames) + + +# resizing utils +# TODO: clean up later +def _resize_with_antialiasing(input, size, interpolation="bicubic", align_corners=True): + h, w = input.shape[-2:] + factors = (h / size[0], w / size[1]) + + # First, we have to determine sigma + # Taken from skimage: https://github.com/scikit-image/scikit-image/blob/v0.19.2/skimage/transform/_warps.py#L171 + sigmas = ( + max((factors[0] - 1.0) / 2.0, 0.001), + max((factors[1] - 1.0) / 2.0, 0.001), + ) + + # Now kernel size. Good results are for 3 sigma, but that is kind of slow. Pillow uses 1 sigma + # https://github.com/python-pillow/Pillow/blob/master/src/libImaging/Resample.c#L206 + # But they do it in the 2 passes, which gives better results. Let's try 2 sigmas for now + ks = int(max(2.0 * 2 * sigmas[0], 3)), int(max(2.0 * 2 * sigmas[1], 3)) + + # Make sure it is odd + if (ks[0] % 2) == 0: + ks = ks[0] + 1, ks[1] + + if (ks[1] % 2) == 0: + ks = ks[0], ks[1] + 1 + + input = _gaussian_blur2d(input, ks, sigmas) + + output = torch.nn.functional.interpolate(input, size=size, mode=interpolation, align_corners=align_corners) + return output + + +def _compute_padding(kernel_size): + """Compute padding tuple.""" + # 4 or 6 ints: (padding_left, padding_right,padding_top,padding_bottom) + # https://pytorch.org/docs/stable/nn.html#torch.nn.functional.pad + if len(kernel_size) < 2: + raise AssertionError(kernel_size) + computed = [k - 1 for k in kernel_size] + + # for even kernels we need to do asymmetric padding :( + out_padding = 2 * len(kernel_size) * [0] + + for i in range(len(kernel_size)): + computed_tmp = computed[-(i + 1)] + + pad_front = computed_tmp // 2 + pad_rear = computed_tmp - pad_front + + out_padding[2 * i + 0] = pad_front + out_padding[2 * i + 1] = pad_rear + + return out_padding + + +def _filter2d(input, kernel): + # prepare kernel + b, c, h, w = input.shape + tmp_kernel = kernel[:, None, ...].to(device=input.device, dtype=input.dtype) + + tmp_kernel = tmp_kernel.expand(-1, c, -1, -1) + + height, width = tmp_kernel.shape[-2:] + + padding_shape: list[int] = _compute_padding([height, width]) + input = torch.nn.functional.pad(input, padding_shape, mode="reflect") + + # kernel and input tensor reshape to align element-wise or batch-wise params + tmp_kernel = tmp_kernel.reshape(-1, 1, height, width) + input = input.view(-1, tmp_kernel.size(0), input.size(-2), input.size(-1)) + + # convolve the tensor with the kernel. + output = torch.nn.functional.conv2d(input, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1) + + out = output.view(b, c, h, w) + return out + + +def _gaussian(window_size: int, sigma): + if isinstance(sigma, float): + sigma = torch.tensor([[sigma]]) + + batch_size = sigma.shape[0] + + x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1) + + if window_size % 2 == 0: + x = x + 0.5 + + gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0))) + + return gauss / gauss.sum(-1, keepdim=True) + + +def _gaussian_blur2d(input, kernel_size, sigma): + if isinstance(sigma, tuple): + sigma = torch.tensor([sigma], dtype=input.dtype) + else: + sigma = sigma.to(dtype=input.dtype) + + ky, kx = int(kernel_size[0]), int(kernel_size[1]) + bs = sigma.shape[0] + kernel_x = _gaussian(kx, sigma[:, 1].view(bs, 1)) + kernel_y = _gaussian(ky, sigma[:, 0].view(bs, 1)) + out_x = _filter2d(input, kernel_x[..., None, :]) + out = _filter2d(out_x, kernel_y[..., None]) + + return out diff --git a/diffusers/pipelines/t2i_adapter/__init__.py b/diffusers/pipelines/t2i_adapter/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..08c22a2707fe55770a519db481954881c1cad26e --- /dev/null +++ b/diffusers/pipelines/t2i_adapter/__init__.py @@ -0,0 +1,47 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_stable_diffusion_adapter"] = ["StableDiffusionAdapterPipeline"] + _import_structure["pipeline_stable_diffusion_xl_adapter"] = ["StableDiffusionXLAdapterPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * # noqa F403 + else: + from .pipeline_stable_diffusion_adapter import StableDiffusionAdapterPipeline + from .pipeline_stable_diffusion_xl_adapter import StableDiffusionXLAdapterPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py b/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py new file mode 100644 index 0000000000000000000000000000000000000000..ffb877cfd0f66b463ea1994cfb1c944431147140 --- /dev/null +++ b/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py @@ -0,0 +1,960 @@ +# Copyright 2025 TencentARC and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from dataclasses import dataclass +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin +from ...models import AutoencoderKL, MultiAdapter, T2IAdapter, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + BaseOutput, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +@dataclass +class StableDiffusionAdapterPipelineOutput(BaseOutput): + """ + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + list of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + nsfw_content_detected (`list[bool]`) + list of flags denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, or `None` if safety checking could not be performed. + """ + + images: list[PIL.Image.Image] | np.ndarray + nsfw_content_detected: list[bool] | None + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> from PIL import Image + >>> from diffusers.utils import load_image + >>> import torch + >>> from diffusers import StableDiffusionAdapterPipeline, T2IAdapter + + >>> image = load_image( + ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/t2i-adapter/color_ref.png" + ... ) + + >>> color_palette = image.resize((8, 8)) + >>> color_palette = color_palette.resize((512, 512), resample=Image.Resampling.NEAREST) + + >>> adapter = T2IAdapter.from_pretrained("TencentARC/t2iadapter_color_sd14v1", torch_dtype=torch.float16) + >>> pipe = StableDiffusionAdapterPipeline.from_pretrained( + ... "CompVis/stable-diffusion-v1-4", + ... adapter=adapter, + ... torch_dtype=torch.float16, + ... ) + + >>> pipe.to("cuda") + + >>> out_image = pipe( + ... "At night, glowing cubes in front of the beach", + ... image=color_palette, + ... ).images[0] + ``` +""" + + +def _preprocess_adapter_image(image, height, width): + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image] + image = [ + i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image + ] # expand [h, w] or [h, w, c] to [b, h, w, c] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + if image[0].ndim == 3: + image = torch.stack(image, dim=0) + elif image[0].ndim == 4: + image = torch.cat(image, dim=0) + else: + raise ValueError( + f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but receive: {image[0].ndim}" + ) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionAdapterPipeline(DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter + https://huggingface.co/papers/2302.08453 + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + Args: + adapter ([`T2IAdapter`] or [`MultiAdapter`] or `list[T2IAdapter]`): + Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a + list, the outputs from each Adapter are added together to create one combined additional conditioning. + adapter_weights (`list[float]`, *optional*, defaults to None): + list of floats representing the weight which will be multiply to each adapter's output before adding them + together. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->adapter->unet->vae" + _optional_components = ["safety_checker", "feature_extractor"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + adapter: T2IAdapter | MultiAdapter | list[T2IAdapter], + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, + ): + super().__init__() + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + if isinstance(adapter, (list, tuple)): + adapter = MultiAdapter(adapter) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + adapter=adapter, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: list[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: process multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + image, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if isinstance(self.adapter, MultiAdapter): + if not isinstance(image, list): + raise ValueError( + "MultiAdapter is enabled, but `image` is not a list. Please pass a list of images to `image`." + ) + + if len(image) != len(self.adapter.adapters): + raise ValueError( + f"MultiAdapter requires passing the same number of images as adapters. Given {len(image)} images and {len(self.adapter.adapters)} adapters." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _default_height_width(self, height, width, image): + # NOTE: It is possible that a list of images have different + # dimensions for each image, so just checking the first image + # is not _exactly_ correct, but it is simple. + while isinstance(image, list): + image = image[0] + + if height is None: + if isinstance(image, PIL.Image.Image): + height = image.height + elif isinstance(image, torch.Tensor): + height = image.shape[-2] + + # round down to nearest multiple of `self.adapter.downscale_factor` + height = (height // self.adapter.downscale_factor) * self.adapter.downscale_factor + + if width is None: + if isinstance(image, PIL.Image.Image): + width = image.width + elif isinstance(image, torch.Tensor): + width = image.shape[-1] + + # round down to nearest multiple of `self.adapter.downscale_factor` + width = (width // self.adapter.downscale_factor) * self.adapter.downscale_factor + + return height, width + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: torch.Tensor | PIL.Image.Image | list[PIL.Image.Image] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + guidance_scale: float = 7.5, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + adapter_conditioning_scale: float | list[float] = 1.0, + clip_skip: int | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor`, `PIL.Image.Image`, `list[torch.Tensor]` or `list[PIL.Image.Image]` or `list[list[PIL.Image.Image]]`): + The Adapter input condition. Adapter uses this input condition to generate guidance to Unet. If the + type is specified as `torch.Tensor`, it is passed to Adapter as is. PIL.Image.Image` can also be + accepted as an image. The control image is automatically resized to fit the output image. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead. + Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + adapter_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the adapter are multiplied by `adapter_conditioning_scale` before they are added to the + residual in the original unet. If multiple adapters are specified in init, you can set the + corresponding scale as a list. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] if `return_dict` is True, otherwise a + `tuple. When returning a tuple, the first element is a list with the generated images, and the second + element is a list of `bool`s denoting whether the corresponding generated image likely represents + "not-safe-for-work" (nsfw) content, according to the `safety_checker`. + """ + # 0. Default height and width to unet + height, width = self._default_height_width(height, width, image) + device = self._execution_device + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, height, width, callback_steps, image, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + self._guidance_scale = guidance_scale + + if isinstance(self.adapter, MultiAdapter): + adapter_input = [] + + for one_image in image: + one_image = _preprocess_adapter_image(one_image, height, width) + one_image = one_image.to(device=device, dtype=self.adapter.dtype) + adapter_input.append(one_image) + else: + adapter_input = _preprocess_adapter_image(image, height, width) + adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + clip_skip=clip_skip, + ) + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.5 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + if isinstance(self.adapter, MultiAdapter): + adapter_state = self.adapter(adapter_input, adapter_conditioning_scale) + for k, v in enumerate(adapter_state): + adapter_state[k] = v + else: + adapter_state = self.adapter(adapter_input) + for k, v in enumerate(adapter_state): + adapter_state[k] = v * adapter_conditioning_scale + if num_images_per_prompt > 1: + for k, v in enumerate(adapter_state): + adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1) + if self.do_classifier_free_guidance: + for k, v in enumerate(adapter_state): + adapter_state[k] = torch.cat([v] * 2, dim=0) + + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + down_intrablock_additional_residuals=[state.clone() for state in adapter_state], + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + has_nsfw_concept = None + elif output_type == "pil": + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # 10. Convert to PIL + image = self.numpy_to_pil(image) + else: + # 8. Post-processing + image = self.decode_latents(latents) + + # 9. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionAdapterPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py b/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py new file mode 100644 index 0000000000000000000000000000000000000000..a6dd07847de2df5e6a1e325ac512b02af44875b2 --- /dev/null +++ b/diffusers/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py @@ -0,0 +1,1303 @@ +# Copyright 2025 TencentARC and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import numpy as np +import PIL.Image +import torch +from transformers import ( + CLIPImageProcessor, + CLIPTextModel, + CLIPTextModelWithProjection, + CLIPTokenizer, + CLIPVisionModelWithProjection, +) + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import ( + FromSingleFileMixin, + IPAdapterMixin, + StableDiffusionXLLoraLoaderMixin, + TextualInversionLoaderMixin, +) +from ...models import AutoencoderKL, ImageProjection, MultiAdapter, T2IAdapter, UNet2DConditionModel +from ...models.lora import adjust_lora_scale_text_encoder +from ...schedulers import KarrasDiffusionSchedulers +from ...utils import ( + PIL_INTERPOLATION, + USE_PEFT_BACKEND, + deprecate, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin +from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import T2IAdapter, StableDiffusionXLAdapterPipeline, DDPMScheduler + >>> from diffusers.utils import load_image + + >>> sketch_image = load_image("https://huggingface.co/Adapter/t2iadapter/resolve/main/sketch.png").convert("L") + + >>> model_id = "stabilityai/stable-diffusion-xl-base-1.0" + + >>> adapter = T2IAdapter.from_pretrained( + ... "Adapter/t2iadapter", + ... subfolder="sketch_sdxl_1.0", + ... torch_dtype=torch.float16, + ... adapter_type="full_adapter_xl", + ... ) + >>> scheduler = DDPMScheduler.from_pretrained(model_id, subfolder="scheduler") + + >>> pipe = StableDiffusionXLAdapterPipeline.from_pretrained( + ... model_id, adapter=adapter, torch_dtype=torch.float16, variant="fp16", scheduler=scheduler + ... ).to("cuda") + + >>> generator = torch.manual_seed(42) + >>> sketch_image_out = pipe( + ... prompt="a photo of a dog in real world, high quality", + ... negative_prompt="extra digit, fewer digits, cropped, worst quality, low quality", + ... image=sketch_image, + ... generator=generator, + ... guidance_scale=7.5, + ... ).images[0] + ``` +""" + + +def _preprocess_adapter_image(image, height, width): + if isinstance(image, torch.Tensor): + return image + elif isinstance(image, PIL.Image.Image): + image = [image] + + if isinstance(image[0], PIL.Image.Image): + image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image] + image = [ + i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image + ] # expand [h, w] or [h, w, c] to [b, h, w, c] + image = np.concatenate(image, axis=0) + image = np.array(image).astype(np.float32) / 255.0 + image = image.transpose(0, 3, 1, 2) + image = torch.from_numpy(image) + elif isinstance(image[0], torch.Tensor): + if image[0].ndim == 3: + image = torch.stack(image, dim=0) + elif image[0].ndim == 4: + image = torch.cat(image, dim=0) + else: + raise ValueError( + f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but receive: {image[0].ndim}" + ) + return image + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + r""" + Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on + Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). + + Args: + noise_cfg (`torch.Tensor`): + The predicted noise tensor for the guided diffusion process. + noise_pred_text (`torch.Tensor`): + The predicted noise tensor for the text-guided diffusion process. + guidance_rescale (`float`, *optional*, defaults to 0.0): + A rescale factor applied to the noise predictions. + + Returns: + noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor. + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionXLAdapterPipeline( + DiffusionPipeline, + StableDiffusionMixin, + TextualInversionLoaderMixin, + StableDiffusionXLLoraLoaderMixin, + IPAdapterMixin, + FromSingleFileMixin, +): + r""" + Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter + https://huggingface.co/papers/2302.08453 + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + adapter ([`T2IAdapter`] or [`MultiAdapter`] or `list[T2IAdapter]`): + Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a + list, the outputs from each Adapter are added together to create one combined additional conditioning. + adapter_weights (`list[float]`, *optional*, defaults to None): + list of floats representing the weight which will be multiply to each adapter's output before adding them + together. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5) for + details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" + _optional_components = [ + "tokenizer", + "tokenizer_2", + "text_encoder", + "text_encoder_2", + "feature_extractor", + "image_encoder", + ] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + adapter: T2IAdapter | MultiAdapter | list[T2IAdapter], + scheduler: KarrasDiffusionSchedulers, + force_zeros_for_empty_prompt: bool = True, + feature_extractor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModelWithProjection = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + adapter=adapter, + scheduler=scheduler, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = ( + self.unet.config.sample_size + if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size") + else 128 + ) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + def encode_prompt( + self, + prompt: str, + prompt_2: str | None = None, + device: torch.device | None = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: str | None = None, + negative_prompt_2: str | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + lora_scale: float | None = None, + clip_skip: int | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) + else: + scale_lora_layers(self.text_encoder_2, lora_scale) + + prompt = [prompt] if isinstance(prompt, str) else prompt + + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 + + # textual inversion: process multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if pooled_prompt_embeds is None and prompt_embeds[0].ndim == 2: + pooled_prompt_embeds = prompt_embeds[0] + + if clip_skip is None: + prompt_embeds = prompt_embeds.hidden_states[-2] + else: + # "2" because SDXL always indexes from the penultimate layer. + prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + # normalize str to list + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + negative_prompt_2 = ( + batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 + ) + + uncond_tokens: list[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + if negative_pooled_prompt_embeds is None and negative_prompt_embeds[0].ndim == 2: + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + if self.text_encoder_2 is not None: + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + if self.text_encoder_2 is not None: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + else: + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + if self.text_encoder is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds + def prepare_ip_adapter_image_embeds( + self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance + ): + image_embeds = [] + if do_classifier_free_guidance: + negative_image_embeds = [] + if ip_adapter_image_embeds is None: + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + + image_embeds.append(single_image_embeds[None, :]) + if do_classifier_free_guidance: + negative_image_embeds.append(single_negative_image_embeds[None, :]) + else: + for single_image_embeds in ip_adapter_image_embeds: + if do_classifier_free_guidance: + single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) + negative_image_embeds.append(single_negative_image_embeds) + image_embeds.append(single_image_embeds) + + ip_adapter_image_embeds = [] + for i, single_image_embeds in enumerate(image_embeds): + single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0) + if do_classifier_free_guidance: + single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0) + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0) + + single_image_embeds = single_image_embeds.to(device=device) + ip_adapter_image_embeds.append(single_image_embeds) + + return ip_adapter_image_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://huggingface.co/papers/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.check_inputs + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ip_adapter_image=None, + ip_adapter_image_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + if ip_adapter_image is not None and ip_adapter_image_embeds is not None: + raise ValueError( + "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." + ) + + if ip_adapter_image_embeds is not None: + if not isinstance(ip_adapter_image_embeds, list): + raise ValueError( + f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" + ) + elif ip_adapter_image_embeds[0].ndim not in [3, 4]: + raise ValueError( + f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = ( + batch_size, + num_channels_latents, + int(height) // self.vae_scale_factor, + int(width) // self.vae_scale_factor, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids + def _get_add_time_ids( + self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None + ): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + deprecate( + "upcast_vae", + "1.0.0", + "`upcast_vae` is deprecated. Please use `pipe.vae.to(torch.float32)`. For more details, please refer to: https://github.com/huggingface/diffusers/pull/12619#issue-3606633695.", + ) + self.vae.to(dtype=torch.float32) + + # Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width + def _default_height_width(self, height, width, image): + # NOTE: It is possible that a list of images have different + # dimensions for each image, so just checking the first image + # is not _exactly_ correct, but it is simple. + while isinstance(image, list): + image = image[0] + + if height is None: + if isinstance(image, PIL.Image.Image): + height = image.height + elif isinstance(image, torch.Tensor): + height = image.shape[-2] + + # round down to nearest multiple of `self.adapter.downscale_factor` + height = (height // self.adapter.downscale_factor) * self.adapter.downscale_factor + + if width is None: + if isinstance(image, PIL.Image.Image): + width = image.width + elif isinstance(image, torch.Tensor): + width = image.shape[-1] + + # round down to nearest multiple of `self.adapter.downscale_factor` + width = (width // self.adapter.downscale_factor) * self.adapter.downscale_factor + + return height, width + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding( + self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 + ) -> torch.Tensor: + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + w (`torch.Tensor`): + Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. + embedding_dim (`int`, *optional*, defaults to 512): + Dimension of the embeddings to generate. + dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): + Data type of the generated embeddings. + + Returns: + `torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`. + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://huggingface.co/papers/2205.11487 . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + prompt_2: str | list[str] | None = None, + image: PipelineImageInput = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + timesteps: list[int] = None, + sigmas: list[float] = None, + denoising_end: float | None = None, + guidance_scale: float = 5.0, + negative_prompt: str | list[str] | None = None, + negative_prompt_2: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + eta: float = 0.0, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + pooled_prompt_embeds: torch.Tensor | None = None, + negative_pooled_prompt_embeds: torch.Tensor | None = None, + ip_adapter_image: PipelineImageInput | None = None, + ip_adapter_image_embeds: list[torch.Tensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + callback: Callable[[int, int, torch.Tensor], None] | None = None, + callback_steps: int = 1, + cross_attention_kwargs: dict[str, Any] | None = None, + guidance_rescale: float = 0.0, + original_size: tuple[int, int] | None = None, + crops_coords_top_left: tuple[int, int] = (0, 0), + target_size: tuple[int, int] | None = None, + negative_original_size: tuple[int, int] | None = None, + negative_crops_coords_top_left: tuple[int, int] = (0, 0), + negative_target_size: tuple[int, int] | None = None, + adapter_conditioning_scale: float | list[float] = 1.0, + adapter_conditioning_factor: float = 1.0, + clip_skip: int | None = None, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + image (`torch.Tensor`, `PIL.Image.Image`, `list[torch.Tensor]` or `list[PIL.Image.Image]` or `list[list[PIL.Image.Image]]`): + The Adapter input condition. Adapter uses this input condition to generate guidance to Unet. If the + type is specified as `torch.Tensor`, it is passed to Adapter as is. PIL.Image.Image` can also be + accepted as an image. The control image is automatically resized to fit the output image. + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. Anything below 512 pixels won't work well for + [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) + and checkpoints that are not specifically fine-tuned on low resolutions. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://huggingface.co/papers/2010.02502. Only + applies to [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + ip_adapter_image_embeds (`list[torch.Tensor]`, *optional*): + Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of + IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should + contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not + provided, embeddings are computed from the `ip_adapter_image` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionAdapterPipelineOutput`] + instead of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.Tensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + negative_original_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a specific image resolution. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_crops_coords_top_left (`tuple[int]`, *optional*, defaults to (0, 0)): + To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's + micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + negative_target_size (`tuple[int]`, *optional*, defaults to (1024, 1024)): + To negatively condition the generation process based on a target image resolution. It should be as same + as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more + information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. + adapter_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 1.0): + The outputs of the adapter are multiplied by `adapter_conditioning_scale` before they are added to the + residual in the original unet. If multiple adapters are specified in init, you can set the + corresponding scale as a list. + adapter_conditioning_factor (`float`, *optional*, defaults to 1.0): + The fraction of timesteps for which adapter should be applied. If `adapter_conditioning_factor` is + `0.0`, adapter is not applied at all. If `adapter_conditioning_factor` is `1.0`, adapter is applied for + all timesteps. If `adapter_conditioning_factor` is `0.5`, adapter is applied for half of the timesteps. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + + height, width = self._default_height_width(height, width, image) + device = self._execution_device + + if isinstance(self.adapter, MultiAdapter): + adapter_input = [] + + for one_image in image: + one_image = _preprocess_adapter_image(one_image, height, width) + one_image = one_image.to(device=device, dtype=self.adapter.dtype) + adapter_input.append(one_image) + else: + adapter_input = _preprocess_adapter_image(image, height, width) + adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype) + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ip_adapter_image, + ip_adapter_image_embeds, + ) + + self._guidance_scale = guidance_scale + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3.1 Encode input prompt + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + clip_skip=clip_skip, + ) + + # 3.2 Encode ip_adapter_image + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, + ip_adapter_image_embeds, + device, + batch_size * num_images_per_prompt, + self.do_classifier_free_guidance, + ) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps, sigmas + ) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6.1 Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Prepare added time ids & embeddings & adapter features + if isinstance(self.adapter, MultiAdapter): + adapter_state = self.adapter(adapter_input, adapter_conditioning_scale) + for k, v in enumerate(adapter_state): + adapter_state[k] = v + else: + adapter_state = self.adapter(adapter_input) + for k, v in enumerate(adapter_state): + adapter_state[k] = v * adapter_conditioning_scale + if num_images_per_prompt > 1: + for k, v in enumerate(adapter_state): + adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1) + if self.do_classifier_free_guidance: + for k, v in enumerate(adapter_state): + adapter_state[k] = torch.cat([v] * 2, dim=0) + + add_text_embeds = pooled_prompt_embeds + if self.text_encoder_2 is None: + text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) + else: + text_encoder_projection_dim = self.text_encoder_2.config.projection_dim + + add_time_ids = self._get_add_time_ids( + original_size, + crops_coords_top_left, + target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + if negative_original_size is not None and negative_target_size is not None: + negative_add_time_ids = self._get_add_time_ids( + negative_original_size, + negative_crops_coords_top_left, + negative_target_size, + dtype=prompt_embeds.dtype, + text_encoder_projection_dim=text_encoder_projection_dim, + ) + else: + negative_add_time_ids = add_time_ids + + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + # Apply denoising_end + if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + if ip_adapter_image is not None or ip_adapter_image_embeds is not None: + added_cond_kwargs["image_embeds"] = image_embeds + + # predict the noise residual + if i < int(num_inference_steps * adapter_conditioning_factor): + down_intrablock_additional_residuals = [state.clone() for state in adapter_state] + else: + down_intrablock_additional_residuals = None + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=cross_attention_kwargs, + down_intrablock_additional_residuals=down_intrablock_additional_residuals, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://huggingface.co/papers/2305.08891 + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if XLA_AVAILABLE: + xm.mark_step() + + if not output_type == "latent": + # make sure the VAE is in float32 mode, as it overflows in float16 + needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast + + if needs_upcasting: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + + # cast back to fp16 if needed + if needs_upcasting: + self.vae.to(dtype=torch.float16) + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) diff --git a/diffusers/pipelines/transformers_loading_utils.py b/diffusers/pipelines/transformers_loading_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..02f8ee612601ff42887fbe0cc19778665c9660aa --- /dev/null +++ b/diffusers/pipelines/transformers_loading_utils.py @@ -0,0 +1,123 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import contextlib +import os +import tempfile +from typing import TYPE_CHECKING + +from huggingface_hub import DDUFEntry +from tqdm import tqdm + +from ..utils import is_safetensors_available, is_transformers_available, is_transformers_version + + +if TYPE_CHECKING: + from transformers import PreTrainedModel, PreTrainedTokenizer + +if is_transformers_available(): + from transformers import PreTrainedModel, PreTrainedTokenizer + +if is_safetensors_available(): + import safetensors.torch + + +def _load_tokenizer_from_dduf( + cls: "PreTrainedTokenizer", name: str, dduf_entries: dict[str, DDUFEntry], **kwargs +) -> "PreTrainedTokenizer": + """ + Load a tokenizer from a DDUF archive. + + In practice, `transformers` do not provide a way to load a tokenizer from a DDUF archive. This function is a + workaround by extracting the tokenizer files from the DDUF archive and loading the tokenizer from the extracted + files. There is an extra cost of extracting the files, but of limited impact as the tokenizer files are usually + small-ish. + """ + with tempfile.TemporaryDirectory() as tmp_dir: + for entry_name, entry in dduf_entries.items(): + if entry_name.startswith(name + "/"): + tmp_entry_path = os.path.join(tmp_dir, *entry_name.split("/")) + # need to create intermediary directory if they don't exist + os.makedirs(os.path.dirname(tmp_entry_path), exist_ok=True) + with open(tmp_entry_path, "wb") as f: + with entry.as_mmap() as mm: + f.write(mm) + return cls.from_pretrained(os.path.dirname(tmp_entry_path), **kwargs) + + +def _load_transformers_model_from_dduf( + cls: "PreTrainedModel", name: str, dduf_entries: dict[str, DDUFEntry], **kwargs +) -> "PreTrainedModel": + """ + Load a transformers model from a DDUF archive. + + In practice, `transformers` do not provide a way to load a model from a DDUF archive. This function is a workaround + by instantiating a model from the config file and loading the weights from the DDUF archive directly. + """ + config_file = dduf_entries.get(f"{name}/config.json") + if config_file is None: + raise EnvironmentError( + f"Could not find a config.json file for component {name} in DDUF file (contains {dduf_entries.keys()})." + ) + generation_config = dduf_entries.get(f"{name}/generation_config.json", None) + + weight_files = [ + entry + for entry_name, entry in dduf_entries.items() + if entry_name.startswith(f"{name}/") and entry_name.endswith(".safetensors") + ] + if not weight_files: + raise EnvironmentError( + f"Could not find any weight file for component {name} in DDUF file (contains {dduf_entries.keys()})." + ) + if not is_safetensors_available(): + raise EnvironmentError( + "Safetensors is not available, cannot load model from DDUF. Please `pip install safetensors`." + ) + if is_transformers_version("<", "4.47.0"): + raise ImportError( + "You need to install `transformers>4.47.0` in order to load a transformers model from a DDUF file. " + "You can install it with: `pip install --upgrade transformers`" + ) + + with tempfile.TemporaryDirectory() as tmp_dir: + from transformers import AutoConfig, GenerationConfig + + tmp_config_file = os.path.join(tmp_dir, "config.json") + with open(tmp_config_file, "w") as f: + f.write(config_file.read_text()) + config = AutoConfig.from_pretrained(tmp_config_file) + if generation_config is not None: + tmp_generation_config_file = os.path.join(tmp_dir, "generation_config.json") + with open(tmp_generation_config_file, "w") as f: + f.write(generation_config.read_text()) + generation_config = GenerationConfig.from_pretrained(tmp_generation_config_file) + state_dict = {} + with contextlib.ExitStack() as stack: + for entry in tqdm(weight_files, desc="Loading state_dict"): # Loop over safetensors files + # Memory-map the safetensors file + mmap = stack.enter_context(entry.as_mmap()) + # Load tensors from the memory-mapped file + tensors = safetensors.torch.load(mmap) + # Update the state dictionary with tensors + state_dict.update(tensors) + # `from_pretrained` sets the model to eval mode by default, which is the + # correct behavior for inference. Do not call `model.train()` here. + return cls.from_pretrained( + pretrained_model_name_or_path=None, + config=config, + generation_config=generation_config, + state_dict=state_dict, + **kwargs, + ) diff --git a/diffusers/pipelines/visualcloze/__init__.py b/diffusers/pipelines/visualcloze/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ab765a1bbad9ae0b8c7f2e7dec5019e3a5e4ea01 --- /dev/null +++ b/diffusers/pipelines/visualcloze/__init__.py @@ -0,0 +1,52 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_visualcloze_combined"] = ["VisualClozePipeline"] + _import_structure["pipeline_visualcloze_generation"] = ["VisualClozeGenerationPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_visualcloze_combined import VisualClozePipeline + from .pipeline_visualcloze_generation import VisualClozeGenerationPipeline + + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py b/diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py new file mode 100644 index 0000000000000000000000000000000000000000..f640fddc2bc5cb12b545b0754b18ead903bc963e --- /dev/null +++ b/diffusers/pipelines/visualcloze/pipeline_visualcloze_combined.py @@ -0,0 +1,440 @@ +# Copyright 2025 VisualCloze team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Any, Callable + +import torch +from PIL import Image +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ..flux.pipeline_flux_fill import FluxFillPipeline as VisualClozeUpsamplingPipeline +from ..flux.pipeline_output import FluxPipelineOutput +from ..pipeline_utils import DiffusionPipeline +from .pipeline_visualcloze_generation import VisualClozeGenerationPipeline + + +if is_torch_xla_available(): + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import VisualClozePipeline + >>> from diffusers.utils import load_image + + >>> image_paths = [ + ... # in-context examples + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_mask.jpg" + ... ), + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_image.jpg" + ... ), + ... ], + ... # query with the target image + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_query_mask.jpg" + ... ), + ... None, # No image needed for the target image + ... ], + ... ] + >>> task_prompt = "In each row, a logical task is demonstrated to achieve [IMAGE2] an aesthetically pleasing photograph based on [IMAGE1] sam 2-generated masks with rich color coding." + >>> content_prompt = "Majestic photo of a golden eagle perched on a rocky outcrop in a mountainous landscape. The eagle is positioned in the right foreground, facing left, with its sharp beak and keen eyes prominently visible. Its plumage is a mix of dark brown and golden hues, with intricate feather details. The background features a soft-focus view of snow-capped mountains under a cloudy sky, creating a serene and grandiose atmosphere. The foreground includes rugged rocks and patches of green moss. Photorealistic, medium depth of field, soft natural lighting, cool color palette, high contrast, sharp focus on the eagle, blurred background, tranquil, majestic, wildlife photography." + >>> pipe = VisualClozePipeline.from_pretrained( + ... "VisualCloze/VisualClozePipeline-384", resolution=384, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe( + ... task_prompt=task_prompt, + ... content_prompt=content_prompt, + ... image=image_paths, + ... upsampling_width=1344, + ... upsampling_height=768, + ... upsampling_strength=0.4, + ... guidance_scale=30, + ... num_inference_steps=30, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0][0] + >>> image.save("visualcloze.png") + ``` +""" + + +class VisualClozePipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The VisualCloze pipeline for image generation with visual context. Reference: + https://github.com/lzyhha/VisualCloze/tree/main. This pipeline is designed to generate images based on visual + in-context examples. + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + resolution (`int`, *optional*, defaults to 384): + The resolution of each image when concatenating images from the query and in-context examples. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + resolution: int = 384, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + + self.generation_pipe = VisualClozeGenerationPipeline( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + resolution=resolution, + ) + self.upsampling_pipe = VisualClozeUpsamplingPipeline( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + + def check_inputs( + self, + image, + task_prompt, + content_prompt, + upsampling_height, + upsampling_width, + strength, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") + + if upsampling_height is not None and upsampling_height % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`upsampling_height`has to be divisible by {self.vae_scale_factor * 2} but are {upsampling_height}. Dimensions will be resized accordingly" + ) + if upsampling_width is not None and upsampling_width % (self.vae_scale_factor * 2) != 0: + logger.warning( + f"`upsampling_width` have to be divisible by {self.vae_scale_factor * 2} but are {upsampling_width}. Dimensions will be resized accordingly" + ) + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Validate prompt inputs + if (task_prompt is not None or content_prompt is not None) and prompt_embeds is not None: + raise ValueError("Cannot provide both text `task_prompt` + `content_prompt` and `prompt_embeds`. ") + + if task_prompt is None and content_prompt is None and prompt_embeds is None: + raise ValueError("Must provide either `task_prompt` + `content_prompt` or pre-computed `prompt_embeds`. ") + + # Validate prompt types and consistency + if task_prompt is None: + raise ValueError("`task_prompt` is missing.") + + if task_prompt is not None and not isinstance(task_prompt, (str, list)): + raise ValueError(f"`task_prompt` must be str or list, got {type(task_prompt)}") + + if content_prompt is not None and not isinstance(content_prompt, (str, list)): + raise ValueError(f"`content_prompt` must be str or list, got {type(content_prompt)}") + + if isinstance(task_prompt, list) or isinstance(content_prompt, list): + if not isinstance(task_prompt, list) or not isinstance(content_prompt, list): + raise ValueError( + f"`task_prompt` and `content_prompt` must both be lists, or both be of type str or None, " + f"got {type(task_prompt)} and {type(content_prompt)}" + ) + if len(content_prompt) != len(task_prompt): + raise ValueError("`task_prompt` and `content_prompt` must have the same length whe they are lists.") + + for sample in image: + if not isinstance(sample, list) or not isinstance(sample[0], list): + raise ValueError("Each sample in the batch must have a 2D list of images.") + if len({len(row) for row in sample}) != 1: + raise ValueError("Each in-context example and query should contain the same number of images.") + if not any(img is None for img in sample[-1]): + raise ValueError("There are no targets in the query, which should be represented as None.") + for row in sample[:-1]: + if any(img is None for img in row): + raise ValueError("Images are missing in in-context examples.") + + # Validate embeddings + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + # Validate sequence length + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"max_sequence_length cannot exceed 512, got {max_sequence_length}") + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + task_prompt: str | list[str] = None, + content_prompt: str | list[str] = None, + image: torch.FloatTensor | None = None, + upsampling_height: int | None = None, + upsampling_width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 30.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + upsampling_strength: float = 1.0, + ): + r""" + Function invoked when calling the VisualCloze pipeline for generation. + + Args: + task_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the task intention. + content_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the content or caption of the target image to be generated. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + upsampling_height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image (i.e., output image) after upsampling via SDEdit. By + default, the image is upsampled by a factor of three, and the base resolution is determined by the + resolution parameter of the pipeline. When only one of `upsampling_height` or `upsampling_width` is + specified, the other will be automatically set based on the aspect ratio. + upsampling_width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image (i.e., output image) after upsampling via SDEdit. By + default, the image is upsampled by a factor of three, and the base resolution is determined by the + resolution parameter of the pipeline. When only one of `upsampling_height` or `upsampling_width` is + specified, the other will be automatically set based on the aspect ratio. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 30.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + upsampling_strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image` when upsampling the results. Must be between 0 and + 1. The generated image is used as a starting point and more noise is added the higher the + `upsampling_strength`. The number of denoising steps depends on the amount of noise initially added. + When `upsampling_strength` is 1, added noise is maximum and the denoising process runs for the full + number of iterations specified in `num_inference_steps`. A value of 0 skips the upsampling step and + output the results at the resolution of `self.resolution`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + generation_output = self.generation_pipe( + task_prompt=task_prompt, + content_prompt=content_prompt, + image=image, + num_inference_steps=num_inference_steps, + sigmas=sigmas, + guidance_scale=guidance_scale, + num_images_per_prompt=num_images_per_prompt, + generator=generator, + latents=latents, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + joint_attention_kwargs=joint_attention_kwargs, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + output_type=output_type if upsampling_strength == 0 else "pil", + ) + if upsampling_strength == 0: + if not return_dict: + return (generation_output,) + + return FluxPipelineOutput(images=generation_output) + + # Upsampling the generated images + # 1. Prepare the input images and prompts + if not isinstance(content_prompt, (list)): + content_prompt = [content_prompt] + n_target_per_sample = [] + upsampling_image = [] + upsampling_mask = [] + upsampling_prompt = [] + upsampling_generator = generator if isinstance(generator, (torch.Generator,)) else [] + for i in range(len(generation_output.images)): + n_target_per_sample.append(len(generation_output.images[i])) + for image in generation_output.images[i]: + upsampling_image.append(image) + upsampling_mask.append(Image.new("RGB", image.size, (255, 255, 255))) + upsampling_prompt.append( + content_prompt[i % len(content_prompt)] if content_prompt[i % len(content_prompt)] else "" + ) + if not isinstance(generator, (torch.Generator,)): + upsampling_generator.append(generator[i % len(content_prompt)]) + + # 2. Apply the denosing loop + upsampling_output = self.upsampling_pipe( + prompt=upsampling_prompt, + image=upsampling_image, + mask_image=upsampling_mask, + height=upsampling_height, + width=upsampling_width, + strength=upsampling_strength, + num_inference_steps=num_inference_steps, + sigmas=sigmas, + guidance_scale=guidance_scale, + generator=upsampling_generator, + output_type=output_type, + joint_attention_kwargs=joint_attention_kwargs, + callback_on_step_end=callback_on_step_end, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + image = upsampling_output.images + + output = [] + if output_type == "pil": + # Each sample in the batch may have multiple output images. When returning as PIL images, + # these images cannot be concatenated. Therefore, for each sample, + # a list is used to represent all the output images. + output = [] + start = 0 + for n in n_target_per_sample: + output.append(image[start : start + n]) + start += n + else: + output = image + + if not return_dict: + return (output,) + + return FluxPipelineOutput(images=output) diff --git a/diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py b/diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py new file mode 100644 index 0000000000000000000000000000000000000000..17aa366ebacd726603c6122e895984847fcc060d --- /dev/null +++ b/diffusers/pipelines/visualcloze/pipeline_visualcloze_generation.py @@ -0,0 +1,923 @@ +# Copyright 2025 VisualCloze team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Any, Callable + +import numpy as np +import torch +from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast + +from ...loaders import FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import FluxTransformer2DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import ( + USE_PEFT_BACKEND, + is_torch_xla_available, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from ...utils.torch_utils import randn_tensor +from ..flux.pipeline_flux_fill import calculate_shift, retrieve_latents, retrieve_timesteps +from ..flux.pipeline_output import FluxPipelineOutput +from ..pipeline_utils import DiffusionPipeline +from .visualcloze_utils import VisualClozeProcessor + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers import VisualClozeGenerationPipeline, FluxFillPipeline as VisualClozeUpsamplingPipeline + >>> from diffusers.utils import load_image + >>> from PIL import Image + + >>> image_paths = [ + ... # in-context examples + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_mask.jpg" + ... ), + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_incontext-example-1_image.jpg" + ... ), + ... ], + ... # query with the target image + ... [ + ... load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/visualcloze/visualcloze_mask2image_query_mask.jpg" + ... ), + ... None, # No image needed for the target image + ... ], + ... ] + >>> task_prompt = "In each row, a logical task is demonstrated to achieve [IMAGE2] an aesthetically pleasing photograph based on [IMAGE1] sam 2-generated masks with rich color coding." + >>> content_prompt = "Majestic photo of a golden eagle perched on a rocky outcrop in a mountainous landscape. The eagle is positioned in the right foreground, facing left, with its sharp beak and keen eyes prominently visible. Its plumage is a mix of dark brown and golden hues, with intricate feather details. The background features a soft-focus view of snow-capped mountains under a cloudy sky, creating a serene and grandiose atmosphere. The foreground includes rugged rocks and patches of green moss. Photorealistic, medium depth of field, soft natural lighting, cool color palette, high contrast, sharp focus on the eagle, blurred background, tranquil, majestic, wildlife photography." + >>> pipe = VisualClozeGenerationPipeline.from_pretrained( + ... "VisualCloze/VisualClozePipeline-384", resolution=384, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = pipe( + ... task_prompt=task_prompt, + ... content_prompt=content_prompt, + ... image=image_paths, + ... guidance_scale=30, + ... num_inference_steps=30, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0][0] + + >>> # optional, upsampling the generated image + >>> pipe_upsample = VisualClozeUpsamplingPipeline.from_pipe(pipe) + >>> pipe_upsample.to("cuda") + + >>> mask_image = Image.new("RGB", image.size, (255, 255, 255)) + + >>> image = pipe_upsample( + ... image=image, + ... mask_image=mask_image, + ... prompt=content_prompt, + ... width=1344, + ... height=768, + ... strength=0.4, + ... guidance_scale=30, + ... num_inference_steps=30, + ... max_sequence_length=512, + ... generator=torch.Generator("cpu").manual_seed(0), + ... ).images[0] + + >>> image.save("visualcloze.png") + ``` +""" + + +class VisualClozeGenerationPipeline( + DiffusionPipeline, + FluxLoraLoaderMixin, + FromSingleFileMixin, + TextualInversionLoaderMixin, +): + r""" + The VisualCloze pipeline for image generation with visual context. Reference: + https://github.com/lzyhha/VisualCloze/tree/main This pipeline is designed to generate images based on visual + in-context examples. + + Args: + transformer ([`FluxTransformer2DModel`]): + Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`T5TokenizerFast`): + Second Tokenizer of class + [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast). + resolution (`int`, *optional*, defaults to 384): + The resolution of each image when concatenating images from the query and in-context examples. + """ + + model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + text_encoder_2: T5EncoderModel, + tokenizer_2: T5TokenizerFast, + transformer: FluxTransformer2DModel, + resolution: int = 384, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + transformer=transformer, + scheduler=scheduler, + ) + self.resolution = resolution + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8 + # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible + # by the patch size. So the vae scale factor is multiplied by the patch size to account for this + self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16 + self.image_processor = VisualClozeProcessor( + vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels, resolution=resolution + ) + self.tokenizer_max_length = ( + self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77 + ) + self.default_sample_size = 128 + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_images_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2) + + text_inputs = self.tokenizer_2( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_length=False, + return_overflowing_tokens=False, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because `max_sequence_length` is set to " + f" {max_sequence_length} tokens: {removed_text}" + ) + + prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0] + + dtype = self.text_encoder_2.dtype + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + + _, seq_len, _ = prompt_embeds.shape + + # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._get_clip_prompt_embeds + def _get_clip_prompt_embeds( + self, + prompt: str | list[str], + num_images_per_prompt: int = 1, + device: torch.device | None = None, + ): + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + batch_size = len(prompt) + + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer_max_length, + truncation=True, + return_overflowing_tokens=False, + return_length=False, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids): + removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_max_length} tokens: {removed_text}" + ) + prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False) + + # Use pooled output of CLIPTextModel + prompt_embeds = prompt_embeds.pooler_output + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt) + prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1) + + return prompt_embeds + + # Modified from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.encode_prompt + def encode_prompt( + self, + layout_prompt: str | list[str], + task_prompt: str | list[str], + content_prompt: str | list[str], + device: torch.device | None = None, + num_images_per_prompt: int = 1, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + lora_scale: float | None = None, + ): + r""" + + Args: + layout_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the number of in-context examples and the number of images involved in + the task. + task_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the task intention. + content_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the content or caption of the target image to be generated. + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if self.text_encoder is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder, lora_scale) + if self.text_encoder_2 is not None and USE_PEFT_BACKEND: + scale_lora_layers(self.text_encoder_2, lora_scale) + + if isinstance(layout_prompt, str): + layout_prompt = [layout_prompt] + task_prompt = [task_prompt] + content_prompt = [content_prompt] + + def _preprocess(prompt, content=False): + if prompt is not None: + return f"The last image of the last row depicts: {prompt}" if content else prompt + else: + return "" + + prompt = [ + f"{_preprocess(layout_prompt[i])} {_preprocess(task_prompt[i])} {_preprocess(content_prompt[i], content=True)}".strip() + for i in range(len(layout_prompt)) + ] + pooled_prompt_embeds = self._get_clip_prompt_embeds( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + ) + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + ) + + if self.text_encoder is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + if self.text_encoder_2 is not None: + if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder_2, lora_scale) + + dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype + text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype) + + return prompt_embeds, pooled_prompt_embeds, text_ids + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_inpaint.StableDiffusion3InpaintPipeline._encode_vae_image + def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + return image_latents + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def check_inputs( + self, + image, + task_prompt, + content_prompt, + prompt_embeds=None, + pooled_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + max_sequence_length=None, + ): + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + # Validate prompt inputs + if (task_prompt is not None or content_prompt is not None) and prompt_embeds is not None: + raise ValueError("Cannot provide both text `task_prompt` + `content_prompt` and `prompt_embeds`. ") + + if task_prompt is None and content_prompt is None and prompt_embeds is None: + raise ValueError("Must provide either `task_prompt` + `content_prompt` or pre-computed `prompt_embeds`. ") + + # Validate prompt types and consistency + if task_prompt is None: + raise ValueError("`task_prompt` is missing.") + + if task_prompt is not None and not isinstance(task_prompt, (str, list)): + raise ValueError(f"`task_prompt` must be str or list, got {type(task_prompt)}") + + if content_prompt is not None and not isinstance(content_prompt, (str, list)): + raise ValueError(f"`content_prompt` must be str or list, got {type(content_prompt)}") + + if isinstance(task_prompt, list) or isinstance(content_prompt, list): + if not isinstance(task_prompt, list) or not isinstance(content_prompt, list): + raise ValueError( + f"`task_prompt` and `content_prompt` must both be lists, or both be of type str or None, " + f"got {type(task_prompt)} and {type(content_prompt)}" + ) + if len(content_prompt) != len(task_prompt): + raise ValueError("`task_prompt` and `content_prompt` must have the same length whe they are lists.") + + for sample in image: + if not isinstance(sample, list) or not isinstance(sample[0], list): + raise ValueError("Each sample in the batch must have a 2D list of images.") + if len({len(row) for row in sample}) != 1: + raise ValueError("Each in-context example and query should contain the same number of images.") + if not any(img is None for img in sample[-1]): + raise ValueError("There are no targets in the query, which should be represented as None.") + for row in sample[:-1]: + if any(img is None for img in row): + raise ValueError("Images are missing in in-context examples.") + + # Validate embeddings + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + # Validate sequence length + if max_sequence_length is not None and max_sequence_length > 512: + raise ValueError(f"max_sequence_length cannot exceed 512, got {max_sequence_length}") + + @staticmethod + def _prepare_latent_image_ids(image, vae_scale_factor, device, dtype): + latent_image_ids = [] + + for idx, img in enumerate(image, start=1): + img = img.squeeze(0) + channels, height, width = img.shape + + num_patches_h = height // vae_scale_factor // 2 + num_patches_w = width // vae_scale_factor // 2 + + patch_ids = torch.zeros(num_patches_h, num_patches_w, 3, device=device, dtype=dtype) + patch_ids[..., 0] = idx + patch_ids[..., 1] = torch.arange(num_patches_h, device=device, dtype=dtype)[:, None] + patch_ids[..., 2] = torch.arange(num_patches_w, device=device, dtype=dtype)[None, :] + + patch_ids = patch_ids.reshape(-1, 3) + latent_image_ids.append(patch_ids) + + return torch.cat(latent_image_ids, dim=0) + + @staticmethod + # Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents + def _pack_latents(latents, batch_size, num_channels_latents, height, width): + latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) + latents = latents.permute(0, 2, 4, 1, 3, 5) + latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) + + return latents + + @staticmethod + def _unpack_latents(latents, sizes, vae_scale_factor): + batch_size, num_patches, channels = latents.shape + + start = 0 + unpacked_latents = [] + for i in range(len(sizes)): + cur_size = sizes[i] + height = cur_size[0][0] // vae_scale_factor + width = sum([size[1] for size in cur_size]) // vae_scale_factor + + end = start + (height * width) // 4 + + cur_latents = latents[:, start:end] + cur_latents = cur_latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2) + cur_latents = cur_latents.permute(0, 3, 1, 4, 2, 5) + cur_latents = cur_latents.reshape(batch_size, channels // (2 * 2), height, width) + + unpacked_latents.append(cur_latents) + + start = end + + return unpacked_latents + + def _prepare_latents(self, image, mask, gen, vae_scale_factor, device, dtype): + """Helper function to prepare latents for a single batch.""" + # Concatenate images and masks along width dimension + image = [torch.cat(img, dim=3).to(device=device, dtype=dtype) for img in image] + mask = [torch.cat(m, dim=3).to(device=device, dtype=dtype) for m in mask] + + # Generate latent image IDs + latent_image_ids = self._prepare_latent_image_ids(image, vae_scale_factor, device, dtype) + + # For initial encoding, use actual images + image_latent = [self._encode_vae_image(img, gen) for img in image] + masked_image_latent = [img.clone() for img in image_latent] + + for i in range(len(image_latent)): + # Rearrange latents and masks for patch processing + num_channels_latents, height, width = image_latent[i].shape[1:] + image_latent[i] = self._pack_latents(image_latent[i], 1, num_channels_latents, height, width) + masked_image_latent[i] = self._pack_latents(masked_image_latent[i], 1, num_channels_latents, height, width) + + # Rearrange masks for patch processing + num_channels_latents, height, width = mask[i].shape[1:] + mask[i] = mask[i].view( + 1, + num_channels_latents, + height // vae_scale_factor, + vae_scale_factor, + width // vae_scale_factor, + vae_scale_factor, + ) + mask[i] = mask[i].permute(0, 1, 3, 5, 2, 4) + mask[i] = mask[i].reshape( + 1, + num_channels_latents * (vae_scale_factor**2), + height // vae_scale_factor, + width // vae_scale_factor, + ) + mask[i] = self._pack_latents( + mask[i], + 1, + num_channels_latents * (vae_scale_factor**2), + height // vae_scale_factor, + width // vae_scale_factor, + ) + + # Concatenate along batch dimension + image_latent = torch.cat(image_latent, dim=1) + masked_image_latent = torch.cat(masked_image_latent, dim=1) + mask = torch.cat(mask, dim=1) + + return image_latent, masked_image_latent, mask, latent_image_ids + + def prepare_latents( + self, + input_image, + input_mask, + timestep, + batch_size, + dtype, + device, + generator, + vae_scale_factor, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + # Process each batch + masked_image_latents = [] + image_latents = [] + masks = [] + latent_image_ids = [] + + for i in range(len(input_image)): + _image_latent, _masked_image_latent, _mask, _latent_image_ids = self._prepare_latents( + input_image[i], + input_mask[i], + generator if isinstance(generator, torch.Generator) else generator[i], + vae_scale_factor, + device, + dtype, + ) + masked_image_latents.append(_masked_image_latent) + image_latents.append(_image_latent) + masks.append(_mask) + latent_image_ids.append(_latent_image_ids) + + # Concatenate all batches + masked_image_latents = torch.cat(masked_image_latents, dim=0) + image_latents = torch.cat(image_latents, dim=0) + masks = torch.cat(masks, dim=0) + + # Handle batch size expansion + if batch_size > masked_image_latents.shape[0]: + if batch_size % masked_image_latents.shape[0] == 0: + # Expand batches by repeating + additional_image_per_prompt = batch_size // masked_image_latents.shape[0] + masked_image_latents = torch.cat([masked_image_latents] * additional_image_per_prompt, dim=0) + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + masks = torch.cat([masks] * additional_image_per_prompt, dim=0) + else: + raise ValueError( + f"Cannot expand batch size from {masked_image_latents.shape[0]} to {batch_size}. " + "Batch sizes must be multiples of each other." + ) + + # Add noise to latents + noises = randn_tensor(image_latents.shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noises).to(dtype=dtype) + + # Combine masked latents with masks + masked_image_latents = torch.cat((masked_image_latents, masks), dim=-1).to(dtype=dtype) + + return latents, masked_image_latents, latent_image_ids[0] + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + task_prompt: str | list[str] = None, + content_prompt: str | list[str] = None, + image: torch.FloatTensor | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 30.0, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: torch.FloatTensor | None = None, + pooled_prompt_embeds: torch.FloatTensor | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the VisualCloze pipeline for generation. + + Args: + task_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the task intention. + content_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to define the content or caption of the target image to be generated. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 30.0): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict` + is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated + images. + """ + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + image, + task_prompt, + content_prompt, + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + max_sequence_length=max_sequence_length, + ) + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + + processor_output = self.image_processor.preprocess( + task_prompt, content_prompt, image, vae_scale_factor=self.vae_scale_factor + ) + + # 2. Define call parameters + if processor_output["task_prompt"] is not None and isinstance(processor_output["task_prompt"], str): + batch_size = 1 + elif processor_output["task_prompt"] is not None and isinstance(processor_output["task_prompt"], list): + batch_size = len(processor_output["task_prompt"]) + + device = self._execution_device + + # 3. Prepare prompt embeddings + lora_scale = ( + self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None + ) + prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt( + layout_prompt=processor_output["layout_prompt"], + task_prompt=processor_output["task_prompt"], + content_prompt=processor_output["content_prompt"], + prompt_embeds=prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + device=device, + num_images_per_prompt=num_images_per_prompt, + max_sequence_length=max_sequence_length, + lora_scale=lora_scale, + ) + + # 4. Prepare timesteps + # Calculate sequence length and shift factor + image_seq_len = sum( + (size[0] // self.vae_scale_factor // 2) * (size[1] // self.vae_scale_factor // 2) + for sample in processor_output["image_size"][0] + for size in sample + ) + + # Calculate noise schedule parameters + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + + # Get timesteps + sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + mu=mu, + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, 1.0, device) + + # 5. Prepare latent variables + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + latents, masked_image_latents, latent_image_ids = self.prepare_latents( + processor_output["init_image"], + processor_output["mask"], + latent_timestep, + batch_size * num_images_per_prompt, + prompt_embeds.dtype, + device, + generator, + vae_scale_factor=self.vae_scale_factor, + ) + + # Calculate warmup steps + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # Prepare guidance + if self.transformer.config.guidance_embeds: + guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32) + guidance = guidance.expand(latents.shape[0]) + else: + guidance = None + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # Broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]).to(latents.dtype) + latent_model_input = torch.cat((latents, masked_image_latents), dim=2) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep / 1000, + guidance=guidance, + pooled_projections=pooled_prompt_embeds, + encoder_hidden_states=prompt_embeds, + txt_ids=text_ids, + img_ids=latent_image_ids, + joint_attention_kwargs=self.joint_attention_kwargs, + return_dict=False, + )[0] + + # Compute the previous noisy sample x_t -> x_t-1 + latents_dtype = latents.dtype + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if latents.dtype != latents_dtype: + if torch.backends.mps.is_available(): + # Some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 + latents = latents.to(latents_dtype) + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + + # Call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + # XLA optimization + if XLA_AVAILABLE: + xm.mark_step() + + # 7. Post-process the image + # Crop the target image + # Since the generated image is a concatenation of the conditional and target regions, + # we need to extract only the target regions based on their positions + image = [] + if output_type == "latent": + image = latents + else: + for b in range(len(latents)): + cur_image_size = processor_output["image_size"][b % batch_size] + cur_target_position = processor_output["target_position"][b % batch_size] + cur_latent = self._unpack_latents(latents[b].unsqueeze(0), cur_image_size, self.vae_scale_factor)[-1] + cur_latent = (cur_latent / self.vae.config.scaling_factor) + self.vae.config.shift_factor + cur_image = self.vae.decode(cur_latent, return_dict=False)[0] + cur_image = self.image_processor.postprocess(cur_image, output_type=output_type)[0] + + start = 0 + cropped = [] + for i, size in enumerate(cur_image_size[-1]): + if cur_target_position[i]: + if output_type == "pil": + cropped.append(cur_image.crop((start, 0, start + size[1], size[0]))) + else: + cropped.append(cur_image[0 : size[0], start : start + size[1]]) + start += size[1] + image.append(cropped) + if output_type != "pil": + image = np.concatenate([arr[None] for sub_image in image for arr in sub_image], axis=0) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return FluxPipelineOutput(images=image) diff --git a/diffusers/pipelines/visualcloze/visualcloze_utils.py b/diffusers/pipelines/visualcloze/visualcloze_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..1a96ff8d2d1357440d49a15f88832802581b7b8d --- /dev/null +++ b/diffusers/pipelines/visualcloze/visualcloze_utils.py @@ -0,0 +1,250 @@ +# Copyright 2025 VisualCloze team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import torch +from PIL import Image + +from ...image_processor import VaeImageProcessor + + +class VisualClozeProcessor(VaeImageProcessor): + """ + Image processor for the VisualCloze pipeline. + + This processor handles the preprocessing of images for visual cloze tasks, including resizing, normalization, and + mask generation. + + Args: + resolution (int, optional): + Target resolution for processing images. Each image will be resized to this resolution before being + concatenated to avoid the out-of-memory error. Defaults to 384. + *args: Additional arguments passed to [~image_processor.VaeImageProcessor] + **kwargs: Additional keyword arguments passed to [~image_processor.VaeImageProcessor] + """ + + def __init__(self, *args, resolution: int = 384, **kwargs): + super().__init__(*args, **kwargs) + self.resolution = resolution + + def preprocess_image( + self, input_images: list[list[Image.Image | None]], vae_scale_factor: int + ) -> tuple[list[list[torch.Tensor]], list[list[list[int]]], list[int]]: + """ + Preprocesses input images for the VisualCloze pipeline. + + This function handles the preprocessing of input images by: + 1. Resizing and cropping images to maintain consistent dimensions + 2. Converting images to the Tensor format for the VAE + 3. Normalizing pixel values + 4. Tracking image sizes and positions of target images + + Args: + input_images (list[list[Image.Image | None]]): + A nested list of PIL Images where: + - Outer list represents different samples, including in-context examples and the query + - Inner list contains images for the task + - In the last row, condition images are provided and the target images are placed as None + vae_scale_factor (int): + The scale factor used by the VAE for resizing images + + Returns: + tuple containing: + - list[list[torch.Tensor]]: Preprocessed images in tensor format + - list[list[list[int]]]: Dimensions of each processed image [height, width] + - list[int]: Target positions indicating which images are to be generated + """ + n_samples, n_task_images = len(input_images), len(input_images[0]) + divisible = 2 * vae_scale_factor + + processed_images: list[list[Image.Image]] = [[] for _ in range(n_samples)] + resize_size: list[tuple[int, int] | None] = [None for _ in range(n_samples)] + target_position: list[int] = [] + + # Process each sample + for i in range(n_samples): + # Determine size from first non-None image + for j in range(n_task_images): + if input_images[i][j] is not None: + aspect_ratio = input_images[i][j].width / input_images[i][j].height + target_area = self.resolution * self.resolution + new_h = int((target_area / aspect_ratio) ** 0.5) + new_w = int(new_h * aspect_ratio) + + new_w = max(new_w // divisible, 1) * divisible + new_h = max(new_h // divisible, 1) * divisible + resize_size[i] = (new_w, new_h) + break + + # Process all images in the sample + for j in range(n_task_images): + if input_images[i][j] is not None: + target = self._resize_and_crop(input_images[i][j], resize_size[i][0], resize_size[i][1]) + processed_images[i].append(target) + if i == n_samples - 1: + target_position.append(0) + else: + blank = Image.new("RGB", resize_size[i] or (self.resolution, self.resolution), (0, 0, 0)) + processed_images[i].append(blank) + if i == n_samples - 1: + target_position.append(1) + + # Ensure consistent width for multiple target images when there are multiple target images + if len(target_position) > 1 and sum(target_position) > 1: + new_w = resize_size[n_samples - 1][0] or 384 + for i in range(len(processed_images)): + for j in range(len(processed_images[i])): + if processed_images[i][j] is not None: + new_h = int(processed_images[i][j].height * (new_w / processed_images[i][j].width)) + new_w = int(new_w / 16) * 16 + new_h = int(new_h / 16) * 16 + processed_images[i][j] = self._resize_and_crop(processed_images[i][j], new_h, new_w) + + # Convert to tensors and normalize + image_sizes = [] + for i in range(len(processed_images)): + image_sizes.append([[img.height, img.width] for img in processed_images[i]]) + for j, image in enumerate(processed_images[i]): + image = self.pil_to_numpy(image) + image = self.numpy_to_pt(image) + image = self.normalize(image) + processed_images[i][j] = image + + return processed_images, image_sizes, target_position + + def preprocess_mask( + self, input_images: list[list[Image.Image]], target_position: list[int] + ) -> list[list[torch.Tensor]]: + """ + Generate masks for the VisualCloze pipeline. + + Args: + input_images (list[list[Image.Image]]): + Processed images from preprocess_image + target_position (list[int]): + Binary list marking the positions of target images (1 for target, 0 for condition) + + Returns: + list[list[torch.Tensor]]: + A nested list of mask tensors (1 for target positions, 0 for condition images) + """ + mask = [] + for i, row in enumerate(input_images): + if i == len(input_images) - 1: # Query row + row_masks = [ + torch.full((1, 1, row[0].shape[2], row[0].shape[3]), fill_value=m) for m in target_position + ] + else: # In-context examples + row_masks = [ + torch.full((1, 1, row[0].shape[2], row[0].shape[3]), fill_value=0) for _ in target_position + ] + mask.append(row_masks) + return mask + + def preprocess_image_upsampling( + self, + input_images: list[list[Image.Image]], + height: int, + width: int, + ) -> tuple[list[list[Image.Image]], list[list[list[int]]]]: + """Process images for the upsampling stage in the VisualCloze pipeline. + + Args: + input_images: Input image to process + height: Target height + width: Target width + + Returns: + tuple of processed image and its size + """ + image = self.resize(input_images[0][0], height, width) + image = self.pil_to_numpy(image) # to np + image = self.numpy_to_pt(image) # to pt + image = self.normalize(image) + + input_images[0][0] = image + image_sizes = [[[height, width]]] + return input_images, image_sizes + + def preprocess_mask_upsampling(self, input_images: list[list[Image.Image]]) -> list[list[torch.Tensor]]: + return [[torch.ones((1, 1, input_images[0][0].shape[2], input_images[0][0].shape[3]))]] + + def get_layout_prompt(self, size: tuple[int, int]) -> str: + layout_instruction = ( + f"A grid layout with {size[0]} rows and {size[1]} columns, displaying {size[0] * size[1]} images arranged side by side.", + ) + return layout_instruction + + def preprocess( + self, + task_prompt: str | list[str], + content_prompt: str | list[str], + input_images: list[list[list[str | None]]] | None = None, + height: int | None = None, + width: int | None = None, + upsampling: bool = False, + vae_scale_factor: int = 16, + ) -> dict: + """Process visual cloze inputs. + + Args: + task_prompt: Task description(s) + content_prompt: Content description(s) + input_images: list of images or None for the target images + height: Optional target height for upsampling stage + width: Optional target width for upsampling stage + upsampling: Whether this is in the upsampling processing stage + + Returns: + dictionary containing processed images, masks, prompts and metadata + """ + if isinstance(task_prompt, str): + task_prompt = [task_prompt] + content_prompt = [content_prompt] + input_images = [input_images] + + output = { + "init_image": [], + "mask": [], + "task_prompt": task_prompt if not upsampling else [None for _ in range(len(task_prompt))], + "content_prompt": content_prompt, + "layout_prompt": [], + "target_position": [], + "image_size": [], + } + for i in range(len(task_prompt)): + if upsampling: + layout_prompt = None + else: + layout_prompt = self.get_layout_prompt((len(input_images[i]), len(input_images[i][0]))) + + if upsampling: + cur_processed_images, cur_image_size = self.preprocess_image_upsampling( + input_images[i], height=height, width=width + ) + cur_mask = self.preprocess_mask_upsampling(cur_processed_images) + else: + cur_processed_images, cur_image_size, cur_target_position = self.preprocess_image( + input_images[i], vae_scale_factor=vae_scale_factor + ) + cur_mask = self.preprocess_mask(cur_processed_images, cur_target_position) + + output["target_position"].append(cur_target_position) + + output["image_size"].append(cur_image_size) + output["init_image"].append(cur_processed_images) + output["mask"].append(cur_mask) + output["layout_prompt"].append(layout_prompt) + + return output diff --git a/diffusers/pipelines/wan/__init__.py b/diffusers/pipelines/wan/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ad51a52f9242052d1424a360c59cac72f2d8129e --- /dev/null +++ b/diffusers/pipelines/wan/__init__.py @@ -0,0 +1,54 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_wan"] = ["WanPipeline"] + _import_structure["pipeline_wan_animate"] = ["WanAnimatePipeline"] + _import_structure["pipeline_wan_i2v"] = ["WanImageToVideoPipeline"] + _import_structure["pipeline_wan_vace"] = ["WanVACEPipeline"] + _import_structure["pipeline_wan_video2video"] = ["WanVideoToVideoPipeline"] +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_wan import WanPipeline + from .pipeline_wan_animate import WanAnimatePipeline + from .pipeline_wan_i2v import WanImageToVideoPipeline + from .pipeline_wan_vace import WanVACEPipeline + from .pipeline_wan_video2video import WanVideoToVideoPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/wan/image_processor.py b/diffusers/pipelines/wan/image_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..fa18150fcc6e66d4d986474e4dd3f68c13ba494c --- /dev/null +++ b/diffusers/pipelines/wan/image_processor.py @@ -0,0 +1,184 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import numpy as np +import PIL.Image +import torch + +from ...configuration_utils import register_to_config +from ...image_processor import VaeImageProcessor +from ...utils import PIL_INTERPOLATION + + +class WanAnimateImageProcessor(VaeImageProcessor): + r""" + Image processor to preprocess the reference (character) image for the Wan Animate model. + + Args: + do_resize (`bool`, *optional*, defaults to `True`): + Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept + `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method. + vae_scale_factor (`int`, *optional*, defaults to `8`): + VAE (spatial) scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of + this factor. + vae_latent_channels (`int`, *optional*, defaults to `16`): + VAE latent channels. + spatial_patch_size (`tuple[int, int]`, *optional*, defaults to `(2, 2)`): + The spatial patch size used by the diffusion transformer. For Wan models, this is typically (2, 2). + resample (`str`, *optional*, defaults to `lanczos`): + Resampling filter to use when resizing the image. + do_normalize (`bool`, *optional*, defaults to `True`): + Whether to normalize the image to [-1,1]. + do_binarize (`bool`, *optional*, defaults to `False`): + Whether to binarize the image to 0/1. + do_convert_rgb (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to RGB format. + do_convert_grayscale (`bool`, *optional*, defaults to be `False`): + Whether to convert the images to grayscale format. + fill_color (`str` or `float` or `tuple[float, ...]`, *optional*, defaults to `None`): + An optional fill color when `resize_mode` is set to `"fill"`. This will fill the empty space with that + color instead of filling with data from the image. Any valid `color` argument to `PIL.Image.new` is valid; + if `None`, will default to filling with data from `image`. + """ + + @register_to_config + def __init__( + self, + do_resize: bool = True, + vae_scale_factor: int = 8, + vae_latent_channels: int = 16, + spatial_patch_size: tuple[int, int] = (2, 2), + resample: str = "lanczos", + reducing_gap: int = None, + do_normalize: bool = True, + do_binarize: bool = False, + do_convert_rgb: bool = False, + do_convert_grayscale: bool = False, + fill_color: str | float | tuple[float, ...] | None = 0, + ): + super().__init__() + if do_convert_rgb and do_convert_grayscale: + raise ValueError( + "`do_convert_rgb` and `do_convert_grayscale` can not both be set to `True`," + " if you intended to convert the image into RGB format, please set `do_convert_grayscale = False`.", + " if you intended to convert the image into grayscale format, please set `do_convert_rgb = False`", + ) + + def _resize_and_fill( + self, + image: PIL.Image.Image, + width: int, + height: int, + ) -> PIL.Image.Image: + r""" + Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center + the image within the dimensions, filling empty with data from image. + + Args: + image (`PIL.Image.Image`): + The image to resize and fill. + width (`int`): + The width to resize the image to. + height (`int`): + The height to resize the image to. + + Returns: + `PIL.Image.Image`: + The resized and filled image. + """ + + ratio = width / height + src_ratio = image.width / image.height + fill_with_image_data = self.config.fill_color is None + fill_color = self.config.fill_color or 0 + + src_w = width if ratio < src_ratio else image.width * height // image.height + src_h = height if ratio >= src_ratio else image.height * width // image.width + + resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION[self.config.resample]) + res = PIL.Image.new("RGB", (width, height), color=fill_color) + res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2)) + + if fill_with_image_data: + if ratio < src_ratio: + fill_height = height // 2 - src_h // 2 + if fill_height > 0: + res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0)) + res.paste( + resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)), + box=(0, fill_height + src_h), + ) + elif ratio > src_ratio: + fill_width = width // 2 - src_w // 2 + if fill_width > 0: + res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0)) + res.paste( + resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)), + box=(fill_width + src_w, 0), + ) + + return res + + def get_default_height_width( + self, + image: PIL.Image.Image | np.ndarray | torch.Tensor, + height: int | None = None, + width: int | None = None, + ) -> tuple[int, int]: + r""" + Returns the height and width of the image, downscaled to the next integer multiple of `vae_scale_factor`. + + Args: + image (`PIL.Image.Image | np.ndarray | torch.Tensor`): + The image input, which can be a PIL image, NumPy array, or PyTorch tensor. If it is a NumPy array, it + should have shape `[batch, height, width]` or `[batch, height, width, channels]`. If it is a PyTorch + tensor, it should have shape `[batch, channels, height, width]`. + height (`int | None`, *optional*, defaults to `None`): + The height of the preprocessed image. If `None`, the height of the `image` input will be used. + width (`int | None`, *optional*, defaults to `None`): + The width of the preprocessed image. If `None`, the width of the `image` input will be used. + + Returns: + `tuple[int, int]`: + A tuple containing the height and width, both resized to the nearest integer multiple of + `vae_scale_factor * spatial_patch_size`. + """ + + if height is None: + if isinstance(image, PIL.Image.Image): + height = image.height + elif isinstance(image, torch.Tensor): + height = image.shape[2] + else: + height = image.shape[1] + + if width is None: + if isinstance(image, PIL.Image.Image): + width = image.width + elif isinstance(image, torch.Tensor): + width = image.shape[3] + else: + width = image.shape[2] + + max_area = width * height + aspect_ratio = height / width + mod_value_h = self.config.vae_scale_factor * self.config.spatial_patch_size[0] + mod_value_w = self.config.vae_scale_factor * self.config.spatial_patch_size[1] + + # Try to preserve the aspect ratio + height = round(np.sqrt(max_area * aspect_ratio)) // mod_value_h * mod_value_h + width = round(np.sqrt(max_area / aspect_ratio)) // mod_value_w * mod_value_w + + return height, width diff --git a/diffusers/pipelines/wan/pipeline_output.py b/diffusers/pipelines/wan/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..7c7f8a4f0ad9601786439718ffbd8e446c6549fb --- /dev/null +++ b/diffusers/pipelines/wan/pipeline_output.py @@ -0,0 +1,20 @@ +from dataclasses import dataclass + +import torch + +from diffusers.utils import BaseOutput + + +@dataclass +class WanPipelineOutput(BaseOutput): + r""" + Output class for Wan pipelines. + + Args: + frames (`torch.Tensor`, `np.ndarray`, or list[list[PIL.Image.Image]]): + list of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing + denoised PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape + `(batch_size, num_frames, channels, height, width)`. + """ + + frames: torch.Tensor diff --git a/diffusers/pipelines/wan/pipeline_wan.py b/diffusers/pipelines/wan/pipeline_wan.py new file mode 100644 index 0000000000000000000000000000000000000000..be2d53f1793244324e9b9cc5d83d6c7e5ed2b54e --- /dev/null +++ b/diffusers/pipelines/wan/pipeline_wan.py @@ -0,0 +1,678 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Any, Callable + +import regex as re +import torch +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers.utils import export_to_video + >>> from diffusers import AutoencoderKLWan, WanPipeline + >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler + + >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers + >>> model_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> flow_shift = 5.0 # 5.0 for 720P, 3.0 for 480P + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe.to("cuda") + + >>> prompt = "A cat and a dog baking a cake together in a kitchen. The cat is carefully measuring flour, while the dog is stirring the batter with a wooden spoon. The kitchen is cozy, with sunlight streaming through the window." + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + + >>> output = pipe( + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=720, + ... width=1280, + ... num_frames=81, + ... guidance_scale=5.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + if is_ftfy_available(): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +class WanPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for text-to-video generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + transformer_2 ([`WanTransformer3DModel`], *optional*): + Conditional Transformer to denoise the input latents during the low-noise stage. If provided, enables + two-stage denoising where `transformer` handles high-noise stages and `transformer_2` handles low-noise + stages. If not provided, only `transformer` is used. + boundary_ratio (`float`, *optional*, defaults to `None`): + Ratio of total timesteps to use as the boundary for switching between transformers in two-stage denoising. + The actual boundary timestep is calculated as `boundary_ratio * num_train_timesteps`. When provided, + `transformer` handles timesteps >= boundary_timestep and `transformer_2` handles timesteps < + boundary_timestep. If `None`, only `transformer` is used for the entire denoising process. + """ + + model_cpu_offload_seq = "text_encoder->transformer->transformer_2->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + _optional_components = ["transformer", "transformer_2"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + transformer: WanTransformer3DModel | None = None, + transformer_2: WanTransformer3DModel | None = None, + boundary_ratio: float | None = None, + expand_timesteps: bool = False, # Wan2.2 ti2v + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + transformer_2=transformer_2, + ) + self.register_to_config(boundary_ratio=boundary_ratio) + self.register_to_config(expand_timesteps=expand_timesteps) + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + guidance_scale_2=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if self.config.boundary_ratio is None and guidance_scale_2 is not None: + raise ValueError("`guidance_scale_2` is only supported when the pipeline's `boundary_ratio` is not None.") + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 480, + width: int = 832, + num_frames: int = 81, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + guidance_scale_2: float | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, pass `prompt_embeds` instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to avoid during image generation. If not defined, pass `negative_prompt_embeds` + instead. Ignored when not using guidance (`guidance_scale` < `1`). + height (`int`, defaults to `480`): + The height in pixels of the generated image. + width (`int`, defaults to `832`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + guidance_scale_2 (`float`, *optional*, defaults to `None`): + Guidance scale for the low-noise stage transformer (`transformer_2`). If `None` and the pipeline's + `boundary_ratio` is not None, uses the same value as `guidance_scale`. Only used when `transformer_2` + and the pipeline's `boundary_ratio` are not None. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + guidance_scale_2, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + patch_size = ( + self.transformer.config.patch_size + if self.transformer is not None + else self.transformer_2.config.patch_size + ) + h_multiple_of = self.vae_scale_factor_spatial * patch_size[1] + w_multiple_of = self.vae_scale_factor_spatial * patch_size[2] + calc_height = height // h_multiple_of * h_multiple_of + calc_width = width // w_multiple_of * w_multiple_of + if height != calc_height or width != calc_width: + logger.warning( + f"`height` and `width` must be multiples of ({h_multiple_of}, {w_multiple_of}) for proper patchification. " + f"Adjusting ({height}, {width}) -> ({calc_height}, {calc_width})." + ) + height, width = calc_height, calc_width + + if self.config.boundary_ratio is not None and guidance_scale_2 is None: + guidance_scale_2 = guidance_scale + + self._guidance_scale = guidance_scale + self._guidance_scale_2 = guidance_scale_2 + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype if self.transformer is not None else self.transformer_2.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = ( + self.transformer.config.in_channels + if self.transformer is not None + else self.transformer_2.config.in_channels + ) + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + ) + + mask = torch.ones(latents.shape, dtype=torch.float32, device=device) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + + if self.config.boundary_ratio is not None: + boundary_timestep = self.config.boundary_ratio * self.scheduler.config.num_train_timesteps + else: + boundary_timestep = None + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + if boundary_timestep is None or t >= boundary_timestep: + # wan2.1 or high-noise stage in wan2.2 + current_model = self.transformer + current_guidance_scale = guidance_scale + else: + # low-noise stage in wan2.2 + current_model = self.transformer_2 + current_guidance_scale = guidance_scale_2 + + latent_model_input = latents.to(transformer_dtype) + if self.config.expand_timesteps: + # seq_len: num_latent_frames * latent_height//2 * latent_width//2 + temp_ts = (mask[0][0][:, ::2, ::2] * t).flatten() + # batch_size, seq_len + timestep = temp_ts.unsqueeze(0).expand(latents.shape[0], -1) + else: + timestep = t.expand(latents.shape[0]) + + with current_model.cache_context("cond"): + noise_pred = current_model( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with current_model.cache_context("uncond"): + noise_uncond = current_model( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + current_guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/diffusers/pipelines/wan/pipeline_wan_animate.py b/diffusers/pipelines/wan/pipeline_wan_animate.py new file mode 100644 index 0000000000000000000000000000000000000000..5806032c0142d6d04be3605a08bb665c4724b560 --- /dev/null +++ b/diffusers/pipelines/wan/pipeline_wan_animate.py @@ -0,0 +1,1202 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from copy import deepcopy +from typing import Any, Callable + +import PIL +import regex as re +import torch +import torch.nn.functional as F +from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanAnimateTransformer3DModel +from ...schedulers import UniPCMultistepScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .image_processor import WanAnimateImageProcessor +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> import numpy as np + >>> from diffusers import WanAnimatePipeline + >>> from diffusers.utils import export_to_video, load_image, load_video + + >>> model_id = "Wan-AI/Wan2.2-Animate-14B-Diffusers" + >>> pipe = WanAnimatePipeline.from_pretrained(model_id, torch_dtype=torch.bfloat16) + >>> # Optionally upcast the Wan VAE to FP32 + >>> pipe.vae.to(torch.float32) + >>> pipe.to("cuda") + + >>> # Load the reference character image + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + ... ) + + >>> # Load pose and face videos (preprocessed from reference video) + >>> # Note: Videos should be preprocessed to extract pose keypoints and face features + >>> # Refer to the Wan-Animate preprocessing documentation for details + >>> pose_video = load_video("path/to/pose_video.mp4") + >>> face_video = load_video("path/to/face_video.mp4") + + >>> # CFG is generally not used for Wan Animate + >>> prompt = ( + ... "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in " + ... "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot." + ... ) + + >>> # Animation mode: Animate the character with the motion from pose/face videos + >>> output = pipe( + ... image=image, + ... pose_video=pose_video, + ... face_video=face_video, + ... prompt=prompt, + ... height=height, + ... width=width, + ... segment_frame_length=77, # Frame length of each inference segment + ... guidance_scale=1.0, + ... num_inference_steps=20, + ... mode="animate", + ... ).frames[0] + >>> export_to_video(output, "output_animation.mp4", fps=30) + + >>> # Replacement mode: Replace a character in the background video + >>> # Requires additional background_video and mask_video inputs + >>> background_video = load_video("path/to/background_video.mp4") + >>> mask_video = load_video("path/to/mask_video.mp4") # Black areas preserved, white areas generated + >>> output = pipe( + ... image=image, + ... pose_video=pose_video, + ... face_video=face_video, + ... background_video=background_video, + ... mask_video=mask_video, + ... prompt=prompt, + ... height=height, + ... width=width, + ... segment_frame_length=77, # Frame length of each inference segment + ... guidance_scale=1.0, + ... num_inference_steps=20, + ... mode="replace", + ... ).frames[0] + >>> export_to_video(output, "output_replacement.mp4", fps=30) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanAnimatePipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for unified character animation and replacement using Wan-Animate. + + WanAnimatePipeline takes a character image, pose video, and face video as input, and generates a video in two + modes: + + 1. **Animation mode**: The model generates a video of the character image that mimics the human motion in the input + pose and face videos. The character is animated based on the provided motion controls, creating a new animated + video of the character. + + 2. **Replacement mode**: The model replaces a character in a background video with the provided character image, + using the pose and face videos for motion control. This mode requires additional `background_video` and + `mask_video` inputs. The mask video should have black regions where the original content should be preserved and + white regions where the new character should be generated. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.WanLoraLoaderMixin.load_lora_weights`] for loading LoRA weights + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + image_encoder ([`CLIPVisionModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModel), specifically + the + [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large) + variant. + transformer ([`WanAnimateTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + image_processor ([`CLIPImageProcessor`]): + Image processor for preprocessing images before encoding. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + vae: AutoencoderKLWan, + scheduler: UniPCMultistepScheduler, + image_processor: CLIPImageProcessor, + image_encoder: CLIPVisionModel, + transformer: WanAnimateTransformer3DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + image_encoder=image_encoder, + transformer=transformer, + scheduler=scheduler, + image_processor=image_processor, + ) + + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.video_processor_for_mask = VideoProcessor( + vae_scale_factor=self.vae_scale_factor_spatial, do_normalize=False, do_convert_grayscale=True + ) + # In case self.transformer is None (e.g. for some pipeline tests) + spatial_patch_size = self.transformer.config.patch_size[-2:] if self.transformer is not None else (2, 2) + self.vae_image_processor = WanAnimateImageProcessor( + vae_scale_factor=self.vae_scale_factor_spatial, + spatial_patch_size=spatial_patch_size, + resample="bilinear", + fill_color=0, + ) + self.image_processor = image_processor + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan_i2v.WanImageToVideoPipeline.encode_image + def encode_image( + self, + image: PipelineImageInput, + device: torch.device | None = None, + ): + device = device or self._execution_device + image = self.image_processor(images=image, return_tensors="pt").to(device) + image_embeds = self.image_encoder(**image, output_hidden_states=True) + return image_embeds.hidden_states[-2] + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + image, + pose_video, + face_video, + background_video, + mask_video, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + callback_on_step_end_tensor_inputs=None, + mode=None, + prev_segment_conditioning_frames=None, + ): + if image is not None and image_embeds is not None: + raise ValueError( + f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to" + " only forward one of the two." + ) + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `prompt_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}") + if pose_video is None: + raise ValueError("Provide `pose_video`. Cannot leave `pose_video` undefined.") + if face_video is None: + raise ValueError("Provide `face_video`. Cannot leave `face_video` undefined.") + if not isinstance(pose_video, list) or not isinstance(face_video, list): + raise ValueError("`pose_video` and `face_video` must be lists of PIL images.") + if len(pose_video) == 0 or len(face_video) == 0: + raise ValueError("`pose_video` and `face_video` must contain at least one frame.") + if mode == "replace" and (background_video is None or mask_video is None): + raise ValueError( + "Provide `background_video` and `mask_video`. Cannot leave both `background_video` and `mask_video`" + " undefined when mode is `replace`." + ) + if mode == "replace" and (not isinstance(background_video, list) or not isinstance(mask_video, list)): + raise ValueError("`background_video` and `mask_video` must be lists of PIL images when mode is `replace`.") + + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found" + f" {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if mode is not None and (not isinstance(mode, str) or mode not in ("animate", "replace")): + raise ValueError( + f"`mode` has to be of type `str` and in ('animate', 'replace') but its type is {type(mode)} and value is {mode}" + ) + + if prev_segment_conditioning_frames is not None and ( + not isinstance(prev_segment_conditioning_frames, int) or prev_segment_conditioning_frames not in (1, 5) + ): + raise ValueError( + f"`prev_segment_conditioning_frames` has to be of type `int` and 1 or 5 but its type is" + f" {type(prev_segment_conditioning_frames)} and value is {prev_segment_conditioning_frames}" + ) + + def get_i2v_mask( + self, + batch_size: int, + latent_t: int, + latent_h: int, + latent_w: int, + mask_len: int = 1, + mask_pixel_values: torch.Tensor | None = None, + dtype: torch.dtype | None = None, + device: str | torch.device = "cuda", + ) -> torch.Tensor: + # mask_pixel_values shape (if supplied): [B, C = 1, T, latent_h, latent_w] + if mask_pixel_values is None: + mask_lat_size = torch.zeros( + batch_size, 1, (latent_t - 1) * 4 + 1, latent_h, latent_w, dtype=dtype, device=device + ) + else: + mask_lat_size = mask_pixel_values.clone().to(device=device, dtype=dtype) + mask_lat_size[:, :, :mask_len] = 1 + first_frame_mask = mask_lat_size[:, :, 0:1] + # Repeat first frame mask self.vae_scale_factor_temporal (= 4) times in the frame dimension + first_frame_mask = torch.repeat_interleave(first_frame_mask, dim=2, repeats=self.vae_scale_factor_temporal) + mask_lat_size = torch.concat([first_frame_mask, mask_lat_size[:, :, 1:]], dim=2) + mask_lat_size = mask_lat_size.view( + batch_size, -1, self.vae_scale_factor_temporal, latent_h, latent_w + ).transpose(1, 2) # [B, C = 1, 4 * T_lat, H_lat, W_lat] --> [B, C = 4, T_lat, H_lat, W_lat] + + return mask_lat_size + + def prepare_reference_image_latents( + self, + image: torch.Tensor, + batch_size: int = 1, + sample_mode: int = "argmax", + generator: torch.Generator | list[torch.Generator] | None = None, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + ) -> torch.Tensor: + # image shape: (B, C, H, W) or (B, C, T, H, W) + dtype = dtype or self.vae.dtype + if image.ndim == 4: + # Add a singleton frame dimension after the channels dimension + image = image.unsqueeze(2) + + _, _, _, height, width = image.shape + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + # Encode image to latents using VAE + image = image.to(device=device, dtype=dtype) + if isinstance(generator, list): + # Like in prepare_latents, assume len(generator) == batch_size + ref_image_latents = [ + retrieve_latents(self.vae.encode(image), generator=g, sample_mode=sample_mode) for g in generator + ] + ref_image_latents = torch.cat(ref_image_latents) + else: + ref_image_latents = retrieve_latents(self.vae.encode(image), generator, sample_mode) + # Standardize latents in preparation for Wan VAE encode + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(ref_image_latents.device, ref_image_latents.dtype) + ) + latents_recip_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + ref_image_latents.device, ref_image_latents.dtype + ) + ref_image_latents = (ref_image_latents - latents_mean) * latents_recip_std + # Handle the case where we supply one image and one generator, but batch_size > 1 (e.g. generating multiple + # videos per prompt) + if ref_image_latents.shape[0] == 1 and batch_size > 1: + ref_image_latents = ref_image_latents.expand(batch_size, -1, -1, -1, -1) + + # Prepare I2V mask in latent space and prepend to the reference image latents along channel dim + reference_image_mask = self.get_i2v_mask(batch_size, 1, latent_height, latent_width, 1, None, dtype, device) + reference_image_latents = torch.cat([reference_image_mask, ref_image_latents], dim=1) + + return reference_image_latents + + def prepare_prev_segment_cond_latents( + self, + prev_segment_cond_video: torch.Tensor | None = None, + background_video: torch.Tensor | None = None, + mask_video: torch.Tensor | None = None, + batch_size: int = 1, + segment_frame_length: int = 77, + start_frame: int = 0, + height: int = 720, + width: int = 1280, + prev_segment_cond_frames: int = 1, + task: str = "animate", + interpolation_mode: str = "bicubic", + sample_mode: str = "argmax", + generator: torch.Generator | list[torch.Generator] | None = None, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + ) -> torch.Tensor: + # prev_segment_cond_video shape: (B, C, T, H, W) in pixel space if supplied + # background_video shape: (B, C, T, H, W) (same as prev_segment_cond_video shape) + # mask_video shape: (B, 1, T, H, W) (same as prev_segment_cond_video, but with only 1 channel) + dtype = dtype or self.vae.dtype + if prev_segment_cond_video is None: + if task == "replace": + prev_segment_cond_video = background_video[:, :, :prev_segment_cond_frames].to(dtype) + else: + cond_frames_shape = (batch_size, 3, prev_segment_cond_frames, height, width) # In pixel space + prev_segment_cond_video = torch.zeros(cond_frames_shape, dtype=dtype, device=device) + + data_batch_size, channels, _, segment_height, segment_width = prev_segment_cond_video.shape + num_latent_frames = (segment_frame_length - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + if segment_height != height or segment_width != width: + print( + f"Interpolating prev segment cond video from ({segment_width}, {segment_height}) to ({width}, {height})" + ) + # Perform a 4D (spatial) rather than a 5D (spatiotemporal) reshape, following the original code + prev_segment_cond_video = prev_segment_cond_video.transpose(1, 2).flatten(0, 1) # [B * T, C, H, W] + prev_segment_cond_video = F.interpolate( + prev_segment_cond_video, size=(height, width), mode=interpolation_mode + ) + prev_segment_cond_video = prev_segment_cond_video.unflatten(0, (batch_size, -1)).transpose(1, 2) + + # Fill the remaining part of the cond video segment with zeros (if animating) or the background video (if + # replacing). + if task == "replace": + remaining_segment = background_video[:, :, prev_segment_cond_frames:].to(dtype) + else: + remaining_segment_frames = segment_frame_length - prev_segment_cond_frames + remaining_segment = torch.zeros( + batch_size, channels, remaining_segment_frames, height, width, dtype=dtype, device=device + ) + + # Prepend the conditioning frames from the previous segment to the remaining segment video in the frame dim + prev_segment_cond_video = prev_segment_cond_video.to(dtype=dtype) + full_segment_cond_video = torch.cat([prev_segment_cond_video, remaining_segment], dim=2) + + if isinstance(generator, list): + if data_batch_size == len(generator): + prev_segment_cond_latents = [ + retrieve_latents(self.vae.encode(full_segment_cond_video[i].unsqueeze(0)), g, sample_mode) + for i, g in enumerate(generator) + ] + elif data_batch_size == 1: + # Like prepare_latents, assume len(generator) == batch_size + prev_segment_cond_latents = [ + retrieve_latents(self.vae.encode(full_segment_cond_video), g, sample_mode) for g in generator + ] + else: + raise ValueError( + f"The batch size of the prev segment video should be either {len(generator)} or 1 but is" + f" {data_batch_size}" + ) + prev_segment_cond_latents = torch.cat(prev_segment_cond_latents) + else: + prev_segment_cond_latents = retrieve_latents( + self.vae.encode(full_segment_cond_video), generator, sample_mode + ) + # Standardize latents in preparation for Wan VAE encode + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(prev_segment_cond_latents.device, prev_segment_cond_latents.dtype) + ) + latents_recip_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + prev_segment_cond_latents.device, prev_segment_cond_latents.dtype + ) + prev_segment_cond_latents = (prev_segment_cond_latents - latents_mean) * latents_recip_std + + # Prepare I2V mask + if task == "replace": + mask_video = 1 - mask_video + mask_video = mask_video.permute(0, 2, 1, 3, 4) + mask_video = mask_video.flatten(0, 1) + mask_video = F.interpolate(mask_video, size=(latent_height, latent_width), mode="nearest") + mask_pixel_values = mask_video.unflatten(0, (batch_size, -1)) + mask_pixel_values = mask_pixel_values.permute(0, 2, 1, 3, 4) # output shape: [B, C = 1, T, H_lat, W_lat] + else: + mask_pixel_values = None + prev_segment_cond_mask = self.get_i2v_mask( + batch_size, + num_latent_frames, + latent_height, + latent_width, + mask_len=prev_segment_cond_frames if start_frame > 0 else 0, + mask_pixel_values=mask_pixel_values, + dtype=dtype, + device=device, + ) + + # Prepend cond I2V mask to prev segment cond latents along channel dimension + prev_segment_cond_latents = torch.cat([prev_segment_cond_mask, prev_segment_cond_latents], dim=1) + return prev_segment_cond_latents + + def prepare_pose_latents( + self, + pose_video: torch.Tensor, + batch_size: int = 1, + sample_mode: int = "argmax", + generator: torch.Generator | list[torch.Generator] | None = None, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + ) -> torch.Tensor: + # pose_video shape: (B, C, T, H, W) + pose_video = pose_video.to(device=device, dtype=dtype if dtype is not None else self.vae.dtype) + if isinstance(generator, list): + pose_latents = [ + retrieve_latents(self.vae.encode(pose_video), generator=g, sample_mode=sample_mode) for g in generator + ] + pose_latents = torch.cat(pose_latents) + else: + pose_latents = retrieve_latents(self.vae.encode(pose_video), generator, sample_mode) + # Standardize latents in preparation for Wan VAE encode + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(pose_latents.device, pose_latents.dtype) + ) + latents_recip_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + pose_latents.device, pose_latents.dtype + ) + pose_latents = (pose_latents - latents_mean) * latents_recip_std + if pose_latents.shape[0] == 1 and batch_size > 1: + pose_latents = pose_latents.expand(batch_size, -1, -1, -1, -1) + return pose_latents + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 720, + width: int = 1280, + num_frames: int = 77, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + shape = (batch_size, num_channels_latents, num_latent_frames + 1, latent_height, latent_width) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + return latents + + def pad_video_frames(self, frames: list[Any], num_target_frames: int) -> list[Any]: + """ + Pads an array-like video `frames` to `num_target_frames` using a "reflect"-like strategy. The frame dimension + is assumed to be the first dimension. In the 1D case, we can visualize this strategy as follows: + + pad_video_frames([1, 2, 3, 4, 5], 10) -> [1, 2, 3, 4, 5, 4, 3, 2, 1, 2] + """ + idx = 0 + flip = False + target_frames = [] + while len(target_frames) < num_target_frames: + target_frames.append(deepcopy(frames[idx])) + if flip: + idx -= 1 + else: + idx += 1 + if idx == 0 or idx == len(frames) - 1: + flip = not flip + + return target_frames + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + pose_video: list[PIL.Image.Image], + face_video: list[PIL.Image.Image], + background_video: list[PIL.Image.Image] | None = None, + mask_video: list[PIL.Image.Image] | None = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 720, + width: int = 1280, + segment_frame_length: int = 77, + num_inference_steps: int = 20, + mode: str = "animate", + prev_segment_conditioning_frames: int = 1, + motion_encode_batch_size: int | None = None, + guidance_scale: float = 1.0, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + image_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int, None], PipelineCallback | MultiPipelineCallbacks] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input character image to condition the generation on. Must be an image, a list of images or a + `torch.Tensor`. + pose_video (`list[PIL.Image.Image]`): + The input pose video to condition the generation on. Must be a list of PIL images. + face_video (`list[PIL.Image.Image]`): + The input face video to condition the generation on. Must be a list of PIL images. + background_video (`list[PIL.Image.Image]`, *optional*): + When mode is `"replace"`, the input background video to condition the generation on. Must be a list of + PIL images. + mask_video (`list[PIL.Image.Image]`, *optional*): + When mode is `"replace"`, the input mask video to condition the generation on. Must be a list of PIL + images. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + mode (`str`, defaults to `"animation"`): + The mode of the generation. Choose between `"animate"` and `"replace"`. + prev_segment_conditioning_frames (`int`, defaults to `1`): + The number of frames from the previous video segment to be used for temporal guidance. Recommended to + be 1 or 5. In general, should be 4N + 1, where N is a non-negative integer. + motion_encode_batch_size (`int`, *optional*): + The batch size for batched encoding of the face video via the motion encoder. This allows trading off + inference speed for lower memory usage by setting a smaller batch size. Will default to + `self.transformer.config.motion_encoder_batch_size` if not set. + height (`int`, defaults to `720`): + The height of the generated video. + width (`int`, defaults to `1280`): + The width of the generated video. + segment_frame_length (`int`, defaults to `77`): + The number of frames in each generated video segment. The total frames of video generated will be equal + to the number of frames in `pose_video`; we will generate the video in segments until we have hit this + length. In general, should be 4N + 1, where N is a non-negative integer. + num_inference_steps (`int`, defaults to `20`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `1.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. By default, CFG is not used in Wan + Animate inference. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided, + image embeddings are generated from the `image` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + image, + pose_video, + face_video, + background_video, + mask_video, + height, + width, + prompt_embeds, + negative_prompt_embeds, + image_embeds, + callback_on_step_end_tensor_inputs, + mode, + prev_segment_conditioning_frames, + ) + + if segment_frame_length % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`segment_frame_length - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the" + f" nearest number." + ) + segment_frame_length = ( + segment_frame_length // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + ) + segment_frame_length = max(segment_frame_length, 1) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # As we generate in segments of `segment_frame_length`, set the target frame length to be the least multiple + # of the effective segment length greater than or equal to the length of `pose_video`. + cond_video_frames = len(pose_video) + effective_segment_length = segment_frame_length - prev_segment_conditioning_frames + last_segment_frames = (cond_video_frames - prev_segment_conditioning_frames) % effective_segment_length + if last_segment_frames == 0: + num_padding_frames = 0 + else: + num_padding_frames = effective_segment_length - last_segment_frames + num_target_frames = cond_video_frames + num_padding_frames + num_segments = num_target_frames // effective_segment_length + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Preprocess and encode the reference (character) image + image_height, image_width = self.video_processor.get_default_height_width(image) + if image_height != height or image_width != width: + logger.warning(f"Reshaping reference image from ({image_width}, {image_height}) to ({width}, {height})") + image_pixels = self.vae_image_processor.preprocess(image, height=height, width=width, resize_mode="fill").to( + device, dtype=torch.float32 + ) + + # Get CLIP features from the reference image + if image_embeds is None: + image_embeds = self.encode_image(image, device) + image_embeds = image_embeds.repeat(batch_size * num_videos_per_prompt, 1, 1) + image_embeds = image_embeds.to(transformer_dtype) + + # 5. Encode conditioning videos (pose, face) + pose_video = self.pad_video_frames(pose_video, num_target_frames) + face_video = self.pad_video_frames(face_video, num_target_frames) + + # TODO: also support np.ndarray input (e.g. from decord like the original implementation?) + pose_video_width, pose_video_height = pose_video[0].size + if pose_video_height != height or pose_video_width != width: + logger.warning( + f"Reshaping pose video from ({pose_video_width}, {pose_video_height}) to ({width}, {height})" + ) + pose_video = self.video_processor.preprocess_video(pose_video, height=height, width=width).to( + device, dtype=torch.float32 + ) + + face_video_width, face_video_height = face_video[0].size + expected_face_size = self.transformer.config.motion_encoder_size + if face_video_width != expected_face_size or face_video_height != expected_face_size: + logger.warning( + f"Reshaping face video from ({face_video_width}, {face_video_height}) to ({expected_face_size}," + f" {expected_face_size})" + ) + face_video = self.video_processor.preprocess_video( + face_video, height=expected_face_size, width=expected_face_size + ).to(device, dtype=torch.float32) + + if mode == "replace": + background_video = self.pad_video_frames(background_video, num_target_frames) + mask_video = self.pad_video_frames(mask_video, num_target_frames) + + background_video = self.video_processor.preprocess_video(background_video, height=height, width=width).to( + device, dtype=torch.float32 + ) + mask_video = self.video_processor_for_mask.preprocess_video(mask_video, height=height, width=width).to( + device, dtype=torch.float32 + ) + + # 6. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 7. Prepare latent variables which stay constant for all inference segments + num_channels_latents = self.vae.config.z_dim + + # Get VAE-encoded latents of the reference (character) image + reference_image_latents = self.prepare_reference_image_latents( + image_pixels, batch_size * num_videos_per_prompt, generator=generator, device=device + ) + + # 8. Loop over video inference segments + start = 0 + end = segment_frame_length # Data space frames, not latent frames + all_out_frames = [] + out_frames = None + + for _ in range(num_segments): + assert start + prev_segment_conditioning_frames < cond_video_frames + + # Sample noisy latents from prior for the current inference segment + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents=num_channels_latents, + height=height, + width=width, + num_frames=segment_frame_length, + dtype=torch.float32, + device=device, + generator=generator, + latents=latents if start == 0 else None, # Only use pre-calculated latents for first segment + ) + + pose_video_segment = pose_video[:, :, start:end] + face_video_segment = face_video[:, :, start:end] + + face_video_segment = face_video_segment.expand(batch_size * num_videos_per_prompt, -1, -1, -1, -1) + face_video_segment = face_video_segment.to(dtype=transformer_dtype) + + if start > 0: + prev_segment_cond_video = out_frames[:, :, -prev_segment_conditioning_frames:].clone().detach() + else: + prev_segment_cond_video = None + + if mode == "replace": + background_video_segment = background_video[:, :, start:end] + mask_video_segment = mask_video[:, :, start:end] + + background_video_segment = background_video_segment.expand( + batch_size * num_videos_per_prompt, -1, -1, -1, -1 + ) + mask_video_segment = mask_video_segment.expand(batch_size * num_videos_per_prompt, -1, -1, -1, -1) + else: + background_video_segment = None + mask_video_segment = None + + pose_latents = self.prepare_pose_latents( + pose_video_segment, batch_size * num_videos_per_prompt, generator=generator, device=device + ) + pose_latents = pose_latents.to(dtype=transformer_dtype) + + prev_segment_cond_latents = self.prepare_prev_segment_cond_latents( + prev_segment_cond_video, + background_video=background_video_segment, + mask_video=mask_video_segment, + batch_size=batch_size * num_videos_per_prompt, + segment_frame_length=segment_frame_length, + start_frame=start, + height=height, + width=width, + prev_segment_cond_frames=prev_segment_conditioning_frames, + task=mode, + generator=generator, + device=device, + ) + + # Concatenate the reference latents in the frame dimension + reference_latents = torch.cat([reference_image_latents, prev_segment_cond_latents], dim=2) + + # 8.1 Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + # Concatenate the reference image + prev segment conditioning in the channel dim + latent_model_input = torch.cat([latents, reference_latents], dim=1).to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + with self.transformer.cache_context("cond"): + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_image=image_embeds, + pose_hidden_states=pose_latents, + face_pixel_values=face_video_segment, + motion_encode_batch_size=motion_encode_batch_size, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + # Blank out face for unconditional guidance (set all pixels to -1) + face_pixel_values_uncond = face_video_segment * 0 - 1 + with self.transformer.cache_context("uncond"): + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_hidden_states_image=image_embeds, + pose_hidden_states=pose_latents, + face_pixel_values=face_pixel_values_uncond, + motion_encode_batch_size=motion_encode_batch_size, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + latents = latents.to(self.vae.dtype) + # Destandardize latents in preparation for Wan VAE decoding + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_recip_std = 1.0 / torch.tensor(self.vae.config.latents_std).view( + 1, self.vae.config.z_dim, 1, 1, 1 + ).to(latents.device, latents.dtype) + latents = latents / latents_recip_std + latents_mean + # Skip the first latent frame (used for conditioning) + out_frames = self.vae.decode(latents[:, :, 1:], return_dict=False)[0] + + if start > 0: + out_frames = out_frames[:, :, prev_segment_conditioning_frames:] + all_out_frames.append(out_frames) + + start += effective_segment_length + end += effective_segment_length + + # Reset scheduler timesteps / state for next denoising loop + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + self._current_timestep = None + assert start + prev_segment_conditioning_frames >= cond_video_frames + + if not output_type == "latent": + video = torch.cat(all_out_frames, dim=2)[:, :, :cond_video_frames] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/diffusers/pipelines/wan/pipeline_wan_i2v.py b/diffusers/pipelines/wan/pipeline_wan_i2v.py new file mode 100644 index 0000000000000000000000000000000000000000..8061f67ab6b979b3015fb1d9bb6327553c34ff53 --- /dev/null +++ b/diffusers/pipelines/wan/pipeline_wan_i2v.py @@ -0,0 +1,841 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Any, Callable + +import PIL +import regex as re +import torch +from transformers import AutoTokenizer, CLIPImageProcessor, CLIPVisionModel, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> import numpy as np + >>> from diffusers import AutoencoderKLWan, WanImageToVideoPipeline + >>> from diffusers.utils import export_to_video, load_image + >>> from transformers import CLIPVisionModel + + >>> # Available models: Wan-AI/Wan2.1-I2V-14B-480P-Diffusers, Wan-AI/Wan2.1-I2V-14B-720P-Diffusers + >>> model_id = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" + >>> image_encoder = CLIPVisionModel.from_pretrained( + ... model_id, subfolder="image_encoder", torch_dtype=torch.float32 + ... ) + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanImageToVideoPipeline.from_pretrained( + ... model_id, vae=vae, image_encoder=image_encoder, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> image = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/astronaut.jpg" + ... ) + >>> max_area = 480 * 832 + >>> aspect_ratio = image.height / image.width + >>> mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1] + >>> height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value + >>> width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value + >>> image = image.resize((width, height)) + >>> prompt = ( + ... "An astronaut hatching from an egg, on the surface of the moon, the darkness and depth of space realised in " + ... "the background. High quality, ultrarealistic detail and breath-taking movie-like camera shot." + ... ) + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + + >>> output = pipe( + ... image=image, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=height, + ... width=width, + ... num_frames=81, + ... guidance_scale=5.0, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanImageToVideoPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for image-to-video generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + image_encoder ([`CLIPVisionModel`]): + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPVisionModel), specifically + the + [clip-vit-huge-patch14](https://github.com/mlfoundations/open_clip/blob/main/docs/PRETRAINED.md#vit-h14-xlm-roberta-large) + variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + transformer_2 ([`WanTransformer3DModel`], *optional*): + Conditional Transformer to denoise the input latents during the low-noise stage. In two-stage denoising, + `transformer` handles high-noise stages and `transformer_2` handles low-noise stages. If not provided, only + `transformer` is used. + boundary_ratio (`float`, *optional*, defaults to `None`): + Ratio of total timesteps to use as the boundary for switching between transformers in two-stage denoising. + The actual boundary timestep is calculated as `boundary_ratio * num_train_timesteps`. When provided, + `transformer` handles timesteps >= boundary_timestep and `transformer_2` handles timesteps < + boundary_timestep. If `None`, only `transformer` is used for the entire denoising process. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->transformer->transformer_2->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + _optional_components = ["transformer", "transformer_2", "image_encoder", "image_processor"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + image_processor: CLIPImageProcessor = None, + image_encoder: CLIPVisionModel = None, + transformer: WanTransformer3DModel = None, + transformer_2: WanTransformer3DModel = None, + boundary_ratio: float | None = None, + expand_timesteps: bool = False, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + image_encoder=image_encoder, + transformer=transformer, + scheduler=scheduler, + image_processor=image_processor, + transformer_2=transformer_2, + ) + self.register_to_config(boundary_ratio=boundary_ratio, expand_timesteps=expand_timesteps) + + self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + self.image_processor = image_processor + + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 512, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + def encode_image( + self, + image: PipelineImageInput, + device: torch.device | None = None, + ): + device = device or self._execution_device + image = self.image_processor(images=image, return_tensors="pt").to(device) + image_embeds = self.image_encoder(**image, output_hidden_states=True) + return image_embeds.hidden_states[-2] + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + image_embeds=None, + callback_on_step_end_tensor_inputs=None, + guidance_scale_2=None, + ): + if image is not None and image_embeds is not None: + raise ValueError( + f"Cannot forward both `image`: {image} and `image_embeds`: {image_embeds}. Please make sure to" + " only forward one of the two." + ) + if image is None and image_embeds is None: + raise ValueError( + "Provide either `image` or `prompt_embeds`. Cannot leave both `image` and `image_embeds` undefined." + ) + if image is not None and not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image): + raise ValueError(f"`image` has to be of type `torch.Tensor` or `PIL.Image.Image` but is {type(image)}") + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if self.config.boundary_ratio is None and guidance_scale_2 is not None: + raise ValueError("`guidance_scale_2` is only supported when the pipeline's `boundary_ratio` is not None.") + + if self.config.boundary_ratio is not None and image_embeds is not None: + raise ValueError("Cannot forward `image_embeds` when the pipeline's `boundary_ratio` is not configured.") + + def prepare_latents( + self, + image: PipelineImageInput, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + last_image: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor]: + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + latent_height = height // self.vae_scale_factor_spatial + latent_width = width // self.vae_scale_factor_spatial + + shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device=device, dtype=dtype) + + image = image.unsqueeze(2) # [batch_size, channels, 1, height, width] + + if self.config.expand_timesteps: + video_condition = image + + elif last_image is None: + video_condition = torch.cat( + [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 1, height, width)], dim=2 + ) + else: + last_image = last_image.unsqueeze(2) + video_condition = torch.cat( + [image, image.new_zeros(image.shape[0], image.shape[1], num_frames - 2, height, width), last_image], + dim=2, + ) + video_condition = video_condition.to(device=device, dtype=self.vae.dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + + if isinstance(generator, list): + latent_condition = [ + retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") for _ in generator + ] + latent_condition = torch.cat(latent_condition) + else: + latent_condition = retrieve_latents(self.vae.encode(video_condition), sample_mode="argmax") + latent_condition = latent_condition.repeat(batch_size, 1, 1, 1, 1) + + latent_condition = latent_condition.to(dtype) + latent_condition = (latent_condition - latents_mean) * latents_std + + if self.config.expand_timesteps: + first_frame_mask = torch.ones( + 1, 1, num_latent_frames, latent_height, latent_width, dtype=dtype, device=device + ) + first_frame_mask[:, :, 0] = 0 + return latents, latent_condition, first_frame_mask + + mask_lat_size = torch.ones(batch_size, 1, num_frames, latent_height, latent_width) + + if last_image is None: + mask_lat_size[:, :, list(range(1, num_frames))] = 0 + else: + mask_lat_size[:, :, list(range(1, num_frames - 1))] = 0 + first_frame_mask = mask_lat_size[:, :, 0:1] + first_frame_mask = torch.repeat_interleave(first_frame_mask, dim=2, repeats=self.vae_scale_factor_temporal) + mask_lat_size = torch.concat([first_frame_mask, mask_lat_size[:, :, 1:, :]], dim=2) + mask_lat_size = mask_lat_size.view(batch_size, -1, self.vae_scale_factor_temporal, latent_height, latent_width) + mask_lat_size = mask_lat_size.transpose(1, 2) + mask_lat_size = mask_lat_size.to(latent_condition.device) + + return latents, torch.concat([mask_lat_size, latent_condition], dim=1) + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: PipelineImageInput, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 480, + width: int = 832, + num_frames: int = 81, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + guidance_scale_2: float | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + image_embeds: torch.Tensor | None = None, + last_image: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + image (`PipelineImageInput`): + The input image to condition the generation on. Must be an image, a list of images or a `torch.Tensor`. + last_image (`torch.Tensor`, *optional*): + Optional last frame to condition the generated video on. When provided, the model interpolates between + `image` (first frame) and `last_image` (last frame). + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `480`): + The height of the generated video. + width (`int`, defaults to `832`): + The width of the generated video. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + guidance_scale_2 (`float`, *optional*, defaults to `None`): + Guidance scale for the low-noise stage transformer (`transformer_2`). If `None` and the pipeline's + `boundary_ratio` is not None, uses the same value as `guidance_scale`. Only used when `transformer_2` + and the pipeline's `boundary_ratio` are not None. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `negative_prompt` input argument. + image_embeds (`torch.Tensor`, *optional*): + Pre-generated image embeddings. Can be used to easily tweak image inputs (weighting). If not provided, + image embeddings are generated from the `image` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + image, + height, + width, + prompt_embeds, + negative_prompt_embeds, + image_embeds, + callback_on_step_end_tensor_inputs, + guidance_scale_2, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + patch_size = ( + self.transformer.config.patch_size + if self.transformer is not None + else self.transformer_2.config.patch_size + ) + h_multiple_of = self.vae_scale_factor_spatial * patch_size[1] + w_multiple_of = self.vae_scale_factor_spatial * patch_size[2] + calc_height = height // h_multiple_of * h_multiple_of + calc_width = width // w_multiple_of * w_multiple_of + if height != calc_height or width != calc_width: + logger.warning( + f"`height` and `width` must be multiples of ({h_multiple_of}, {w_multiple_of}) for proper patchification. " + f"Adjusting ({height}, {width}) -> ({calc_height}, {calc_width})." + ) + height, width = calc_height, calc_width + + if self.config.boundary_ratio is not None and guidance_scale_2 is None: + guidance_scale_2 = guidance_scale + + self._guidance_scale = guidance_scale + self._guidance_scale_2 = guidance_scale_2 + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + # Encode image embedding + transformer_dtype = self.transformer.dtype if self.transformer is not None else self.transformer_2.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # only wan 2.1 i2v transformer accepts image_embeds + if self.transformer is not None and self.transformer.config.image_dim is not None: + if image_embeds is None: + if last_image is None: + image_embeds = self.encode_image(image, device) + else: + image_embeds = self.encode_image([image, last_image], device) + image_embeds = image_embeds.repeat(batch_size, 1, 1) + image_embeds = image_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + num_channels_latents = self.vae.config.z_dim + image = self.video_processor.preprocess(image, height=height, width=width).to(device, dtype=torch.float32) + if last_image is not None: + last_image = self.video_processor.preprocess(last_image, height=height, width=width).to( + device, dtype=torch.float32 + ) + + latents_outputs = self.prepare_latents( + image, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames, + torch.float32, + device, + generator, + latents, + last_image, + ) + if self.config.expand_timesteps: + # wan 2.2 5b i2v use firt_frame_mask to mask timesteps + latents, condition, first_frame_mask = latents_outputs + else: + latents, condition = latents_outputs + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + if self.config.boundary_ratio is not None: + boundary_timestep = self.config.boundary_ratio * self.scheduler.config.num_train_timesteps + else: + boundary_timestep = None + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + if boundary_timestep is None or t >= boundary_timestep: + # wan2.1 or high-noise stage in wan2.2 + current_model = self.transformer + current_guidance_scale = guidance_scale + else: + # low-noise stage in wan2.2 + current_model = self.transformer_2 + current_guidance_scale = guidance_scale_2 + + if self.config.expand_timesteps: + latent_model_input = (1 - first_frame_mask) * condition + first_frame_mask * latents + latent_model_input = latent_model_input.to(transformer_dtype) + + # seq_len: num_latent_frames * (latent_height // patch_size) * (latent_width // patch_size) + temp_ts = (first_frame_mask[0][0][:, ::2, ::2] * t).flatten() + # batch_size, seq_len + timestep = temp_ts.unsqueeze(0).expand(latents.shape[0], -1) + else: + latent_model_input = torch.cat([latents, condition], dim=1).to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + with current_model.cache_context("cond"): + noise_pred = current_model( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + encoder_hidden_states_image=image_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with current_model.cache_context("uncond"): + noise_uncond = current_model( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + encoder_hidden_states_image=image_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + current_guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if self.config.expand_timesteps: + latents = (1 - first_frame_mask) * condition + first_frame_mask * latents + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/diffusers/pipelines/wan/pipeline_wan_vace.py b/diffusers/pipelines/wan/pipeline_wan_vace.py new file mode 100644 index 0000000000000000000000000000000000000000..b0896d382d674428e08573dfaa4e9b0d424d8bdb --- /dev/null +++ b/diffusers/pipelines/wan/pipeline_wan_vace.py @@ -0,0 +1,1046 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +from typing import Any, Callable + +import PIL.Image +import regex as re +import torch +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...image_processor import PipelineImageInput +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanVACETransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> import PIL.Image + >>> from diffusers import AutoencoderKLWan, WanVACEPipeline + >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler + >>> from diffusers.utils import export_to_video, load_image + def prepare_video_and_mask(first_img: PIL.Image.Image, last_img: PIL.Image.Image, height: int, width: int, num_frames: int): + first_img = first_img.resize((width, height)) + last_img = last_img.resize((width, height)) + frames = [] + frames.append(first_img) + # Ideally, this should be 127.5 to match original code, but they perform computation on numpy arrays + # whereas we are passing PIL images. If you choose to pass numpy arrays, you can set it to 127.5 to + # match the original code. + frames.extend([PIL.Image.new("RGB", (width, height), (128, 128, 128))] * (num_frames - 2)) + frames.append(last_img) + mask_black = PIL.Image.new("L", (width, height), 0) + mask_white = PIL.Image.new("L", (width, height), 255) + mask = [mask_black, *[mask_white] * (num_frames - 2), mask_black] + return frames, mask + + >>> # Available checkpoints: Wan-AI/Wan2.1-VACE-1.3B-diffusers, Wan-AI/Wan2.1-VACE-14B-diffusers + >>> model_id = "Wan-AI/Wan2.1-VACE-1.3B-diffusers" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanVACEPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> flow_shift = 3.0 # 5.0 for 720P, 3.0 for 480P + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe.to("cuda") + + >>> prompt = "CG animation style, a small blue bird takes off from the ground, flapping its wings. The bird's feathers are delicate, with a unique pattern on its chest. The background shows a blue sky with white clouds under bright sunshine. The camera follows the bird upward, capturing its flight and the vastness of the sky from a close-up, low-angle perspective." + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + >>> first_frame = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_first_frame.png" + ... ) + >>> last_frame = load_image( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/flf2v_input_last_frame.png>>> " + ... ) + + >>> height = 512 + >>> width = 512 + >>> num_frames = 81 + >>> video, mask = prepare_video_and_mask(first_frame, last_frame, height, width, num_frames) + + >>> output = pipe( + ... video=video, + ... mask=mask, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=height, + ... width=width, + ... num_frames=num_frames, + ... num_inference_steps=30, + ... guidance_scale=5.0, + ... generator=torch.Generator().manual_seed(42), + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanVACEPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for controllable generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + transformer ([`WanVACETransformer3DModel`], *optional*): + Conditional Transformer to denoise the input latents during the high-noise stage. In two-stage denoising, + `transformer` handles high-noise stages and `transformer_2` handles low-noise stages. At least one of + `transformer` or `transformer_2` must be provided. + transformer_2 ([`WanVACETransformer3DModel`], *optional*): + Conditional Transformer to denoise the input latents during the low-noise stage. In two-stage denoising, + `transformer` handles high-noise stages and `transformer_2` handles low-noise stages. At least one of + `transformer` or `transformer_2` must be provided. + boundary_ratio (`float`, *optional*, defaults to `None`): + Ratio of total timesteps to use as the boundary for switching between transformers in two-stage denoising. + The actual boundary timestep is calculated as `boundary_ratio * num_train_timesteps`. When provided, + `transformer` handles timesteps >= boundary_timestep and `transformer_2` handles timesteps < + boundary_timestep. If `None`, only the available transformer is used for the entire denoising process. + """ + + model_cpu_offload_seq = "text_encoder->transformer->transformer_2->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + _optional_components = ["transformer", "transformer_2"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + transformer: WanVACETransformer3DModel = None, + transformer_2: WanVACETransformer3DModel = None, + boundary_ratio: float | None = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + transformer_2=transformer_2, + scheduler=scheduler, + ) + self.register_to_config(boundary_ratio=boundary_ratio) + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + video=None, + mask=None, + reference_images=None, + guidance_scale_2=None, + ): + if self.transformer is not None: + base = self.vae_scale_factor_spatial * self.transformer.config.patch_size[1] + elif self.transformer_2 is not None: + base = self.vae_scale_factor_spatial * self.transformer_2.config.patch_size[1] + else: + raise ValueError( + "`transformer` or `transformer_2` component must be set in order to run inference with this pipeline" + ) + + if height % base != 0 or width % base != 0: + raise ValueError(f"`height` and `width` have to be divisible by {base} but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + if self.config.boundary_ratio is None and guidance_scale_2 is not None: + raise ValueError("`guidance_scale_2` is only supported when the pipeline's `boundary_ratio` is not None.") + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if video is not None: + if mask is not None: + if len(video) != len(mask): + raise ValueError( + f"Length of `video` {len(video)} and `mask` {len(mask)} do not match. Please make sure that" + " they have the same length." + ) + if reference_images is not None: + is_pil_image = isinstance(reference_images, PIL.Image.Image) + is_list_of_pil_images = isinstance(reference_images, list) and all( + isinstance(ref_img, PIL.Image.Image) for ref_img in reference_images + ) + is_list_of_list_of_pil_images = isinstance(reference_images, list) and all( + isinstance(ref_img, list) and all(isinstance(ref_img_, PIL.Image.Image) for ref_img_ in ref_img) + for ref_img in reference_images + ) + if not (is_pil_image or is_list_of_pil_images or is_list_of_list_of_pil_images): + raise ValueError( + "`reference_images` has to be of type `PIL.Image.Image` or `list` of `PIL.Image.Image`, or " + "`list` of `list` of `PIL.Image.Image`, but is {type(reference_images)}" + ) + if is_list_of_list_of_pil_images and len(reference_images) != 1: + raise ValueError( + "The pipeline only supports generating one video at a time at the moment. When passing a list " + "of list of reference images, where the outer list corresponds to the batch size and the inner " + "list corresponds to list of conditioning images per video, please make sure to only pass " + "one inner list of reference images (i.e., `[[, , ...]]`" + ) + elif mask is not None: + raise ValueError("`mask` can only be passed if `video` is passed as well.") + + def preprocess_conditions( + self, + video: list[PipelineImageInput] | None = None, + mask: list[PipelineImageInput] | None = None, + reference_images: PIL.Image.Image | list[PIL.Image.Image] | list[list[PIL.Image.Image]] | None = None, + batch_size: int = 1, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + ): + if video is not None: + base = self.vae_scale_factor_spatial * ( + self.transformer.config.patch_size[1] + if self.transformer is not None + else self.transformer_2.config.patch_size[1] + ) + video_height, video_width = self.video_processor.get_default_height_width(video[0]) + + if video_height * video_width > height * width: + scale = min(width / video_width, height / video_height) + video_height, video_width = int(video_height * scale), int(video_width * scale) + + if video_height % base != 0 or video_width % base != 0: + logger.warning( + f"Video height and width should be divisible by {base}, but got {video_height} and {video_width}. " + ) + video_height = (video_height // base) * base + video_width = (video_width // base) * base + + assert video_height * video_width <= height * width + + video = self.video_processor.preprocess_video(video, video_height, video_width) + image_size = (video_height, video_width) # Use the height/width of video (with possible rescaling) + else: + video = torch.zeros(batch_size, 3, num_frames, height, width, dtype=dtype, device=device) + image_size = (height, width) # Use the height/width provider by user + + if mask is not None: + mask = self.video_processor.preprocess_video(mask, image_size[0], image_size[1]) + mask = torch.clamp((mask + 1) / 2, min=0, max=1) + else: + mask = torch.ones_like(video) + + video = video.to(dtype=dtype, device=device) + mask = mask.to(dtype=dtype, device=device) + + # Make a list of list of images where the outer list corresponds to video batch size and the inner list + # corresponds to list of conditioning images per video + if reference_images is None or isinstance(reference_images, PIL.Image.Image): + reference_images = [[reference_images] for _ in range(video.shape[0])] + elif isinstance(reference_images, (list, tuple)) and isinstance(next(iter(reference_images)), PIL.Image.Image): + reference_images = [reference_images] + elif ( + isinstance(reference_images, (list, tuple)) + and isinstance(next(iter(reference_images)), list) + and isinstance(next(iter(reference_images[0])), PIL.Image.Image) + ): + reference_images = reference_images + else: + raise ValueError( + "`reference_images` has to be of type `PIL.Image.Image` or `list` of `PIL.Image.Image`, or " + "`list` of `list` of `PIL.Image.Image`, but is {type(reference_images)}" + ) + + if video.shape[0] != len(reference_images): + raise ValueError( + f"Batch size of `video` {video.shape[0]} and length of `reference_images` {len(reference_images)} does not match." + ) + + ref_images_lengths = [len(reference_images_batch) for reference_images_batch in reference_images] + if any(l != ref_images_lengths[0] for l in ref_images_lengths): + raise ValueError( + f"All batches of `reference_images` should have the same length, but got {ref_images_lengths}. Support for this " + "may be added in the future." + ) + + reference_images_preprocessed = [] + for i, reference_images_batch in enumerate(reference_images): + preprocessed_images = [] + for j, image in enumerate(reference_images_batch): + if image is None: + continue + image = self.video_processor.preprocess(image, None, None) + img_height, img_width = image.shape[-2:] + scale = min(image_size[0] / img_height, image_size[1] / img_width) + new_height, new_width = int(img_height * scale), int(img_width * scale) + resized_image = torch.nn.functional.interpolate( + image, size=(new_height, new_width), mode="bilinear", align_corners=False + ).squeeze(0) # [C, H, W] + top = (image_size[0] - new_height) // 2 + left = (image_size[1] - new_width) // 2 + canvas = torch.ones(3, *image_size, device=device, dtype=dtype) + canvas[:, top : top + new_height, left : left + new_width] = resized_image + preprocessed_images.append(canvas) + reference_images_preprocessed.append(preprocessed_images) + + return video, mask, reference_images_preprocessed + + def prepare_video_latents( + self, + video: torch.Tensor, + mask: torch.Tensor, + reference_images: list[list[torch.Tensor]] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + device: torch.device | None = None, + ) -> torch.Tensor: + device = device or self._execution_device + + if isinstance(generator, list): + # TODO: support this + raise ValueError("Passing a list of generators is not yet supported. This may be supported in the future.") + + if reference_images is None: + # For each batch of video, we set no re + # ference image (as one or more can be passed by user) + reference_images = [[None] for _ in range(video.shape[0])] + else: + if video.shape[0] != len(reference_images): + raise ValueError( + f"Batch size of `video` {video.shape[0]} and length of `reference_images` {len(reference_images)} does not match." + ) + + if video.shape[0] != 1: + # TODO: support this + raise ValueError( + "Generating with more than one video is not yet supported. This may be supported in the future." + ) + + vae_dtype = self.vae.dtype + video = video.to(dtype=vae_dtype) + + latents_mean = torch.tensor(self.vae.config.latents_mean, device=device, dtype=torch.float32).view( + 1, self.vae.config.z_dim, 1, 1, 1 + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std, device=device, dtype=torch.float32).view( + 1, self.vae.config.z_dim, 1, 1, 1 + ) + + if mask is None: + latents = retrieve_latents(self.vae.encode(video), generator, sample_mode="argmax").unbind(0) + latents = ((latents.float() - latents_mean) * latents_std).to(vae_dtype) + else: + mask = torch.where(mask > 0.5, 1.0, 0.0).to(dtype=vae_dtype) + inactive = video * (1 - mask) + reactive = video * mask + inactive = retrieve_latents(self.vae.encode(inactive), generator, sample_mode="argmax") + reactive = retrieve_latents(self.vae.encode(reactive), generator, sample_mode="argmax") + inactive = ((inactive.float() - latents_mean) * latents_std).to(vae_dtype) + reactive = ((reactive.float() - latents_mean) * latents_std).to(vae_dtype) + latents = torch.cat([inactive, reactive], dim=1) + + latent_list = [] + for latent, reference_images_batch in zip(latents, reference_images): + for reference_image in reference_images_batch: + assert reference_image.ndim == 3 + reference_image = reference_image.to(dtype=vae_dtype) + reference_image = reference_image[None, :, None, :, :] # [1, C, 1, H, W] + reference_latent = retrieve_latents(self.vae.encode(reference_image), generator, sample_mode="argmax") + reference_latent = ((reference_latent.float() - latents_mean) * latents_std).to(vae_dtype) + reference_latent = reference_latent.squeeze(0) # [C, 1, H, W] + reference_latent = torch.cat([reference_latent, torch.zeros_like(reference_latent)], dim=0) + latent = torch.cat([reference_latent.squeeze(0), latent], dim=1) + latent_list.append(latent) + return torch.stack(latent_list) + + def prepare_masks( + self, + mask: torch.Tensor, + reference_images: list[torch.Tensor] | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + ) -> torch.Tensor: + if isinstance(generator, list): + # TODO: support this + raise ValueError("Passing a list of generators is not yet supported. This may be supported in the future.") + + if reference_images is None: + # For each batch of video, we set no reference image (as one or more can be passed by user) + reference_images = [[None] for _ in range(mask.shape[0])] + else: + if mask.shape[0] != len(reference_images): + raise ValueError( + f"Batch size of `mask` {mask.shape[0]} and length of `reference_images` {len(reference_images)} does not match." + ) + + if mask.shape[0] != 1: + # TODO: support this + raise ValueError( + "Generating with more than one video is not yet supported. This may be supported in the future." + ) + + transformer_patch_size = ( + self.transformer.config.patch_size[1] + if self.transformer is not None + else self.transformer_2.config.patch_size[1] + ) + + mask_list = [] + for mask_, reference_images_batch in zip(mask, reference_images): + num_channels, num_frames, height, width = mask_.shape + new_num_frames = (num_frames + self.vae_scale_factor_temporal - 1) // self.vae_scale_factor_temporal + new_height = height // (self.vae_scale_factor_spatial * transformer_patch_size) * transformer_patch_size + new_width = width // (self.vae_scale_factor_spatial * transformer_patch_size) * transformer_patch_size + mask_ = mask_[0, :, :, :] + mask_ = mask_.view( + num_frames, new_height, self.vae_scale_factor_spatial, new_width, self.vae_scale_factor_spatial + ) + mask_ = mask_.permute(2, 4, 0, 1, 3).flatten(0, 1) # [8x8, num_frames, new_height, new_width] + mask_ = torch.nn.functional.interpolate( + mask_.unsqueeze(0), size=(new_num_frames, new_height, new_width), mode="nearest-exact" + ).squeeze(0) + num_ref_images = len(reference_images_batch) + if num_ref_images > 0: + mask_padding = torch.zeros_like(mask_[:, :num_ref_images, :, :]) + mask_ = torch.cat([mask_padding, mask_], dim=1) + mask_list.append(mask_) + return torch.stack(mask_list) + + def prepare_latents( + self, + batch_size: int, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + num_frames: int = 81, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + ) -> torch.Tensor: + if latents is not None: + return latents.to(device=device, dtype=dtype) + + num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1 + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + int(height) // self.vae_scale_factor_spatial, + int(width) // self.vae_scale_factor_spatial, + ) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + video: list[PipelineImageInput] | None = None, + mask: list[PipelineImageInput] | None = None, + reference_images: list[PipelineImageInput] | None = None, + conditioning_scale: float | list[float] | torch.Tensor = 1.0, + height: int = 480, + width: int = 832, + num_frames: int = 81, + num_inference_steps: int = 50, + guidance_scale: float = 5.0, + guidance_scale_2: float | None = None, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + video (`list[PIL.Image.Image]`, *optional*): + The input video or videos to be used as a starting point for the generation. The video should be a list + of PIL images, a numpy array, or a torch tensor. Currently, the pipeline only supports generating one + video at a time. + mask (`list[PIL.Image.Image]`, *optional*): + The input mask defines which video regions to condition on and which to generate. Black areas in the + mask indicate conditioning regions, while white areas indicate regions for generation. The mask should + be a list of PIL images, a numpy array, or a torch tensor. Currently supports generating a single video + at a time. + reference_images (`list[PIL.Image.Image]`, *optional*): + A list of one or more reference images as extra conditioning for the generation. For example, if you + are trying to inpaint a video to change the character, you can pass reference images of the new + character here. Refer to the Diffusers [examples](https://github.com/huggingface/diffusers/pull/11582) + and original [user + guide](https://github.com/ali-vilab/VACE/blob/0897c6d055d7d9ea9e191dce763006664d9780f8/UserGuide.md) + for a full list of supported tasks and use cases. + conditioning_scale (`float`, `list[float]`, `torch.Tensor`, defaults to `1.0`): + The conditioning scale to be applied when adding the control conditioning latent stream to the + denoising latent stream in each control layer of the model. If a float is provided, it will be applied + uniformly to all layers. If a list or tensor is provided, it should have the same length as the number + of control layers in the model (`len(transformer.config.vace_layers)`). + height (`int`, defaults to `480`): + The height in pixels of the generated image. + width (`int`, defaults to `832`): + The width in pixels of the generated image. + num_frames (`int`, defaults to `81`): + The number of frames in the generated video. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + guidance_scale_2 (`float`, *optional*, defaults to `None`): + Guidance scale for the low-noise stage transformer (`transformer_2`). If `None` and the pipeline's + `boundary_ratio` is not None, uses the same value as `guidance_scale`. Only used when `transformer_2` + and the pipeline's `boundary_ratio` are not None. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + # Simplification of implementation for now + if prompt is not None and not isinstance(prompt, str): + raise ValueError("Passing a list of prompts is not yet supported. This may be supported in the future.") + if num_videos_per_prompt != 1: + raise ValueError( + "Generating multiple videos per prompt is not yet supported. This may be supported in the future." + ) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + video, + mask, + reference_images, + guidance_scale_2, + ) + + if num_frames % self.vae_scale_factor_temporal != 1: + logger.warning( + f"`num_frames - 1` has to be divisible by {self.vae_scale_factor_temporal}. Rounding to the nearest number." + ) + num_frames = num_frames // self.vae_scale_factor_temporal * self.vae_scale_factor_temporal + 1 + num_frames = max(num_frames, 1) + + if self.config.boundary_ratio is not None and guidance_scale_2 is None: + guidance_scale_2 = guidance_scale + + self._guidance_scale = guidance_scale + self._guidance_scale_2 = guidance_scale_2 + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + vae_dtype = self.vae.dtype + transformer_dtype = self.transformer.dtype if self.transformer is not None else self.transformer_2.dtype + + vace_layers = ( + self.transformer.config.vace_layers + if self.transformer is not None + else self.transformer_2.config.vace_layers + ) + if isinstance(conditioning_scale, (int, float)): + conditioning_scale = [conditioning_scale] * len(vace_layers) + if isinstance(conditioning_scale, list): + if len(conditioning_scale) != len(vace_layers): + raise ValueError( + f"Length of `conditioning_scale` {len(conditioning_scale)} does not match number of layers {len(vace_layers)}." + ) + conditioning_scale = torch.tensor(conditioning_scale) + if isinstance(conditioning_scale, torch.Tensor): + if conditioning_scale.size(0) != len(vace_layers): + raise ValueError( + f"Length of `conditioning_scale` {conditioning_scale.size(0)} does not match number of layers {len(vace_layers)}." + ) + conditioning_scale = conditioning_scale.to(device=device, dtype=transformer_dtype) + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps = self.scheduler.timesteps + + # 5. Prepare latent variables + video, mask, reference_images = self.preprocess_conditions( + video, + mask, + reference_images, + batch_size, + height, + width, + num_frames, + torch.float32, + device, + ) + num_reference_images = len(reference_images[0]) + + conditioning_latents = self.prepare_video_latents(video, mask, reference_images, generator, device) + mask = self.prepare_masks(mask, reference_images, generator) + conditioning_latents = torch.cat([conditioning_latents, mask], dim=1) + conditioning_latents = conditioning_latents.to(transformer_dtype) + + num_channels_latents = ( + self.transformer.config.in_channels + if self.transformer is not None + else self.transformer_2.config.in_channels + ) + latents = self.prepare_latents( + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + num_frames + num_reference_images * self.vae_scale_factor_temporal, + torch.float32, + device, + generator, + latents, + ) + + if conditioning_latents.shape[2] != latents.shape[2]: + logger.warning( + "The number of frames in the conditioning latents does not match the number of frames to be generated. Generation quality may be affected." + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + if self.config.boundary_ratio is not None: + boundary_timestep = self.config.boundary_ratio * self.scheduler.config.num_train_timesteps + else: + boundary_timestep = None + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + + if boundary_timestep is None or t >= boundary_timestep: + # wan2.1 or high-noise stage in wan2.2 + current_model = self.transformer + current_guidance_scale = guidance_scale + else: + # low-noise stage in wan2.2 + current_model = self.transformer_2 + current_guidance_scale = guidance_scale_2 + + latent_model_input = latents.to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + with current_model.cache_context("cond"): + noise_pred = current_model( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + control_hidden_states=conditioning_latents, + control_hidden_states_scale=conditioning_scale, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + with current_model.cache_context("uncond"): + noise_uncond = current_model( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + control_hidden_states=conditioning_latents, + control_hidden_states_scale=conditioning_scale, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + current_guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents[:, :, num_reference_images:] + latents = latents.to(vae_dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/diffusers/pipelines/wan/pipeline_wan_video2video.py b/diffusers/pipelines/wan/pipeline_wan_video2video.py new file mode 100644 index 0000000000000000000000000000000000000000..8993475a285123a39caf30619eec0e7a16ce2a42 --- /dev/null +++ b/diffusers/pipelines/wan/pipeline_wan_video2video.py @@ -0,0 +1,741 @@ +# Copyright 2025 The Wan Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import html +import inspect +from typing import Any, Callable + +import regex as re +import torch +from PIL import Image +from transformers import AutoTokenizer, UMT5EncoderModel + +from ...callbacks import MultiPipelineCallbacks, PipelineCallback +from ...loaders import WanLoraLoaderMixin +from ...models import AutoencoderKLWan, WanTransformer3DModel +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_ftfy_available, is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ...video_processor import VideoProcessor +from ..pipeline_utils import DiffusionPipeline +from .pipeline_output import WanPipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +if is_ftfy_available(): + import ftfy + + +EXAMPLE_DOC_STRING = """ + Examples: + ```python + >>> import torch + >>> from diffusers.utils import export_to_video, load_video + >>> from diffusers import AutoencoderKLWan, WanVideoToVideoPipeline + >>> from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler + + >>> # Available models: Wan-AI/Wan2.1-T2V-14B-Diffusers, Wan-AI/Wan2.1-T2V-1.3B-Diffusers + >>> model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers" + >>> vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32) + >>> pipe = WanVideoToVideoPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16) + >>> flow_shift = 3.0 # 5.0 for 720P, 3.0 for 480P + >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=flow_shift) + >>> pipe.to("cuda") + + >>> prompt = "A robot standing on a mountain top. The sun is setting in the background" + >>> negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards" + >>> video = load_video( + ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/hiker.mp4" + ... ) + >>> output = pipe( + ... video=video, + ... prompt=prompt, + ... negative_prompt=negative_prompt, + ... height=480, + ... width=720, + ... guidance_scale=5.0, + ... strength=0.7, + ... ).frames[0] + >>> export_to_video(output, "output.mp4", fps=16) + ``` +""" + + +def basic_clean(text): + text = ftfy.fix_text(text) + text = html.unescape(html.unescape(text)) + return text.strip() + + +def whitespace_clean(text): + text = re.sub(r"\s+", " ", text) + text = text.strip() + return text + + +def prompt_clean(text): + text = whitespace_clean(basic_clean(text)) + return text + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +class WanVideoToVideoPipeline(DiffusionPipeline, WanLoraLoaderMixin): + r""" + Pipeline for video-to-video generation using Wan. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + Args: + tokenizer ([`T5Tokenizer`]): + Tokenizer from [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5Tokenizer), + specifically the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + text_encoder ([`T5EncoderModel`]): + [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically + the [google/umt5-xxl](https://huggingface.co/google/umt5-xxl) variant. + transformer ([`WanTransformer3DModel`]): + Conditional Transformer to denoise the input latents. + scheduler ([`UniPCMultistepScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKLWan`]): + Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + tokenizer: AutoTokenizer, + text_encoder: UMT5EncoderModel, + transformer: WanTransformer3DModel, + vae: AutoencoderKLWan, + scheduler: FlowMatchEulerDiscreteScheduler, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + transformer=transformer, + scheduler=scheduler, + ) + + self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4 + self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8 + self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial) + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline._get_t5_prompt_embeds + def _get_t5_prompt_embeds( + self, + prompt: str | list[str] = None, + num_videos_per_prompt: int = 1, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + device = device or self._execution_device + dtype = dtype or self.text_encoder.dtype + + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt = [prompt_clean(u) for u in prompt] + batch_size = len(prompt) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + add_special_tokens=True, + return_attention_mask=True, + return_tensors="pt", + ) + text_input_ids, mask = text_inputs.input_ids, text_inputs.attention_mask + seq_lens = mask.gt(0).sum(dim=1).long() + + prompt_embeds = self.text_encoder(text_input_ids.to(device), mask.to(device)).last_hidden_state + prompt_embeds = prompt_embeds.to(dtype=dtype, device=device) + prompt_embeds = [u[:v] for u, v in zip(prompt_embeds, seq_lens)] + prompt_embeds = torch.stack( + [torch.cat([u, u.new_zeros(max_sequence_length - u.size(0), u.size(1))]) for u in prompt_embeds], dim=0 + ) + + # duplicate text embeddings for each generation per prompt, using mps friendly method + _, seq_len, _ = prompt_embeds.shape + prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1) + prompt_embeds = prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1) + + return prompt_embeds + + # Copied from diffusers.pipelines.wan.pipeline_wan.WanPipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + negative_prompt: str | list[str] | None = None, + do_classifier_free_guidance: bool = True, + num_videos_per_prompt: int = 1, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + max_sequence_length: int = 226, + device: torch.device | None = None, + dtype: torch.dtype | None = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `list[str]`, *optional*): + prompt to be encoded + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + do_classifier_free_guidance (`bool`, *optional*, defaults to `True`): + Whether to use classifier free guidance or not. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + Number of videos that should be generated per prompt. torch device to place the resulting embeddings on + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + device: (`torch.device`, *optional*): + torch device + dtype: (`torch.dtype`, *optional*): + torch dtype + """ + device = device or self._execution_device + + prompt = [prompt] if isinstance(prompt, str) else prompt + if prompt is not None: + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + prompt_embeds = self._get_t5_prompt_embeds( + prompt=prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + + negative_prompt_embeds = self._get_t5_prompt_embeds( + prompt=negative_prompt, + num_videos_per_prompt=num_videos_per_prompt, + max_sequence_length=max_sequence_length, + device=device, + dtype=dtype, + ) + + return prompt_embeds, negative_prompt_embeds + + def check_inputs( + self, + prompt, + negative_prompt, + height, + width, + video=None, + latents=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 16 != 0 or width % 16 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`: {negative_prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif negative_prompt is not None and ( + not isinstance(negative_prompt, str) and not isinstance(negative_prompt, list) + ): + raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") + + if video is not None and latents is not None: + raise ValueError("Only one of `video` or `latents` should be provided") + + def prepare_latents( + self, + video: torch.Tensor | None = None, + batch_size: int = 1, + num_channels_latents: int = 16, + height: int = 480, + width: int = 832, + dtype: torch.dtype | None = None, + device: torch.device | None = None, + generator: torch.Generator | None = None, + latents: torch.Tensor | None = None, + timestep: torch.Tensor | None = None, + ): + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + num_latent_frames = ( + (video.size(2) - 1) // self.vae_scale_factor_temporal + 1 if latents is None else latents.size(1) + ) + shape = ( + batch_size, + num_channels_latents, + num_latent_frames, + height // self.vae_scale_factor_spatial, + width // self.vae_scale_factor_spatial, + ) + + if latents is None: + init_latents = [retrieve_latents(self.vae.encode(vid.unsqueeze(0)), sample_mode="argmax") for vid in video] + + init_latents = torch.cat(init_latents, dim=0).to(dtype) + + latents_mean = ( + torch.tensor(self.vae.config.latents_mean).view(1, self.vae.config.z_dim, 1, 1, 1).to(device, dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + device, dtype + ) + + init_latents = (init_latents - latents_mean) * latents_std + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + if hasattr(self.scheduler, "add_noise"): + latents = self.scheduler.add_noise(init_latents, noise, timestep) + else: + latents = self.scheduler.scale_noise(init_latents, timestep, noise) + else: + latents = latents.to(device) + + return latents + + # Copied from diffusers.pipelines.animatediff.pipeline_animatediff_video2video.AnimateDiffVideoToVideoPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, timesteps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + + t_start = max(num_inference_steps - init_timestep, 0) + timesteps = timesteps[t_start * self.scheduler.order :] + + return timesteps, num_inference_steps - t_start + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1.0 + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def current_timestep(self): + return self._current_timestep + + @property + def interrupt(self): + return self._interrupt + + @property + def attention_kwargs(self): + return self._attention_kwargs + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + video: list[Image.Image] = None, + prompt: str | list[str] = None, + negative_prompt: str | list[str] = None, + height: int = 480, + width: int = 832, + num_inference_steps: int = 50, + timesteps: list[int] | None = None, + guidance_scale: float = 5.0, + strength: float = 0.8, + num_videos_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.Tensor | None = None, + prompt_embeds: torch.Tensor | None = None, + negative_prompt_embeds: torch.Tensor | None = None, + output_type: str | None = "np", + return_dict: bool = True, + attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | PipelineCallback | MultiPipelineCallbacks | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + The call function to the pipeline for generation. + + Args: + video (`list[PIL.Image.Image]`): + The input video used as the starting point for video-to-video generation. The video should be provided + as a list of PIL images, a numpy array, or a torch tensor. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds` + instead. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + height (`int`, defaults to `480`): + The height in pixels of the generated image. + width (`int`, defaults to `832`): + The width in pixels of the generated image. + num_inference_steps (`int`, defaults to `50`): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`list[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, defaults to `5.0`): + Guidance scale as defined in [Classifier-Free Diffusion + Guidance](https://huggingface.co/papers/2207.12598). `guidance_scale` is defined as `w` of equation 2. + of [Imagen Paper](https://huggingface.co/papers/2205.11487). Guidance scale is enabled by setting + `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to + the text `prompt`, usually at the expense of lower image quality. + strength (`float`, defaults to `0.8`): + Higher strength leads to more differences between original image and generated video. + num_videos_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.Tensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.Tensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + output_type (`str`, *optional*, defaults to `"np"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`WanPipelineOutput`] instead of a plain tuple. + attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*): + A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of + each denoising step during the inference. with the following arguments: `callback_on_step_end(self: + DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a + list of all tensors as specified by `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`list`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, defaults to `512`): + The maximum sequence length of the text encoder. If the prompt is longer than this, it will be + truncated. If the prompt is shorter, it will be padded to this length. + + Examples: + + Returns: + [`~WanPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`WanPipelineOutput`] is returned, otherwise a `tuple` is returned where + the first element is a list with the generated images and the second element is a list of `bool`s + indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content. + """ + + if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)): + callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs + + height = height or self.transformer.config.sample_height * self.vae_scale_factor_spatial + width = width or self.transformer.config.sample_width * self.vae_scale_factor_spatial + num_videos_per_prompt = 1 + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + negative_prompt, + height, + width, + video, + latents, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._attention_kwargs = attention_kwargs + self._current_timestep = None + self._interrupt = False + + device = self._execution_device + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # 3. Encode input prompt + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + num_videos_per_prompt=num_videos_per_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + device=device, + ) + + transformer_dtype = self.transformer.dtype + prompt_embeds = prompt_embeds.to(transformer_dtype) + if negative_prompt_embeds is not None: + negative_prompt_embeds = negative_prompt_embeds.to(transformer_dtype) + + # 4. Prepare timesteps + if XLA_AVAILABLE: + timestep_device = "cpu" + else: + timestep_device = device + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, num_inference_steps, timestep_device, timesteps + ) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device) + latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt) + self._num_timesteps = len(timesteps) + + if latents is None: + video = self.video_processor.preprocess_video(video, height=height, width=width).to( + device, dtype=torch.float32 + ) + + # 5. Prepare latent variables + num_channels_latents = self.transformer.config.in_channels + latents = self.prepare_latents( + video, + batch_size * num_videos_per_prompt, + num_channels_latents, + height, + width, + torch.float32, + device, + generator, + latents, + latent_timestep, + ) + + # 6. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + self._current_timestep = t + latent_model_input = latents.to(transformer_dtype) + timestep = t.expand(latents.shape[0]) + + noise_pred = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + + if self.do_classifier_free_guidance: + noise_uncond = self.transformer( + hidden_states=latent_model_input, + timestep=timestep, + encoder_hidden_states=negative_prompt_embeds, + attention_kwargs=attention_kwargs, + return_dict=False, + )[0] + noise_pred = noise_uncond + guidance_scale * (noise_pred - noise_uncond) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + self._current_timestep = None + + if not output_type == "latent": + latents = latents.to(self.vae.dtype) + latents_mean = ( + torch.tensor(self.vae.config.latents_mean) + .view(1, self.vae.config.z_dim, 1, 1, 1) + .to(latents.device, latents.dtype) + ) + latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to( + latents.device, latents.dtype + ) + latents = latents / latents_std + latents_mean + video = self.vae.decode(latents, return_dict=False)[0] + video = self.video_processor.postprocess_video(video, output_type=output_type) + else: + video = latents + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (video,) + + return WanPipelineOutput(frames=video) diff --git a/diffusers/pipelines/z_image/__init__.py b/diffusers/pipelines/z_image/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..14629a6e216058c1800de8a2dbfe28f4ec6d2d66 --- /dev/null +++ b/diffusers/pipelines/z_image/__init__.py @@ -0,0 +1,59 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_torch_and_transformers_objects # noqa: F403 + + _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects)) +else: + _import_structure["pipeline_output"] = ["ZImagePipelineOutput"] + _import_structure["pipeline_z_image"] = ["ZImagePipeline"] + _import_structure["pipeline_z_image_controlnet"] = ["ZImageControlNetPipeline"] + _import_structure["pipeline_z_image_controlnet_inpaint"] = ["ZImageControlNetInpaintPipeline"] + _import_structure["pipeline_z_image_img2img"] = ["ZImageImg2ImgPipeline"] + _import_structure["pipeline_z_image_inpaint"] = ["ZImageInpaintPipeline"] + _import_structure["pipeline_z_image_omni"] = ["ZImageOmniPipeline"] + + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_torch_and_transformers_objects import * + else: + from .pipeline_output import ZImagePipelineOutput + from .pipeline_z_image import ZImagePipeline + from .pipeline_z_image_controlnet import ZImageControlNetPipeline + from .pipeline_z_image_controlnet_inpaint import ZImageControlNetInpaintPipeline + from .pipeline_z_image_img2img import ZImageImg2ImgPipeline + from .pipeline_z_image_inpaint import ZImageInpaintPipeline + from .pipeline_z_image_omni import ZImageOmniPipeline +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/pipelines/z_image/pipeline_output.py b/diffusers/pipelines/z_image/pipeline_output.py new file mode 100644 index 0000000000000000000000000000000000000000..5cbcbee4bfab7a28f33dcbffd0863749320d763a --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_output.py @@ -0,0 +1,34 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import PIL.Image + +from diffusers.utils import BaseOutput + + +@dataclass +class ZImagePipelineOutput(BaseOutput): + """ + Output class for Z-Image pipelines. + + Args: + images (`list[PIL.Image.Image]` or `np.ndarray`) + List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width, + num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline. + """ + + images: list[PIL.Image.Image, np.ndarray] diff --git a/diffusers/pipelines/z_image/pipeline_z_image.py b/diffusers/pipelines/z_image/pipeline_z_image.py new file mode 100644 index 0000000000000000000000000000000000000000..3e2055c6257f3ba8963c464dbf0ab93ca19504f3 --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_z_image.py @@ -0,0 +1,600 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import AutoTokenizer, PreTrainedModel + +from ...image_processor import VaeImageProcessor +from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import ZImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ZImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImagePipeline + + >>> pipe = ZImagePipeline.from_pretrained("Z-a-o/Z-Image-Turbo", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch. + >>> # (1) Use flash attention 2 + >>> # pipe.transformer.set_attention_backend("flash") + >>> # (2) Use flash attention 3 + >>> # pipe.transformer.set_attention_backend("_flash_3") + + >>> prompt = "一幅为名为“造相「Z-IMAGE-TURBO」”的项目设计的创意海报。画面巧妙地将文字概念视觉化：一辆复古蒸汽小火车化身为巨大的拉链头，正拉开厚厚的冬日积雪，展露出一个生机盎然的春天。" + >>> image = pipe( + ... prompt, + ... height=1024, + ... width=1024, + ... num_inference_steps=8, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(42), + ... ).images[0] + >>> image.save("zimage.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +def get_default_z_image_sigmas(num_inference_steps: int) -> list[float]: + return torch.linspace(1.0, 1 / num_inference_steps, num_inference_steps).tolist() + + +class ZImagePipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: ZImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._encode_prompt( + prompt=prompt, + device=device, + prompt_embeds=prompt_embeds, + max_sequence_length=max_sequence_length, + ) + + if do_classifier_free_guidance: + if negative_prompt is None: + negative_prompt = ["" for _ in prompt] + else: + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + assert len(prompt) == len(negative_prompt) + negative_prompt_embeds = self._encode_prompt( + prompt=negative_prompt, + device=device, + prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + ) + else: + negative_prompt_embeds = [] + return prompt_embeds, negative_prompt_embeds + + def _encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + max_sequence_length: int = 512, + ) -> list[torch.FloatTensor]: + device = device or self._execution_device + + if prompt_embeds is not None: + return prompt_embeds + + if isinstance(prompt, str): + prompt = [prompt] + + for i, prompt_item in enumerate(prompt): + messages = [ + {"role": "user", "content": prompt_item}, + ] + prompt_item = self.tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=True, + ) + prompt[i] = prompt_item + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device) + prompt_masks = text_inputs.attention_mask.to(device).bool() + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_masks, + output_hidden_states=True, + ).hidden_states[-2] + + embeddings_list = [] + + for i in range(len(prompt_embeds)): + embeddings_list.append(prompt_embeds[i][prompt_masks[i]]) + + return embeddings_list + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + cfg_normalization: bool = False, + cfg_truncation: float = 1.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: list[torch.FloatTensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + cfg_normalization (`bool`, *optional*, defaults to False): + Whether to apply configuration normalization. + cfg_truncation (`float`, *optional*, defaults to 1.0): + The truncation value for configuration. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain + tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + height = height or 1024 + width = width or 1024 + + vae_scale = self.vae_scale_factor * 2 + if height % vae_scale != 0: + raise ValueError( + f"Height must be divisible by {vae_scale} (got {height}). " + f"Please adjust the height to a multiple of {vae_scale}." + ) + if width % vae_scale != 0: + raise ValueError( + f"Width must be divisible by {vae_scale} (got {width}). " + f"Please adjust the width to a multiple of {vae_scale}." + ) + + device = self._execution_device + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._cfg_normalization = cfg_normalization + self._cfg_truncation = cfg_truncation + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is not None and prompt is None: + if self.do_classifier_free_guidance and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided without `prompt`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + else: + ( + prompt_embeds, + negative_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.in_channels + + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # Repeat prompt_embeds for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] + if self.do_classifier_free_guidance and negative_prompt_embeds: + negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] + + actual_batch_size = batch_size * num_images_per_prompt + image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2) + + # 5. Prepare timesteps + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if sigmas is None: + sigmas = get_default_z_image_sigmas(num_inference_steps) + scheduler_kwargs = {"mu": mu} + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # We set the index here to remove DtoH sync, helpful especially during compilation. + # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 + self.scheduler.set_begin_index(0) + + if self.do_classifier_free_guidance and self._cfg_truncation is not None and float(self._cfg_truncation) <= 1: + _precomputed_t_norms = ((1000 - timesteps.float()) / 1000).tolist() + else: + _precomputed_t_norms = None + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + timestep = (1000 - timestep) / 1000 + current_guidance_scale = self.guidance_scale + if _precomputed_t_norms is not None: + if _precomputed_t_norms[i] > self._cfg_truncation: + current_guidance_scale = 0.0 + + # Run CFG only if configured AND scale is non-zero + apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 + + if apply_cfg: + latents_typed = latents.to(self.transformer.dtype) + latent_model_input = latents_typed.repeat(2, 1, 1, 1) + prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds + timestep_model_input = timestep.repeat(2) + else: + latent_model_input = latents.to(self.transformer.dtype) + prompt_embeds_model_input = prompt_embeds + timestep_model_input = timestep + + latent_model_input = latent_model_input.unsqueeze(2) + latent_model_input_list = list(latent_model_input.unbind(dim=0)) + + model_out_list = self.transformer( + latent_model_input_list, timestep_model_input, prompt_embeds_model_input, return_dict=False + )[0] + + if apply_cfg: + # Perform CFG + pos_out = model_out_list[:actual_batch_size] + neg_out = model_out_list[actual_batch_size:] + + noise_pred = [] + for j in range(actual_batch_size): + pos = pos_out[j].float() + neg = neg_out[j].float() + + pred = pos + current_guidance_scale * (pos - neg) + + # Renormalization + if self._cfg_normalization and float(self._cfg_normalization) > 0.0: + ori_pos_norm = torch.linalg.vector_norm(pos) + new_pos_norm = torch.linalg.vector_norm(pred) + max_new_norm = ori_pos_norm * float(self._cfg_normalization) + if new_pos_norm > max_new_norm: + pred = pred * (max_new_norm / new_pos_norm) + + noise_pred.append(pred) + + noise_pred = torch.stack(noise_pred, dim=0) + else: + noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) + + noise_pred = noise_pred.squeeze(2) + noise_pred = -noise_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] + assert latents.dtype == torch.float32 + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if output_type == "latent": + image = latents + + else: + latents = latents.to(self.vae.dtype) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ZImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/z_image/pipeline_z_image_controlnet.py b/diffusers/pipelines/z_image/pipeline_z_image_controlnet.py new file mode 100644 index 0000000000000000000000000000000000000000..81373ffb56ffd9b4c8d87b7ae45588cc0a7c87ae --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_z_image_controlnet.py @@ -0,0 +1,740 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import AutoTokenizer, PreTrainedModel + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.controlnets import ZImageControlNetModel +from ...models.transformers import ZImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ZImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImageControlNetPipeline + >>> from diffusers import ZImageControlNetModel + >>> from diffusers.utils import load_image + >>> from huggingface_hub import hf_hub_download + + >>> controlnet = ZImageControlNetModel.from_single_file( + ... hf_hub_download( + ... "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union", + ... filename="Z-Image-Turbo-Fun-Controlnet-Union.safetensors", + ... ), + ... torch_dtype=torch.bfloat16, + ... ) + + >>> # 2.1 + >>> # controlnet = ZImageControlNetModel.from_single_file( + >>> # hf_hub_download( + >>> # "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0", + >>> # filename="Z-Image-Turbo-Fun-Controlnet-Union-2.1.safetensors", + >>> # ), + >>> # torch_dtype=torch.bfloat16, + >>> # ) + + >>> # 2.0 + >>> # controlnet = ZImageControlNetModel.from_single_file( + >>> # hf_hub_download( + >>> # "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0", + >>> # filename="Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors", + >>> # ), + >>> # torch_dtype=torch.bfloat16, + >>> # ) + + >>> pipe = ZImageControlNetPipeline.from_pretrained( + ... "Tongyi-MAI/Z-Image-Turbo", controlnet=controlnet, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch. + >>> # (1) Use flash attention 2 + >>> # pipe.transformer.set_attention_backend("flash") + >>> # (2) Use flash attention 3 + >>> # pipe.transformer.set_attention_backend("_flash_3") + + >>> control_image = load_image( + ... "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union/resolve/main/asset/pose.jpg?download=true" + ... ) + >>> prompt = "一位年轻女子站在阳光明媚的海岸线上，白裙在轻拂的海风中微微飘动。她拥有一头鲜艳的紫色长发，在风中轻盈舞动，发间系着一个精致的黑色蝴蝶结，与身后柔和的蔚蓝天空形成鲜明对比。她面容清秀，眉目精致，透着一股甜美的青春气息；神情柔和，略带羞涩，目光静静地凝望着远方的地平线，双手自然交叠于身前，仿佛沉浸在思绪之中。在她身后，是辽阔无垠、波光粼粼的大海，阳光洒在海面上，映出温暖的金色光晕。" + >>> image = pipe( + ... prompt, + ... control_image=control_image, + ... controlnet_conditioning_scale=0.75, + ... height=1728, + ... width=992, + ... num_inference_steps=8, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(43), + ... ).images[0] + >>> image.save("zimage.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.z_image.pipeline_z_image.get_default_z_image_sigmas +def get_default_z_image_sigmas(num_inference_steps: int) -> list[float]: + return torch.linspace(1.0, 1 / num_inference_steps, num_inference_steps).tolist() + + +class ZImageControlNetPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: ZImageTransformer2DModel, + controlnet: ZImageControlNetModel, + ): + super().__init__() + controlnet = ZImageControlNetModel.from_transformer(controlnet, transformer) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + controlnet=controlnet, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._encode_prompt( + prompt=prompt, + device=device, + prompt_embeds=prompt_embeds, + max_sequence_length=max_sequence_length, + ) + + if do_classifier_free_guidance: + if negative_prompt is None: + negative_prompt = ["" for _ in prompt] + else: + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + assert len(prompt) == len(negative_prompt) + negative_prompt_embeds = self._encode_prompt( + prompt=negative_prompt, + device=device, + prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + ) + else: + negative_prompt_embeds = [] + return prompt_embeds, negative_prompt_embeds + + def _encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + max_sequence_length: int = 512, + ) -> list[torch.FloatTensor]: + device = device or self._execution_device + + if prompt_embeds is not None: + return prompt_embeds + + if isinstance(prompt, str): + prompt = [prompt] + + for i, prompt_item in enumerate(prompt): + messages = [ + {"role": "user", "content": prompt_item}, + ] + prompt_item = self.tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=True, + ) + prompt[i] = prompt_item + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device) + prompt_masks = text_inputs.attention_mask.to(device).bool() + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_masks, + output_hidden_states=True, + ).hidden_states[-2] + + embeddings_list = [] + + for i in range(len(prompt_embeds)): + embeddings_list.append(prompt_embeds[i][prompt_masks[i]]) + + return embeddings_list + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + return latents + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + control_image: PipelineImageInput = None, + controlnet_conditioning_scale: float | list[float] = 0.75, + cfg_normalization: bool = False, + cfg_truncation: float = 1.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: list[torch.FloatTensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + control_image (`PipelineImageInput`): + The ControlNet input condition to provide guidance to the `transformer` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. The dimensions of the output image defaults to `control_image`'s dimensions. If height + and/or width are passed, `control_image` is resized accordingly. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 0.75): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `transformer`. + cfg_normalization (`bool`, *optional*, defaults to False): + Whether to apply configuration normalization. + cfg_truncation (`float`, *optional*, defaults to 1.0): + The truncation value for configuration. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain + tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + height = height or 1024 + width = width or 1024 + + vae_scale = self.vae_scale_factor * 2 + if height % vae_scale != 0: + raise ValueError( + f"Height must be divisible by {vae_scale} (got {height}). " + f"Please adjust the height to a multiple of {vae_scale}." + ) + if width % vae_scale != 0: + raise ValueError( + f"Width must be divisible by {vae_scale} (got {width}). " + f"Please adjust the width to a multiple of {vae_scale}." + ) + + device = self._execution_device + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._cfg_normalization = cfg_normalization + self._cfg_truncation = cfg_truncation + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is not None and prompt is None: + if self.do_classifier_free_guidance and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided without `prompt`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + else: + ( + prompt_embeds, + negative_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.in_channels + + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image.shape[-2:] + control_image = retrieve_latents(self.vae.encode(control_image), generator=generator, sample_mode="argmax") + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + control_image = control_image.unsqueeze(2) + + if num_channels_latents != self.controlnet.config.control_in_dim: + # For model version 2.0 + control_image = torch.cat( + [ + control_image, + torch.zeros( + control_image.shape[0], + self.controlnet.config.control_in_dim - num_channels_latents, + *control_image.shape[2:], + ).to(device=control_image.device, dtype=control_image.dtype), + ], + dim=1, + ) + + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # Repeat prompt_embeds for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] + if self.do_classifier_free_guidance and negative_prompt_embeds: + negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] + + actual_batch_size = batch_size * num_images_per_prompt + image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2) + + # 5. Prepare timesteps + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if sigmas is None: + sigmas = get_default_z_image_sigmas(num_inference_steps) + scheduler_kwargs = {"mu": mu} + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + timestep = (1000 - timestep) / 1000 + # Normalized time for time-aware config (0 at start, 1 at end) + t_norm = timestep[0].item() + + # Handle cfg truncation + current_guidance_scale = self.guidance_scale + if ( + self.do_classifier_free_guidance + and self._cfg_truncation is not None + and float(self._cfg_truncation) <= 1 + ): + if t_norm > self._cfg_truncation: + current_guidance_scale = 0.0 + + # Run CFG only if configured AND scale is non-zero + apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 + + if apply_cfg: + latents_typed = latents.to(self.transformer.dtype) + latent_model_input = latents_typed.repeat(2, 1, 1, 1) + prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds + timestep_model_input = timestep.repeat(2) + control_image_input = control_image.repeat(2, 1, 1, 1, 1) + else: + latent_model_input = latents.to(self.transformer.dtype) + prompt_embeds_model_input = prompt_embeds + timestep_model_input = timestep + control_image_input = control_image + + latent_model_input = latent_model_input.unsqueeze(2) + latent_model_input_list = list(latent_model_input.unbind(dim=0)) + + controlnet_block_samples = self.controlnet( + latent_model_input_list, + timestep_model_input, + prompt_embeds_model_input, + control_image_input, + conditioning_scale=controlnet_conditioning_scale, + ) + + model_out_list = self.transformer( + latent_model_input_list, + timestep_model_input, + prompt_embeds_model_input, + controlnet_block_samples=controlnet_block_samples, + )[0] + + if apply_cfg: + # Perform CFG + pos_out = model_out_list[:actual_batch_size] + neg_out = model_out_list[actual_batch_size:] + + noise_pred = [] + for j in range(actual_batch_size): + pos = pos_out[j].float() + neg = neg_out[j].float() + + pred = pos + current_guidance_scale * (pos - neg) + + # Renormalization + if self._cfg_normalization and float(self._cfg_normalization) > 0.0: + ori_pos_norm = torch.linalg.vector_norm(pos) + new_pos_norm = torch.linalg.vector_norm(pred) + max_new_norm = ori_pos_norm * float(self._cfg_normalization) + if new_pos_norm > max_new_norm: + pred = pred * (max_new_norm / new_pos_norm) + + noise_pred.append(pred) + + noise_pred = torch.stack(noise_pred, dim=0) + else: + noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) + + noise_pred = noise_pred.squeeze(2) + noise_pred = -noise_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] + assert latents.dtype == torch.float32 + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if output_type == "latent": + image = latents + + else: + latents = latents.to(self.vae.dtype) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ZImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/z_image/pipeline_z_image_controlnet_inpaint.py b/diffusers/pipelines/z_image/pipeline_z_image_controlnet_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..178e74dea4fa45f225853672abaeba4d1c285f8d --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_z_image_controlnet_inpaint.py @@ -0,0 +1,767 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +import torch.nn.functional as F +from transformers import AutoTokenizer, PreTrainedModel + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.controlnets import ZImageControlNetModel +from ...models.transformers import ZImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ZImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImageControlNetInpaintPipeline + >>> from diffusers import ZImageControlNetModel + >>> from diffusers.utils import load_image + >>> from huggingface_hub import hf_hub_download + + >>> controlnet = ZImageControlNetModel.from_single_file( + ... hf_hub_download( + ... "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0", + ... filename="Z-Image-Turbo-Fun-Controlnet-Union-2.1.safetensors", + ... ), + ... torch_dtype=torch.bfloat16, + ... ) + + >>> # 2.0 + >>> # controlnet = ZImageControlNetModel.from_single_file( + >>> # hf_hub_download( + >>> # "alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0", + >>> # filename="Z-Image-Turbo-Fun-Controlnet-Union-2.0.safetensors", + >>> # ), + >>> # torch_dtype=torch.bfloat16, + >>> # ) + + >>> pipe = ZImageControlNetInpaintPipeline.from_pretrained( + ... "Tongyi-MAI/Z-Image-Turbo", controlnet=controlnet, torch_dtype=torch.bfloat16 + ... ) + >>> pipe.to("cuda") + + >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch. + >>> # (1) Use flash attention 2 + >>> # pipe.transformer.set_attention_backend("flash") + >>> # (2) Use flash attention 3 + >>> # pipe.transformer.set_attention_backend("_flash_3") + + >>> image = load_image( + ... "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/asset/inpaint.jpg?download=true" + ... ) + >>> mask_image = load_image( + ... "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/asset/mask.jpg?download=true" + ... ) + >>> control_image = load_image( + ... "https://huggingface.co/alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0/resolve/main/asset/pose.jpg?download=true" + ... ) + >>> prompt = "一位年轻女子站在阳光明媚的海岸线上，画面为全身竖构图，身体微微侧向右侧，左手自然下垂，右臂弯曲扶在腰间，她的手指清晰可见，站姿放松而略带羞涩。她身穿轻盈的白色连衣裙，裙摆在海风中轻轻飘动，布料半透、质感柔软。女子拥有一头鲜艳的及腰紫色长发，被海风吹起，在身侧轻盈飞舞，发间系着一个精致的黑色蝴蝶结，与发色形成对比。她面容清秀，眉目精致，肤色白皙细腻，表情温柔略显羞涩，微微低头，眼神静静望向远处的海平线，流露出甜美的青春气息与若有所思的神情。背景是辽阔无垠的海洋与蔚蓝天空，阳光从侧前方洒下，海面波光粼粼，泛着温暖的金色光晕，天空清澈明亮，云朵稀薄，整体色调清新唯美。" + >>> image = pipe( + ... prompt, + ... image=image, + ... mask_image=mask_image, + ... control_image=control_image, + ... controlnet_conditioning_scale=0.75, + ... height=1728, + ... width=992, + ... num_inference_steps=25, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(43), + ... ).images[0] + >>> image.save("zimage-inpaint.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.z_image.pipeline_z_image.get_default_z_image_sigmas +def get_default_z_image_sigmas(num_inference_steps: int) -> list[float]: + return torch.linspace(1.0, 1 / num_inference_steps, num_inference_steps).tolist() + + +class ZImageControlNetInpaintPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: ZImageTransformer2DModel, + controlnet: ZImageControlNetModel, + ): + super().__init__() + if transformer.in_channels == controlnet.config.control_in_dim: + raise ValueError( + "ZImageControlNetInpaintPipeline is not compatible with `alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union`, use `alibaba-pai/Z-Image-Turbo-Fun-Controlnet-Union-2.0`." + ) + controlnet = ZImageControlNetModel.from_transformer(controlnet, transformer) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + controlnet=controlnet, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + ) + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._encode_prompt( + prompt=prompt, + device=device, + prompt_embeds=prompt_embeds, + max_sequence_length=max_sequence_length, + ) + + if do_classifier_free_guidance: + if negative_prompt is None: + negative_prompt = ["" for _ in prompt] + else: + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + assert len(prompt) == len(negative_prompt) + negative_prompt_embeds = self._encode_prompt( + prompt=negative_prompt, + device=device, + prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + ) + else: + negative_prompt_embeds = [] + return prompt_embeds, negative_prompt_embeds + + def _encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + max_sequence_length: int = 512, + ) -> list[torch.FloatTensor]: + device = device or self._execution_device + + if prompt_embeds is not None: + return prompt_embeds + + if isinstance(prompt, str): + prompt = [prompt] + + for i, prompt_item in enumerate(prompt): + messages = [ + {"role": "user", "content": prompt_item}, + ] + prompt_item = self.tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=True, + ) + prompt[i] = prompt_item + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device) + prompt_masks = text_inputs.attention_mask.to(device).bool() + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_masks, + output_hidden_states=True, + ).hidden_states[-2] + + embeddings_list = [] + + for i in range(len(prompt_embeds)): + embeddings_list.append(prompt_embeds[i][prompt_masks[i]]) + + return embeddings_list + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + return latents + + # Copied from diffusers.pipelines.controlnet_sd3.pipeline_stable_diffusion_3_controlnet.StableDiffusion3ControlNetPipeline.prepare_image + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + if isinstance(image, torch.Tensor): + pass + else: + image = self.image_processor.preprocess(image, height=height, width=width) + + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + control_image: PipelineImageInput = None, + controlnet_conditioning_scale: float | list[float] = 0.75, + cfg_normalization: bool = False, + cfg_truncation: float = 1.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: list[torch.FloatTensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + image (`PipelineImageInput`): + `Image`, numpy array or tensor representing an image batch to be inpainted (which parts of the image to + be masked out with `mask_image` and repainted according to `prompt`). + mask_image (`PipelineImageInput`): + `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask + are repainted while black pixels are preserved. + control_image (`PipelineImageInput`): + The ControlNet input condition to provide guidance to the `transformer` for generation. If the type is + specified as `torch.Tensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted + as an image. + controlnet_conditioning_scale (`float` or `list[float]`, *optional*, defaults to 0.75): + The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added + to the residual in the original `transformer`. + cfg_normalization (`bool`, *optional*, defaults to False): + Whether to apply configuration normalization. + cfg_truncation (`float`, *optional*, defaults to 1.0): + The truncation value for configuration. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain + tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + height = height or 1024 + width = width or 1024 + + vae_scale = self.vae_scale_factor * 2 + if height % vae_scale != 0: + raise ValueError( + f"Height must be divisible by {vae_scale} (got {height}). " + f"Please adjust the height to a multiple of {vae_scale}." + ) + if width % vae_scale != 0: + raise ValueError( + f"Width must be divisible by {vae_scale} (got {width}). " + f"Please adjust the width to a multiple of {vae_scale}." + ) + + device = self._execution_device + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._cfg_normalization = cfg_normalization + self._cfg_truncation = cfg_truncation + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is not None and prompt is None: + if self.do_classifier_free_guidance and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided without `prompt`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + else: + ( + prompt_embeds, + negative_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.in_channels + + control_image = self.prepare_image( + image=control_image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = control_image.shape[-2:] + control_image = retrieve_latents(self.vae.encode(control_image), generator=generator, sample_mode="argmax") + control_image = (control_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + control_image = control_image.unsqueeze(2) + + mask_condition = self.mask_processor.preprocess(mask_image, height=height, width=width) + mask_condition = torch.tile(mask_condition, [1, 3, 1, 1]).to( + device=control_image.device, dtype=control_image.dtype + ) + + init_image = self.prepare_image( + image=image, + width=width, + height=height, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=self.vae.dtype, + ) + height, width = init_image.shape[-2:] + init_image = init_image * (mask_condition < 0.5) + init_image = retrieve_latents(self.vae.encode(init_image), generator=generator, sample_mode="argmax") + init_image = (init_image - self.vae.config.shift_factor) * self.vae.config.scaling_factor + init_image = init_image.unsqueeze(2) + + mask_condition = F.interpolate(1 - mask_condition[:, :1], size=init_image.size()[-2:], mode="nearest").to( + device=control_image.device, dtype=control_image.dtype + ) + mask_condition = mask_condition.unsqueeze(2) + + control_image = torch.cat([control_image, mask_condition, init_image], dim=1) + + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + # Repeat prompt_embeds for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] + if self.do_classifier_free_guidance and negative_prompt_embeds: + negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] + + actual_batch_size = batch_size * num_images_per_prompt + image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2) + + # 5. Prepare timesteps + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if sigmas is None: + sigmas = get_default_z_image_sigmas(num_inference_steps) + scheduler_kwargs = {"mu": mu} + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + timestep = (1000 - timestep) / 1000 + # Normalized time for time-aware config (0 at start, 1 at end) + t_norm = timestep[0].item() + + # Handle cfg truncation + current_guidance_scale = self.guidance_scale + if ( + self.do_classifier_free_guidance + and self._cfg_truncation is not None + and float(self._cfg_truncation) <= 1 + ): + if t_norm > self._cfg_truncation: + current_guidance_scale = 0.0 + + # Run CFG only if configured AND scale is non-zero + apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 + + if apply_cfg: + latents_typed = latents.to(self.transformer.dtype) + latent_model_input = latents_typed.repeat(2, 1, 1, 1) + prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds + timestep_model_input = timestep.repeat(2) + control_image_input = control_image.repeat(2, 1, 1, 1, 1) + else: + latent_model_input = latents.to(self.transformer.dtype) + prompt_embeds_model_input = prompt_embeds + timestep_model_input = timestep + control_image_input = control_image + + latent_model_input = latent_model_input.unsqueeze(2) + latent_model_input_list = list(latent_model_input.unbind(dim=0)) + + controlnet_block_samples = self.controlnet( + latent_model_input_list, + timestep_model_input, + prompt_embeds_model_input, + control_image_input, + conditioning_scale=controlnet_conditioning_scale, + ) + + model_out_list = self.transformer( + latent_model_input_list, + timestep_model_input, + prompt_embeds_model_input, + controlnet_block_samples=controlnet_block_samples, + )[0] + + if apply_cfg: + # Perform CFG + pos_out = model_out_list[:actual_batch_size] + neg_out = model_out_list[actual_batch_size:] + + noise_pred = [] + for j in range(actual_batch_size): + pos = pos_out[j].float() + neg = neg_out[j].float() + + pred = pos + current_guidance_scale * (pos - neg) + + # Renormalization + if self._cfg_normalization and float(self._cfg_normalization) > 0.0: + ori_pos_norm = torch.linalg.vector_norm(pos) + new_pos_norm = torch.linalg.vector_norm(pred) + max_new_norm = ori_pos_norm * float(self._cfg_normalization) + if new_pos_norm > max_new_norm: + pred = pred * (max_new_norm / new_pos_norm) + + noise_pred.append(pred) + + noise_pred = torch.stack(noise_pred, dim=0) + else: + noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) + + noise_pred = noise_pred.squeeze(2) + noise_pred = -noise_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] + assert latents.dtype == torch.float32 + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if output_type == "latent": + image = latents + + else: + latents = latents.to(self.vae.dtype) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ZImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/z_image/pipeline_z_image_img2img.py b/diffusers/pipelines/z_image/pipeline_z_image_img2img.py new file mode 100644 index 0000000000000000000000000000000000000000..b5c7740bb0c1324bd5397871c51a791d8962bd06 --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_z_image_img2img.py @@ -0,0 +1,715 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import torch +from transformers import AutoTokenizer, PreTrainedModel + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import ZImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ZImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImageImg2ImgPipeline + >>> from diffusers.utils import load_image + + >>> pipe = ZImageImg2ImgPipeline.from_pretrained("Z-a-o/Z-Image-Turbo", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + + >>> prompt = "A fantasy landscape with mountains and a river, detailed, vibrant colors" + >>> image = pipe( + ... prompt, + ... image=init_image, + ... strength=0.6, + ... num_inference_steps=8, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(42), + ... ).images[0] + >>> image.save("zimage_img2img.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.z_image.pipeline_z_image.get_default_z_image_sigmas +def get_default_z_image_sigmas(num_inference_steps: int) -> list[float]: + return torch.linspace(1.0, 1 / num_inference_steps, num_inference_steps).tolist() + + +class ZImageImg2ImgPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + r""" + The ZImage pipeline for image-to-image generation. + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`PreTrainedModel`]): + A text encoder model to encode text prompts. + tokenizer ([`AutoTokenizer`]): + A tokenizer to tokenize text prompts. + transformer ([`ZImageTransformer2DModel`]): + A ZImage transformer model to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: ZImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._encode_prompt( + prompt=prompt, + device=device, + prompt_embeds=prompt_embeds, + max_sequence_length=max_sequence_length, + ) + + if do_classifier_free_guidance: + if negative_prompt is None: + negative_prompt = ["" for _ in prompt] + else: + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + assert len(prompt) == len(negative_prompt) + negative_prompt_embeds = self._encode_prompt( + prompt=negative_prompt, + device=device, + prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + ) + else: + negative_prompt_embeds = [] + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline._encode_prompt + def _encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + max_sequence_length: int = 512, + ) -> list[torch.FloatTensor]: + device = device or self._execution_device + + if prompt_embeds is not None: + return prompt_embeds + + if isinstance(prompt, str): + prompt = [prompt] + + for i, prompt_item in enumerate(prompt): + messages = [ + {"role": "user", "content": prompt_item}, + ] + prompt_item = self.tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=True, + ) + prompt[i] = prompt_item + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device) + prompt_masks = text_inputs.attention_mask.to(device).bool() + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_masks, + output_hidden_states=True, + ).hidden_states[-2] + + embeddings_list = [] + + for i in range(len(prompt_embeds)): + embeddings_list.append(prompt_embeds[i][prompt_masks[i]]) + + return embeddings_list + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + return latents.to(device=device, dtype=dtype) + + # Encode the input image + image = image.to(device=device, dtype=dtype) + if image.shape[1] != num_channels_latents: + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + # Apply scaling (inverse of decoding: decode does latents/scaling_factor + shift_factor) + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + else: + image_latents = image + + # Handle batch size expansion + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + + # Add noise using flow matching scale_noise + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + + return latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: str | list[str] = None, + image: PipelineImageInput = None, + strength: float = 0.6, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + cfg_normalization: bool = False, + cfg_truncation: float = 1.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: list[torch.FloatTensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for image-to-image generation. + + Args: + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a + list of tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or + a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + strength (`float`, *optional*, defaults to 0.6): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image`. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. If not provided, uses the input image height. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. If not provided, uses the input image width. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + cfg_normalization (`bool`, *optional*, defaults to False): + Whether to apply configuration normalization. + cfg_truncation (`float`, *optional*, defaults to 1.0): + The truncation value for configuration. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain + tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + # 1. Check inputs and validate strength + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should be in [0.0, 1.0] but is {strength}") + + # 2. Preprocess image + init_image = self.image_processor.preprocess(image) + init_image = init_image.to(dtype=torch.float32) + + # Get dimensions from the preprocessed image if not specified + if height is None: + height = init_image.shape[-2] + if width is None: + width = init_image.shape[-1] + + vae_scale = self.vae_scale_factor * 2 + if height % vae_scale != 0: + raise ValueError( + f"Height must be divisible by {vae_scale} (got {height}). " + f"Please adjust the height to a multiple of {vae_scale}." + ) + if width % vae_scale != 0: + raise ValueError( + f"Width must be divisible by {vae_scale} (got {width}). " + f"Please adjust the width to a multiple of {vae_scale}." + ) + + device = self._execution_device + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._cfg_normalization = cfg_normalization + self._cfg_truncation = cfg_truncation + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is not None and prompt is None: + if self.do_classifier_free_guidance and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided without `prompt`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + else: + ( + prompt_embeds, + negative_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.in_channels + + # Repeat prompt_embeds for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] + if self.do_classifier_free_guidance and negative_prompt_embeds: + negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] + + actual_batch_size = batch_size * num_images_per_prompt + + # Calculate latent dimensions for image_seq_len + latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) + latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) + image_seq_len = (latent_height // 2) * (latent_width // 2) + + # 5. Prepare timesteps + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if sigmas is None: + sigmas = get_default_z_image_sigmas(num_inference_steps) + scheduler_kwargs = {"mu": mu} + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + + # 6. Adjust timesteps based on strength + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline " + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(actual_batch_size) + + # 7. Prepare latents from image + latents = self.prepare_latents( + init_image, + latent_timestep, + actual_batch_size, + num_channels_latents, + height, + width, + prompt_embeds[0].dtype, + device, + generator, + latents, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 8. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + timestep = (1000 - timestep) / 1000 + # Normalized time for time-aware config (0 at start, 1 at end) + t_norm = timestep[0].item() + + # Handle cfg truncation + current_guidance_scale = self.guidance_scale + if ( + self.do_classifier_free_guidance + and self._cfg_truncation is not None + and float(self._cfg_truncation) <= 1 + ): + if t_norm > self._cfg_truncation: + current_guidance_scale = 0.0 + + # Run CFG only if configured AND scale is non-zero + apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 + + if apply_cfg: + latents_typed = latents.to(self.transformer.dtype) + latent_model_input = latents_typed.repeat(2, 1, 1, 1) + prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds + timestep_model_input = timestep.repeat(2) + else: + latent_model_input = latents.to(self.transformer.dtype) + prompt_embeds_model_input = prompt_embeds + timestep_model_input = timestep + + latent_model_input = latent_model_input.unsqueeze(2) + latent_model_input_list = list(latent_model_input.unbind(dim=0)) + + model_out_list = self.transformer( + latent_model_input_list, + timestep_model_input, + prompt_embeds_model_input, + )[0] + + if apply_cfg: + # Perform CFG + pos_out = model_out_list[:actual_batch_size] + neg_out = model_out_list[actual_batch_size:] + + noise_pred = [] + for j in range(actual_batch_size): + pos = pos_out[j].float() + neg = neg_out[j].float() + + pred = pos + current_guidance_scale * (pos - neg) + + # Renormalization + if self._cfg_normalization and float(self._cfg_normalization) > 0.0: + ori_pos_norm = torch.linalg.vector_norm(pos) + new_pos_norm = torch.linalg.vector_norm(pred) + max_new_norm = ori_pos_norm * float(self._cfg_normalization) + if new_pos_norm > max_new_norm: + pred = pred * (max_new_norm / new_pos_norm) + + noise_pred.append(pred) + + noise_pred = torch.stack(noise_pred, dim=0) + else: + noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) + + noise_pred = noise_pred.squeeze(2) + noise_pred = -noise_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] + assert latents.dtype == torch.float32 + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if output_type == "latent": + image = latents + + else: + latents = latents.to(self.vae.dtype) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ZImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/z_image/pipeline_z_image_inpaint.py b/diffusers/pipelines/z_image/pipeline_z_image_inpaint.py new file mode 100644 index 0000000000000000000000000000000000000000..132c22c0cff3d35fa6a1115ded82c17daff67230 --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_z_image_inpaint.py @@ -0,0 +1,938 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from transformers import AutoTokenizer, PreTrainedModel + +from ...image_processor import PipelineImageInput, VaeImageProcessor +from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import ZImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import is_torch_xla_available, logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from .pipeline_output import ZImagePipelineOutput + + +if is_torch_xla_available(): + import torch_xla.core.xla_model as xm + + XLA_AVAILABLE = True +else: + XLA_AVAILABLE = False + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImageInpaintPipeline + >>> from diffusers.utils import load_image + + >>> pipe = ZImageInpaintPipeline.from_pretrained("Tongyi-MAI/Z-Image-Turbo", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg" + >>> init_image = load_image(url).resize((1024, 1024)) + + >>> # Create a mask (white = inpaint, black = preserve) + >>> import numpy as np + >>> from PIL import Image + + >>> mask = np.zeros((1024, 1024), dtype=np.uint8) + >>> mask[256:768, 256:768] = 255 # Inpaint center region + >>> mask_image = Image.fromarray(mask) + + >>> prompt = "A beautiful lake with mountains in the background" + >>> image = pipe( + ... prompt, + ... image=init_image, + ... mask_image=mask_image, + ... strength=1.0, + ... num_inference_steps=8, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(42), + ... ).images[0] + >>> image.save("zimage_inpaint.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents +def retrieve_latents( + encoder_output: torch.Tensor, generator: torch.Generator | None = None, sample_mode: str = "sample" +): + if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": + return encoder_output.latent_dist.sample(generator) + elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": + return encoder_output.latent_dist.mode() + elif hasattr(encoder_output, "latents"): + return encoder_output.latents + else: + raise AttributeError("Could not access latents of provided encoder_output") + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.z_image.pipeline_z_image.get_default_z_image_sigmas +def get_default_z_image_sigmas(num_inference_steps: int) -> list[float]: + return torch.linspace(1.0, 1 / num_inference_steps, num_inference_steps).tolist() + + +class ZImageInpaintPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + r""" + The ZImage pipeline for inpainting. + + Args: + scheduler ([`FlowMatchEulerDiscreteScheduler`]): + A scheduler to be used in combination with `transformer` to denoise the encoded image latents. + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`PreTrainedModel`]): + A text encoder model to encode text prompts. + tokenizer ([`AutoTokenizer`]): + A tokenizer to tokenize text prompts. + transformer ([`ZImageTransformer2DModel`]): + A ZImage transformer model to denoise the encoded image latents. + """ + + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds", "mask", "masked_image_latents"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: ZImageTransformer2DModel, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.mask_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor * 2, + do_normalize=False, + do_binarize=True, + do_convert_grayscale=True, + ) + + # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline.encode_prompt + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._encode_prompt( + prompt=prompt, + device=device, + prompt_embeds=prompt_embeds, + max_sequence_length=max_sequence_length, + ) + + if do_classifier_free_guidance: + if negative_prompt is None: + negative_prompt = ["" for _ in prompt] + else: + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + assert len(prompt) == len(negative_prompt) + negative_prompt_embeds = self._encode_prompt( + prompt=negative_prompt, + device=device, + prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + ) + else: + negative_prompt_embeds = [] + return prompt_embeds, negative_prompt_embeds + + # Copied from diffusers.pipelines.z_image.pipeline_z_image.ZImagePipeline._encode_prompt + def _encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + max_sequence_length: int = 512, + ) -> list[torch.FloatTensor]: + device = device or self._execution_device + + if prompt_embeds is not None: + return prompt_embeds + + if isinstance(prompt, str): + prompt = [prompt] + + for i, prompt_item in enumerate(prompt): + messages = [ + {"role": "user", "content": prompt_item}, + ] + prompt_item = self.tokenizer.apply_chat_template( + messages, + tokenize=False, + add_generation_prompt=True, + enable_thinking=True, + ) + prompt[i] = prompt_item + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device) + prompt_masks = text_inputs.attention_mask.to(device).bool() + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_masks, + output_hidden_states=True, + ).hidden_states[-2] + + embeddings_list = [] + + for i in range(len(prompt_embeds)): + embeddings_list.append(prompt_embeds[i][prompt_masks[i]]) + + return embeddings_list + + # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps + def get_timesteps(self, num_inference_steps, strength, device): + # get the original timestep using init_timestep + init_timestep = min(num_inference_steps * strength, num_inference_steps) + + t_start = int(max(num_inference_steps - init_timestep, 0)) + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + if hasattr(self.scheduler, "set_begin_index"): + self.scheduler.set_begin_index(t_start * self.scheduler.order) + + return timesteps, num_inference_steps - t_start + + def prepare_mask_latents( + self, + mask, + masked_image, + batch_size, + height, + width, + dtype, + device, + generator, + ): + """Prepare mask and masked image latents for inpainting. + + Args: + mask: Binary mask tensor where 1 = inpaint region, 0 = preserve region. + masked_image: Original image with masked regions zeroed out. + batch_size: Number of images to generate. + height: Output image height. + width: Output image width. + dtype: Data type for the tensors. + device: Device to place tensors on. + generator: Random generator for reproducibility. + + Returns: + Tuple of (mask, masked_image_latents) prepared for the denoising loop. + """ + # Calculate latent dimensions + latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) + latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + # Resize mask to latent dimensions + mask = torch.nn.functional.interpolate(mask, size=(latent_height, latent_width), mode="nearest") + mask = mask.to(device=device, dtype=dtype) + + # Encode masked image to latents + masked_image = masked_image.to(device=device, dtype=dtype) + if isinstance(generator, list): + masked_image_latents = [ + retrieve_latents(self.vae.encode(masked_image[i : i + 1]), generator=generator[i]) + for i in range(masked_image.shape[0]) + ] + masked_image_latents = torch.cat(masked_image_latents, dim=0) + else: + masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator) + + # Apply VAE scaling + masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + + # Expand for batch size + if mask.shape[0] < batch_size: + if not batch_size % mask.shape[0] == 0: + raise ValueError( + "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" + f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" + " of masks that you pass is divisible by the total requested batch size." + ) + mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) + if masked_image_latents.shape[0] < batch_size: + if not batch_size % masked_image_latents.shape[0] == 0: + raise ValueError( + "The passed images and the required batch size don't match. Images are supposed to be duplicated" + f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." + " Make sure the number of images that you pass is divisible by the total requested batch size." + ) + masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) + + return mask, masked_image_latents + + def prepare_latents( + self, + image, + timestep, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + """Prepare latents for inpainting, returning noise and image_latents for blending. + + Returns: + Tuple of (latents, noise, image_latents) where: + - latents: Noised image latents for denoising + - noise: The noise tensor used for blending + - image_latents: Clean image latents for blending + """ + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is not None: + # Generate noise for blending even if latents are provided + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # Encode image for blending + image = image.to(device=device, dtype=dtype) + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + image_latents = torch.cat([image_latents] * (batch_size // image_latents.shape[0]), dim=0) + return latents.to(device=device, dtype=dtype), noise, image_latents + + # Encode the input image + image = image.to(device=device, dtype=dtype) + if image.shape[1] != num_channels_latents: + if isinstance(generator, list): + image_latents = [ + retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) + for i in range(image.shape[0]) + ] + image_latents = torch.cat(image_latents, dim=0) + else: + image_latents = retrieve_latents(self.vae.encode(image), generator=generator) + + # Apply scaling (inverse of decoding: decode does latents/scaling_factor + shift_factor) + image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor + else: + image_latents = image + + # Handle batch size expansion + if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0: + additional_image_per_prompt = batch_size // image_latents.shape[0] + image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0) + elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts." + ) + + # Generate noise for both initial noising and later blending + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # Add noise using flow matching scale_noise + latents = self.scheduler.scale_noise(image_latents, timestep, noise) + + return latents, noise, image_latents + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + def check_inputs( + self, + prompt, + image, + mask_image, + strength, + height, + width, + output_type, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if strength < 0 or strength > 1: + raise ValueError(f"The value of strength should be in [0.0, 1.0] but is {strength}") + + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if image is None: + raise ValueError("`image` input cannot be undefined for inpainting.") + + if mask_image is None: + raise ValueError("`mask_image` input cannot be undefined for inpainting.") + + if output_type not in ["latent", "pil", "np", "pt"]: + raise ValueError(f"`output_type` must be one of 'latent', 'pil', 'np', or 'pt', but got {output_type}") + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + image: PipelineImageInput = None, + mask_image: PipelineImageInput = None, + masked_image_latents: Optional[torch.FloatTensor] = None, + strength: float = 1.0, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + cfg_normalization: bool = False, + cfg_truncation: float = 1.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[List[torch.FloatTensor]] = None, + negative_prompt_embeds: Optional[List[torch.FloatTensor]] = None, + output_type: str = "pil", + return_dict: bool = True, + joint_attention_kwargs: Optional[Dict[str, Any]] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for inpainting. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a + list of tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or + a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. + mask_image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`): + `Image`, numpy array or tensor representing a mask image for inpainting. White pixels (value 1) in the + mask will be inpainted, black pixels (value 0) will be preserved from the original image. + masked_image_latents (`torch.FloatTensor`, *optional*): + Pre-encoded masked image latents. If provided, the masked image encoding step will be skipped. + strength (`float`, *optional*, defaults to 1.0): + Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a + starting point and more noise is added the higher the `strength`. The number of denoising steps depends + on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising + process runs for the full number of iterations specified in `num_inference_steps`. A value of 1 + essentially ignores `image` in the masked region. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. If not provided, uses the input image height. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. If not provided, uses the input image width. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`List[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + cfg_normalization (`bool`, *optional*, defaults to False): + Whether to apply configuration normalization. + cfg_truncation (`float`, *optional*, defaults to 1.0): + The truncation value for configuration. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`List[torch.FloatTensor]`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`List[torch.FloatTensor]`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain + tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + # 1. Check inputs + self.check_inputs( + prompt=prompt, + image=image, + mask_image=mask_image, + strength=strength, + height=height, + width=width, + output_type=output_type, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, + ) + + # 2. Preprocess image and mask + init_image = self.image_processor.preprocess(image) + init_image = init_image.to(dtype=torch.float32) + + # Get dimensions from the preprocessed image if not specified + if height is None: + height = init_image.shape[-2] + if width is None: + width = init_image.shape[-1] + + vae_scale = self.vae_scale_factor * 2 + if height % vae_scale != 0: + raise ValueError( + f"Height must be divisible by {vae_scale} (got {height}). " + f"Please adjust the height to a multiple of {vae_scale}." + ) + if width % vae_scale != 0: + raise ValueError( + f"Width must be divisible by {vae_scale} (got {width}). " + f"Please adjust the width to a multiple of {vae_scale}." + ) + + # Preprocess mask + mask = self.mask_processor.preprocess(mask_image, height=height, width=width) + + device = self._execution_device + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._cfg_normalization = cfg_normalization + self._cfg_truncation = cfg_truncation + + # 3. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is not None and prompt is None: + if self.do_classifier_free_guidance and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided without `prompt`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + else: + ( + prompt_embeds, + negative_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + max_sequence_length=max_sequence_length, + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.in_channels + + # Repeat prompt_embeds for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] + if self.do_classifier_free_guidance and negative_prompt_embeds: + negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] + + actual_batch_size = batch_size * num_images_per_prompt + + # Calculate latent dimensions for image_seq_len + latent_height = 2 * (int(height) // (self.vae_scale_factor * 2)) + latent_width = 2 * (int(width) // (self.vae_scale_factor * 2)) + image_seq_len = (latent_height // 2) * (latent_width // 2) + + # 5. Prepare timesteps + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if sigmas is None: + sigmas = get_default_z_image_sigmas(num_inference_steps) + scheduler_kwargs = {"mu": mu} + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + + # 6. Adjust timesteps based on strength + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device) + if num_inference_steps < 1: + raise ValueError( + f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline " + f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." + ) + latent_timestep = timesteps[:1].repeat(actual_batch_size) + + # 7. Prepare latents from image (returns noise and image_latents for blending) + latents, noise, image_latents = self.prepare_latents( + init_image, + latent_timestep, + actual_batch_size, + num_channels_latents, + height, + width, + prompt_embeds[0].dtype, + device, + generator, + latents, + ) + + # 8. Prepare mask and masked image latents + # Create masked image: preserve only unmasked regions (mask=0) + if masked_image_latents is None: + masked_image = init_image * (mask < 0.5) + else: + masked_image = None # Will use provided masked_image_latents + + mask, masked_image_latents = self.prepare_mask_latents( + mask, + masked_image if masked_image is not None else init_image, + actual_batch_size, + height, + width, + prompt_embeds[0].dtype, + device, + generator, + ) + + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 9. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + timestep = (1000 - timestep) / 1000 + # Normalized time for time-aware config (0 at start, 1 at end) + t_norm = timestep[0].item() + + # Handle cfg truncation + current_guidance_scale = self.guidance_scale + if ( + self.do_classifier_free_guidance + and self._cfg_truncation is not None + and float(self._cfg_truncation) <= 1 + ): + if t_norm > self._cfg_truncation: + current_guidance_scale = 0.0 + + # Run CFG only if configured AND scale is non-zero + apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 + + if apply_cfg: + latents_typed = latents.to(self.transformer.dtype) + latent_model_input = latents_typed.repeat(2, 1, 1, 1) + prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds + timestep_model_input = timestep.repeat(2) + else: + latent_model_input = latents.to(self.transformer.dtype) + prompt_embeds_model_input = prompt_embeds + timestep_model_input = timestep + + latent_model_input = latent_model_input.unsqueeze(2) + latent_model_input_list = list(latent_model_input.unbind(dim=0)) + + model_out_list = self.transformer( + latent_model_input_list, + timestep_model_input, + prompt_embeds_model_input, + )[0] + + if apply_cfg: + # Perform CFG + pos_out = model_out_list[:actual_batch_size] + neg_out = model_out_list[actual_batch_size:] + + noise_pred = [] + for j in range(actual_batch_size): + pos = pos_out[j].float() + neg = neg_out[j].float() + + pred = pos + current_guidance_scale * (pos - neg) + + # Renormalization + if self._cfg_normalization and float(self._cfg_normalization) > 0.0: + ori_pos_norm = torch.linalg.vector_norm(pos) + new_pos_norm = torch.linalg.vector_norm(pred) + max_new_norm = ori_pos_norm * float(self._cfg_normalization) + if new_pos_norm > max_new_norm: + pred = pred * (max_new_norm / new_pos_norm) + + noise_pred.append(pred) + + noise_pred = torch.stack(noise_pred, dim=0) + else: + noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) + + noise_pred = noise_pred.squeeze(2) + noise_pred = -noise_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] + assert latents.dtype == torch.float32 + + # Inpainting blend: combine denoised latents with original image latents + init_latents_proper = image_latents + + # Re-scale original latents to current noise level for proper blending + if i < len(timesteps) - 1: + noise_timestep = timesteps[i + 1] + init_latents_proper = self.scheduler.scale_noise( + init_latents_proper, torch.tensor([noise_timestep]), noise + ) + + # Blend: mask=1 for inpaint region (use denoised), mask=0 for preserve region (use original) + latents = (1 - mask) * init_latents_proper + mask * latents + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + mask = callback_outputs.pop("mask", mask) + masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if XLA_AVAILABLE: + xm.mark_step() + + if output_type == "latent": + image = latents + + else: + latents = latents.to(self.vae.dtype) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ZImagePipelineOutput(images=image) diff --git a/diffusers/pipelines/z_image/pipeline_z_image_omni.py b/diffusers/pipelines/z_image/pipeline_z_image_omni.py new file mode 100644 index 0000000000000000000000000000000000000000..50776ceaf34dce0a928c7a1b05701210118d4dce --- /dev/null +++ b/diffusers/pipelines/z_image/pipeline_z_image_omni.py @@ -0,0 +1,749 @@ +# Copyright 2025 Alibaba Z-Image Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable + +import PIL +import torch +from transformers import AutoTokenizer, PreTrainedModel, Siglip2ImageProcessorFast, Siglip2VisionModel + +from ...loaders import FromSingleFileMixin, ZImageLoraLoaderMixin +from ...models.autoencoders import AutoencoderKL +from ...models.transformers import ZImageTransformer2DModel +from ...pipelines.pipeline_utils import DiffusionPipeline +from ...schedulers import FlowMatchEulerDiscreteScheduler +from ...utils import logging, replace_example_docstring +from ...utils.torch_utils import randn_tensor +from ..flux2.image_processor import Flux2ImageProcessor +from .pipeline_output import ZImagePipelineOutput + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import ZImageOmniPipeline + + >>> pipe = ZImageOmniPipeline.from_pretrained("Z-a-o/Z-Image-Turbo", torch_dtype=torch.bfloat16) + >>> pipe.to("cuda") + + >>> # Optionally, set the attention backend to flash-attn 2 or 3, default is SDPA in PyTorch. + >>> # (1) Use flash attention 2 + >>> # pipe.transformer.set_attention_backend("flash") + >>> # (2) Use flash attention 3 + >>> # pipe.transformer.set_attention_backend("_flash_3") + + >>> prompt = "一幅为名为“造相「Z-IMAGE-TURBO」”的项目设计的创意海报。画面巧妙地将文字概念视觉化：一辆复古蒸汽小火车化身为巨大的拉链头，正拉开厚厚的冬日积雪，展露出一个生机盎然的春天。" + >>> image = pipe( + ... prompt, + ... height=1024, + ... width=1024, + ... num_inference_steps=8, + ... guidance_scale=0.0, + ... generator=torch.Generator("cuda").manual_seed(42), + ... ).images[0] + >>> image.save("zimage.png") + ``` +""" + + +# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift +def calculate_shift( + image_seq_len, + base_seq_len: int = 256, + max_seq_len: int = 4096, + base_shift: float = 0.5, + max_shift: float = 1.15, +): + m = (max_shift - base_shift) / (max_seq_len - base_seq_len) + b = base_shift - m * base_seq_len + mu = image_seq_len * m + b + return mu + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps +def retrieve_timesteps( + scheduler, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + **kwargs, +): + r""" + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` + must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed, + `num_inference_steps` and `sigmas` must be `None`. + sigmas (`list[float]`, *optional*): + Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed, + `num_inference_steps` and `timesteps` must be `None`. + + Returns: + `tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values") + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + elif sigmas is not None: + accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accept_sigmas: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" sigmas schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +# Copied from diffusers.pipelines.z_image.pipeline_z_image.get_default_z_image_sigmas +def get_default_z_image_sigmas(num_inference_steps: int) -> list[float]: + return torch.linspace(1.0, 1 / num_inference_steps, num_inference_steps).tolist() + + +class ZImageOmniPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin): + model_cpu_offload_seq = "text_encoder->transformer->vae" + _optional_components = [] + _callback_tensor_inputs = ["latents", "prompt_embeds"] + + def __init__( + self, + scheduler: FlowMatchEulerDiscreteScheduler, + vae: AutoencoderKL, + text_encoder: PreTrainedModel, + tokenizer: AutoTokenizer, + transformer: ZImageTransformer2DModel, + siglip: Siglip2VisionModel, + siglip_processor: Siglip2ImageProcessorFast, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + scheduler=scheduler, + transformer=transformer, + siglip=siglip, + siglip_processor=siglip_processor, + ) + self.vae_scale_factor = ( + 2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8 + ) + # self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2) + + def encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + do_classifier_free_guidance: bool = True, + negative_prompt: str | list[str] | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: torch.FloatTensor | None = None, + max_sequence_length: int = 512, + num_condition_images: int = 0, + ): + prompt = [prompt] if isinstance(prompt, str) else prompt + prompt_embeds = self._encode_prompt( + prompt=prompt, + device=device, + prompt_embeds=prompt_embeds, + max_sequence_length=max_sequence_length, + num_condition_images=num_condition_images, + ) + + if do_classifier_free_guidance: + if negative_prompt is None: + negative_prompt = ["" for _ in prompt] + else: + negative_prompt = [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt + assert len(prompt) == len(negative_prompt) + negative_prompt_embeds = self._encode_prompt( + prompt=negative_prompt, + device=device, + prompt_embeds=negative_prompt_embeds, + max_sequence_length=max_sequence_length, + num_condition_images=num_condition_images, + ) + else: + negative_prompt_embeds = [] + return prompt_embeds, negative_prompt_embeds + + def _encode_prompt( + self, + prompt: str | list[str], + device: torch.device | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + max_sequence_length: int = 512, + num_condition_images: int = 0, + ) -> list[torch.FloatTensor]: + device = device or self._execution_device + + if prompt_embeds is not None: + return prompt_embeds + + if isinstance(prompt, str): + prompt = [prompt] + + for i, prompt_item in enumerate(prompt): + if num_condition_images == 0: + prompt[i] = ["<|im_start|>user\n" + prompt_item + "<|im_end|>\n<|im_start|>assistant\n"] + elif num_condition_images > 0: + prompt_list = ["<|im_start|>user\n<|vision_start|>"] + prompt_list += ["<|vision_end|><|vision_start|>"] * (num_condition_images - 1) + prompt_list += ["<|vision_end|>" + prompt_item + "<|im_end|>\n<|im_start|>assistant\n<|vision_start|>"] + prompt_list += ["<|vision_end|><|im_end|>"] + prompt[i] = prompt_list + + flattened_prompt = [] + prompt_list_lengths = [] + + for i in range(len(prompt)): + prompt_list_lengths.append(len(prompt[i])) + flattened_prompt.extend(prompt[i]) + + text_inputs = self.tokenizer( + flattened_prompt, + padding="max_length", + max_length=max_sequence_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids.to(device) + prompt_masks = text_inputs.attention_mask.to(device).bool() + + prompt_embeds = self.text_encoder( + input_ids=text_input_ids, + attention_mask=prompt_masks, + output_hidden_states=True, + ).hidden_states[-2] + + embeddings_list = [] + start_idx = 0 + for i in range(len(prompt_list_lengths)): + batch_embeddings = [] + end_idx = start_idx + prompt_list_lengths[i] + for j in range(start_idx, end_idx): + batch_embeddings.append(prompt_embeds[j][prompt_masks[j]]) + embeddings_list.append(batch_embeddings) + start_idx = end_idx + + return embeddings_list + + def prepare_latents( + self, + batch_size, + num_channels_latents, + height, + width, + dtype, + device, + generator, + latents=None, + ): + height = 2 * (int(height) // (self.vae_scale_factor * 2)) + width = 2 * (int(width) // (self.vae_scale_factor * 2)) + + shape = (batch_size, num_channels_latents, height, width) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + if latents.shape != shape: + raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") + latents = latents.to(device) + return latents + + def prepare_image_latents( + self, + images: list[torch.Tensor], + batch_size, + device, + dtype, + ): + image_latents = [] + for image in images: + image = image.to(device=device, dtype=dtype) + image_latent = ( + self.vae.encode(image.bfloat16()).latent_dist.mode()[0] - self.vae.config.shift_factor + ) * self.vae.config.scaling_factor + image_latent = image_latent.unsqueeze(1).to(dtype) + image_latents.append(image_latent) # (16, 128, 128) + + # image_latents = [image_latents] * batch_size + image_latents = [image_latents.copy() for _ in range(batch_size)] + + return image_latents + + def prepare_siglip_embeds( + self, + images: list[torch.Tensor], + batch_size, + device, + dtype, + ): + siglip_embeds = [] + for image in images: + siglip_inputs = self.siglip_processor(images=[image], return_tensors="pt").to(device) + shape = siglip_inputs.spatial_shapes[0] + hidden_state = self.siglip(**siglip_inputs).last_hidden_state + B, N, C = hidden_state.shape + hidden_state = hidden_state[:, : shape[0] * shape[1]] + hidden_state = hidden_state.view(shape[0], shape[1], C) + siglip_embeds.append(hidden_state.to(dtype)) + + # siglip_embeds = [siglip_embeds] * batch_size + siglip_embeds = [siglip_embeds.copy() for _ in range(batch_size)] + + return siglip_embeds + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 0 + + @property + def joint_attention_kwargs(self): + return self._joint_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + image: list[PIL.Image.Image, PIL.Image.Image] | None = None, + prompt: str | list[str] = None, + height: int | None = None, + width: int | None = None, + num_inference_steps: int = 50, + sigmas: list[float] | None = None, + guidance_scale: float = 5.0, + cfg_normalization: bool = False, + cfg_truncation: float = 1.0, + negative_prompt: str | list[str] | None = None, + num_images_per_prompt: int | None = 1, + generator: torch.Generator | list[torch.Generator] | None = None, + latents: torch.FloatTensor | None = None, + prompt_embeds: list[torch.FloatTensor] | None = None, + negative_prompt_embeds: list[torch.FloatTensor] | None = None, + output_type: str | None = "pil", + return_dict: bool = True, + joint_attention_kwargs: dict[str, Any] | None = None, + callback_on_step_end: Callable[[int, int], None] | None = None, + callback_on_step_end_tensor_inputs: list[str] = ["latents"], + max_sequence_length: int = 512, + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `list[torch.Tensor]`, `list[PIL.Image.Image]`, or `list[np.ndarray]`): + `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both + numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list + or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a + list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image + latents as `image`, but if passing latents directly it is not encoded again. + prompt (`str` or `list[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + height (`int`, *optional*, defaults to 1024): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 1024): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + sigmas (`list[float]`, *optional*): + Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in + their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed + will be used. + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + cfg_normalization (`bool`, *optional*, defaults to False): + Whether to apply configuration normalization. + cfg_truncation (`float`, *optional*, defaults to 1.0): + The truncation value for configuration. + negative_prompt (`str` or `list[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will be generated by sampling using the supplied random `generator`. + prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`list[torch.FloatTensor]`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.ZImagePipelineOutput`] instead of a plain + tuple. + joint_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + max_sequence_length (`int`, *optional*, defaults to 512): + Maximum sequence length to use with the `prompt`. + + Examples: + + Returns: + [`~pipelines.z_image.ZImagePipelineOutput`] or `tuple`: [`~pipelines.z_image.ZImagePipelineOutput`] if + `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the + generated images. + """ + + if image is not None and not isinstance(image, list): + image = [image] + num_condition_images = len(image) if image is not None else 0 + + device = self._execution_device + + self._guidance_scale = guidance_scale + self._joint_attention_kwargs = joint_attention_kwargs + self._interrupt = False + self._cfg_normalization = cfg_normalization + self._cfg_truncation = cfg_truncation + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = len(prompt_embeds) + + # If prompt_embeds is provided and prompt is None, skip encoding + if prompt_embeds is not None and prompt is None: + if self.do_classifier_free_guidance and negative_prompt_embeds is None: + raise ValueError( + "When `prompt_embeds` is provided without `prompt`, " + "`negative_prompt_embeds` must also be provided for classifier-free guidance." + ) + else: + ( + prompt_embeds, + negative_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + negative_prompt=negative_prompt, + do_classifier_free_guidance=self.do_classifier_free_guidance, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + device=device, + max_sequence_length=max_sequence_length, + num_condition_images=num_condition_images, + ) + + # 3. Process condition images. Copied from diffusers.pipelines.flux2.pipeline_flux2 + condition_images = [] + resized_images = [] + if image is not None: + for img in image: + self.image_processor.check_image_input(img) + for img in image: + image_width, image_height = img.size + if image_width * image_height > 1024 * 1024: + if height is not None and width is not None: + img = self.image_processor._resize_to_target_area(img, height * width) + else: + img = self.image_processor._resize_to_target_area(img, 1024 * 1024) + image_width, image_height = img.size + resized_images.append(img) + + multiple_of = self.vae_scale_factor * 2 + image_width = (image_width // multiple_of) * multiple_of + image_height = (image_height // multiple_of) * multiple_of + img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop") + condition_images.append(img) + + if len(condition_images) > 0: + height = height or image_height + width = width or image_width + + else: + height = height or 1024 + width = width or 1024 + + vae_scale = self.vae_scale_factor * 2 + if height % vae_scale != 0: + raise ValueError( + f"Height must be divisible by {vae_scale} (got {height}). " + f"Please adjust the height to a multiple of {vae_scale}." + ) + if width % vae_scale != 0: + raise ValueError( + f"Width must be divisible by {vae_scale} (got {width}). " + f"Please adjust the width to a multiple of {vae_scale}." + ) + + # 4. Prepare latent variables + num_channels_latents = self.transformer.in_channels + + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + torch.float32, + device, + generator, + latents, + ) + + condition_latents = self.prepare_image_latents( + images=condition_images, + batch_size=batch_size * num_images_per_prompt, + device=device, + dtype=torch.float32, + ) + condition_latents = [[lat.to(self.transformer.dtype) for lat in lats] for lats in condition_latents] + if self.do_classifier_free_guidance: + negative_condition_latents = [[lat.clone() for lat in batch] for batch in condition_latents] + + condition_siglip_embeds = self.prepare_siglip_embeds( + images=resized_images, + batch_size=batch_size * num_images_per_prompt, + device=device, + dtype=torch.float32, + ) + condition_siglip_embeds = [[se.to(self.transformer.dtype) for se in sels] for sels in condition_siglip_embeds] + if self.do_classifier_free_guidance: + negative_condition_siglip_embeds = [[se.clone() for se in batch] for batch in condition_siglip_embeds] + + # Repeat prompt_embeds for num_images_per_prompt + if num_images_per_prompt > 1: + prompt_embeds = [pe for pe in prompt_embeds for _ in range(num_images_per_prompt)] + if self.do_classifier_free_guidance and negative_prompt_embeds: + negative_prompt_embeds = [npe for npe in negative_prompt_embeds for _ in range(num_images_per_prompt)] + + condition_siglip_embeds = [None if sels == [] else sels + [None] for sels in condition_siglip_embeds] + if self.do_classifier_free_guidance: + negative_condition_siglip_embeds = [ + None if sels == [] else sels + [None] for sels in negative_condition_siglip_embeds + ] + + actual_batch_size = batch_size * num_images_per_prompt + image_seq_len = (latents.shape[2] // 2) * (latents.shape[3] // 2) + + # 5. Prepare timesteps + mu = calculate_shift( + image_seq_len, + self.scheduler.config.get("base_image_seq_len", 256), + self.scheduler.config.get("max_image_seq_len", 4096), + self.scheduler.config.get("base_shift", 0.5), + self.scheduler.config.get("max_shift", 1.15), + ) + if sigmas is None: + sigmas = get_default_z_image_sigmas(num_inference_steps) + scheduler_kwargs = {"mu": mu} + timesteps, num_inference_steps = retrieve_timesteps( + self.scheduler, + num_inference_steps, + device, + sigmas=sigmas, + **scheduler_kwargs, + ) + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + self._num_timesteps = len(timesteps) + + # 6. Denoising loop + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timestep = t.expand(latents.shape[0]) + timestep = (1000 - timestep) / 1000 + # Normalized time for time-aware config (0 at start, 1 at end) + t_norm = timestep[0].item() + + # Handle cfg truncation + current_guidance_scale = self.guidance_scale + if ( + self.do_classifier_free_guidance + and self._cfg_truncation is not None + and float(self._cfg_truncation) <= 1 + ): + if t_norm > self._cfg_truncation: + current_guidance_scale = 0.0 + + # Run CFG only if configured AND scale is non-zero + apply_cfg = self.do_classifier_free_guidance and current_guidance_scale > 0 + + if apply_cfg: + latents_typed = latents.to(self.transformer.dtype) + latent_model_input = latents_typed.repeat(2, 1, 1, 1) + prompt_embeds_model_input = prompt_embeds + negative_prompt_embeds + condition_latents_model_input = condition_latents + negative_condition_latents + condition_siglip_embeds_model_input = condition_siglip_embeds + negative_condition_siglip_embeds + timestep_model_input = timestep.repeat(2) + else: + latent_model_input = latents.to(self.transformer.dtype) + prompt_embeds_model_input = prompt_embeds + condition_latents_model_input = condition_latents + condition_siglip_embeds_model_input = condition_siglip_embeds + timestep_model_input = timestep + + latent_model_input = latent_model_input.unsqueeze(2) + latent_model_input_list = list(latent_model_input.unbind(dim=0)) + + # Combine condition latents with target latent + current_batch_size = len(latent_model_input_list) + x_combined = [ + condition_latents_model_input[i] + [latent_model_input_list[i]] for i in range(current_batch_size) + ] + # Create noise mask: 0 for condition images (clean), 1 for target image (noisy) + image_noise_mask = [ + [0] * len(condition_latents_model_input[i]) + [1] for i in range(current_batch_size) + ] + + model_out_list = self.transformer( + x=x_combined, + t=timestep_model_input, + cap_feats=prompt_embeds_model_input, + siglip_feats=condition_siglip_embeds_model_input, + image_noise_mask=image_noise_mask, + return_dict=False, + )[0] + + if apply_cfg: + # Perform CFG + pos_out = model_out_list[:actual_batch_size] + neg_out = model_out_list[actual_batch_size:] + + noise_pred = [] + for j in range(actual_batch_size): + pos = pos_out[j].float() + neg = neg_out[j].float() + + pred = pos + current_guidance_scale * (pos - neg) + + # Renormalization + if self._cfg_normalization and float(self._cfg_normalization) > 0.0: + ori_pos_norm = torch.linalg.vector_norm(pos) + new_pos_norm = torch.linalg.vector_norm(pred) + max_new_norm = ori_pos_norm * float(self._cfg_normalization) + if new_pos_norm > max_new_norm: + pred = pred * (max_new_norm / new_pos_norm) + + noise_pred.append(pred) + + noise_pred = torch.stack(noise_pred, dim=0) + else: + noise_pred = torch.stack([t.float() for t in model_out_list], dim=0) + + noise_pred = noise_pred.squeeze(2) + noise_pred = -noise_pred + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred.to(torch.float32), t, latents, return_dict=False)[0] + assert latents.dtype == torch.float32 + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + + if output_type == "latent": + image = latents + + else: + latents = latents.to(self.vae.dtype) + latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor + + image = self.vae.decode(latents, return_dict=False)[0] + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image,) + + return ZImagePipelineOutput(images=image) diff --git a/diffusers/py.typed b/diffusers/py.typed new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/diffusers/quantizers/__init__.py b/diffusers/quantizers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..048934945abce3155bbec7b8150cf7a4979fc82a --- /dev/null +++ b/diffusers/quantizers/__init__.py @@ -0,0 +1,18 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from .auto import DiffusersAutoQuantizer +from .base import DiffusersQuantizer +from .pipe_quant_config import PipelineQuantizationConfig diff --git a/diffusers/quantizers/auto.py b/diffusers/quantizers/auto.py new file mode 100644 index 0000000000000000000000000000000000000000..ea6caf91ab80602aa1faa42c2f4f7934eec25613 --- /dev/null +++ b/diffusers/quantizers/auto.py @@ -0,0 +1,174 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Adapted from +https://github.com/huggingface/transformers/blob/c409cd81777fb27aadc043ed3d8339dbc020fb3b/src/transformers/quantizers/auto.py +""" + +import warnings + +from .autoround import AutoRoundQuantizer +from .bitsandbytes import BnB4BitDiffusersQuantizer, BnB8BitDiffusersQuantizer +from .gguf import GGUFQuantizer +from .modelopt import NVIDIAModelOptQuantizer +from .nunchaku import NunchakuLiteQuantizer +from .quantization_config import ( + AutoRoundConfig, + BitsAndBytesConfig, + GGUFQuantizationConfig, + NunchakuLiteQuantizationConfig, + NVIDIAModelOptConfig, + QuantizationConfigMixin, + QuantizationMethod, + QuantoConfig, + SDNQConfig, + TorchAoConfig, +) +from .quanto import QuantoQuantizer +from .sdnq import SDNQQuantizer +from .torchao import TorchAoHfQuantizer + + +AUTO_QUANTIZER_MAPPING = { + "bitsandbytes_4bit": BnB4BitDiffusersQuantizer, + "bitsandbytes_8bit": BnB8BitDiffusersQuantizer, + "gguf": GGUFQuantizer, + "quanto": QuantoQuantizer, + "torchao": TorchAoHfQuantizer, + "modelopt": NVIDIAModelOptQuantizer, + "auto-round": AutoRoundQuantizer, + "nunchaku_lite": NunchakuLiteQuantizer, + "sdnq": SDNQQuantizer, +} + +AUTO_QUANTIZATION_CONFIG_MAPPING = { + "bitsandbytes_4bit": BitsAndBytesConfig, + "bitsandbytes_8bit": BitsAndBytesConfig, + "gguf": GGUFQuantizationConfig, + "quanto": QuantoConfig, + "torchao": TorchAoConfig, + "modelopt": NVIDIAModelOptConfig, + "auto-round": AutoRoundConfig, + "nunchaku_lite": NunchakuLiteQuantizationConfig, + "sdnq": SDNQConfig, +} + + +class DiffusersAutoQuantizer: + """ + The auto diffusers quantizer class that takes care of automatically instantiating to the correct + `DiffusersQuantizer` given the `QuantizationConfig`. + """ + + @classmethod + def from_dict(cls, quantization_config_dict: dict): + quant_method = quantization_config_dict.get("quant_method", None) + # We need a special care for bnb models to make sure everything is BC .. + if quantization_config_dict.get("load_in_8bit", False) or quantization_config_dict.get("load_in_4bit", False): + suffix = "_4bit" if quantization_config_dict.get("load_in_4bit", False) else "_8bit" + quant_method = QuantizationMethod.BITS_AND_BYTES + suffix + elif quant_method is None: + raise ValueError( + "The model's quantization config from the arguments has no `quant_method` attribute. Make sure that the model has been correctly quantized" + ) + + if quant_method not in AUTO_QUANTIZATION_CONFIG_MAPPING.keys(): + raise ValueError( + f"Unknown quantization type, got {quant_method} - supported types are:" + f" {list(AUTO_QUANTIZER_MAPPING.keys())}" + ) + + target_cls = AUTO_QUANTIZATION_CONFIG_MAPPING[quant_method] + return target_cls.from_dict(quantization_config_dict) + + @classmethod + def from_config(cls, quantization_config: QuantizationConfigMixin | dict, **kwargs): + # Convert it to a QuantizationConfig if the q_config is a dict + if isinstance(quantization_config, dict): + quantization_config = cls.from_dict(quantization_config) + + quant_method = quantization_config.quant_method + + # Again, we need a special care for bnb as we have a single quantization config + # class for both 4-bit and 8-bit quantization + if quant_method == QuantizationMethod.BITS_AND_BYTES: + if quantization_config.load_in_8bit: + quant_method += "_8bit" + else: + quant_method += "_4bit" + + if quant_method not in AUTO_QUANTIZER_MAPPING.keys(): + raise ValueError( + f"Unknown quantization type, got {quant_method} - supported types are:" + f" {list(AUTO_QUANTIZER_MAPPING.keys())}" + ) + + target_cls = AUTO_QUANTIZER_MAPPING[quant_method] + return target_cls(quantization_config, **kwargs) + + @classmethod + def from_pretrained(cls, pretrained_model_name_or_path, **kwargs): + model_config = cls.load_config(pretrained_model_name_or_path, **kwargs) + if getattr(model_config, "quantization_config", None) is None: + raise ValueError( + f"Did not found a `quantization_config` in {pretrained_model_name_or_path}. Make sure that the model is correctly quantized." + ) + quantization_config_dict = model_config.quantization_config + quantization_config = cls.from_dict(quantization_config_dict) + # Update with potential kwargs that are passed through from_pretrained. + quantization_config.update(kwargs) + + return cls.from_config(quantization_config) + + @classmethod + def merge_quantization_configs( + cls, + quantization_config: dict | QuantizationConfigMixin, + quantization_config_from_args: QuantizationConfigMixin | None, + ): + """ + handles situations where both quantization_config from args and quantization_config from model config are + present. + """ + if quantization_config_from_args is not None: + warning_msg = ( + "You passed `quantization_config` or equivalent parameters to `from_pretrained` but the model you're loading" + " already has a `quantization_config` attribute. The `quantization_config` from the model will be used." + ) + else: + warning_msg = "" + + if isinstance(quantization_config, dict): + quantization_config = cls.from_dict(quantization_config) + + if isinstance(quantization_config, NVIDIAModelOptConfig): + quantization_config.check_model_patching() + + if quantization_config_from_args is not None and isinstance(quantization_config, AutoRoundConfig): + # For AutoRound, allow overriding fields like `backend` from user args, + # since the model config may store a default value (e.g. backend="auto"). + for key, value in quantization_config_from_args.__dict__.items(): + if key in ("quant_method",): + continue + if hasattr(quantization_config, key) and getattr(quantization_config, key) != value: + warnings.warn( + f"Overriding `{key}` in the model's quantization_config with value {value!r} " + f"from the user-provided `quantization_config`." + ) + setattr(quantization_config, key, value) + + if warning_msg != "": + warnings.warn(warning_msg) + + return quantization_config diff --git a/diffusers/quantizers/autoround/__init__.py b/diffusers/quantizers/autoround/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..2fe2083d4a5fcef42767f6976953ee432e2f8928 --- /dev/null +++ b/diffusers/quantizers/autoround/__init__.py @@ -0,0 +1 @@ +from .autoround_quantizer import AutoRoundQuantizer diff --git a/diffusers/quantizers/autoround/autoround_quantizer.py b/diffusers/quantizers/autoround/autoround_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..f80563fed4069b8db3fc255d3ec0f8e4c39bf711 --- /dev/null +++ b/diffusers/quantizers/autoround/autoround_quantizer.py @@ -0,0 +1,122 @@ +# Copyright 2025 The Intel and The HuggingFace Inc. teams. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import TYPE_CHECKING + +from ...utils import is_auto_round_available, logging +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +logger = logging.get_logger(__name__) + + +class AutoRoundQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for AutoRound (https://github.com/intel/auto-round). + + AutoRound is a weight-only quantization method that uses sign gradient descent to jointly optimize rounding values + and min-max ranges for weights. It supports W4A16 (4-bit weight, 16-bit activation) quantization for efficient + inference. + + This quantizer only supports loading pre-quantized AutoRound models. On-the-fly quantization (calibration) is not + supported through this interface. + """ + + # AutoRound requires data calibration — we only support loading pre-quantized checkpoints. + requires_calibration = True + required_packages = ["auto_round"] + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + def validate_environment(self, *args, **kwargs): + """ + Validates that the auto-round library (>= 0.5) is installed and captures the device_map for later use during + model conversion. + """ + self.device_map = kwargs.get("device_map", None) + if not is_auto_round_available(): + raise ImportError( + "Loading an AutoRound quantized model requires the auto-round library " + "(`pip install 'auto-round>=0.13.0'`)" + ) + if not self.pre_quantized: + raise ValueError( + "AutoRound quantizer in diffusers only supports loading pre-quantized models. " + "To quantize a model from scratch, use the AutoRound CLI or Python API " + "(https://github.com/intel/auto-round) directly, then load the result with Diffusers." + ) + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + """ + Replaces target nn.Linear layers with AutoRound's quantized QuantLinear layers before weights are loaded from + the checkpoint. + + Uses `auto_round.inference.convert_model.convert_hf_model` which: + - Inspects the model architecture and the quantization config (bits, group_size, sym, backend). + - Replaces eligible nn.Linear modules with the appropriate QuantLinear variant (the packed-weight layer that + stores qweight, scales, qzeros). + - Returns the converted model and a set of used backend names. + + `infer_target_device` resolves the device_map into a single target device string that AutoRound uses to select + the correct kernel backend (e.g. "cuda", "cpu"). + """ + from auto_round.inference.convert_model import convert_hf_model, infer_target_device + + target_device = infer_target_device(self.device_map) + model, used_backends = convert_hf_model(model, target_device) + self.used_backends = used_backends + + def _process_model_after_weight_loading(self, model, **kwargs): + """ + Finalizes the model after all quantized weights (qweight, scales, qzeros, etc.) have been loaded into the + QuantLinear layers. + + Uses `auto_round.inference.convert_model.post_init` which: + - Performs backend-specific finalization (e.g. repacking weights into the kernel's expected memory layout, + moving buffers to the correct device). + - Freezes quantized parameters (requires_grad=False). + - Prepares the model for inference. + + """ + from auto_round.inference.convert_model import post_init + + post_init(model, self.used_backends) + + return model + + @property + def is_trainable(self) -> bool: + """AutoRound W4A16 pre-quantized models do not support training.""" + return False + + @property + def is_serializable(self): + """AutoRound quantized models can be serialized (the quantization config may be + updated by the backend, e.g. for GPTQ/AWQ-compatible formats).""" + return True + + @property + def is_compileable(self) -> bool: + return True diff --git a/diffusers/quantizers/base.py b/diffusers/quantizers/base.py new file mode 100644 index 0000000000000000000000000000000000000000..29dd2ba7f9ab4b49f841f71bdf36df884cf05716 --- /dev/null +++ b/diffusers/quantizers/base.py @@ -0,0 +1,265 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Adapted from +https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/quantizers/base.py +""" + +from abc import ABC, abstractmethod +from typing import TYPE_CHECKING, Any + +from ..utils import is_torch_available +from .quantization_config import QuantizationConfigMixin + + +if TYPE_CHECKING: + from ..models.modeling_utils import ModelMixin + +if is_torch_available(): + import torch + + +class DiffusersQuantizer(ABC): + """ + Abstract class of the HuggingFace quantizer. Supports for now quantizing HF diffusers models for inference and/or + quantization. This class is used only for diffusers.models.modeling_utils.ModelMixin.from_pretrained and cannot be + easily used outside the scope of that method yet. + + Attributes + quantization_config (`diffusers.quantizers.quantization_config.QuantizationConfigMixin`): + The quantization config that defines the quantization parameters of your model that you want to quantize. + modules_to_not_convert (`list[str]`, *optional*): + The list of module names to not convert when quantizing the model. + required_packages (`list[str]`, *optional*): + The list of required pip packages to install prior to using the quantizer + requires_calibration (`bool`): + Whether the quantization method requires to calibrate the model before using it. + """ + + requires_calibration = False + required_packages = None + + def __init__(self, quantization_config: QuantizationConfigMixin, **kwargs): + self.quantization_config = quantization_config + + # -- Handle extra kwargs below -- + self.modules_to_not_convert = kwargs.pop("modules_to_not_convert", []) + self.pre_quantized = kwargs.pop("pre_quantized", True) + + if not self.pre_quantized and self.requires_calibration: + raise ValueError( + f"The quantization method {quantization_config.quant_method} does require the model to be pre-quantized." + f" You explicitly passed `pre_quantized=False` meaning your model weights are not quantized. Make sure to " + f"pass `pre_quantized=True` while knowing what you are doing." + ) + + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + """ + Some quantization methods require to explicitly set the dtype of the model to a target dtype. You need to + override this method in case you want to make sure that behavior is preserved + + Args: + torch_dtype (`torch.dtype`): + The input dtype that is passed in `from_pretrained` + """ + return torch_dtype + + def update_device_map(self, device_map: dict[str, Any] | None) -> dict[str, Any] | None: + """ + Override this method if you want to pass a override the existing device map with a new one. E.g. for + bitsandbytes, since `accelerate` is a hard requirement, if no device_map is passed, the device_map is set to + `"auto"`` + + Args: + device_map (`dict | str`, *optional*): + The device_map that is passed through the `from_pretrained` method. + """ + return device_map + + def adjust_target_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + """ + Override this method if you want to adjust the `target_dtype` variable used in `from_pretrained` to compute the + device_map in case the device_map is a `str`. E.g. for bitsandbytes we force-set `target_dtype` to `torch.int8` + and for 4-bit we pass a custom enum `accelerate.CustomDtype.int4`. + + Args: + torch_dtype (`torch.dtype`, *optional*): + The torch_dtype that is used to compute the device_map. + """ + return torch_dtype + + def update_missing_keys(self, model, missing_keys: list[str], prefix: str) -> list[str]: + """ + Override this method if you want to adjust the `missing_keys`. + + Args: + missing_keys (`list[str]`, *optional*): + The list of missing keys in the checkpoint compared to the state dict of the model + """ + return missing_keys + + def get_special_dtypes_update(self, model, torch_dtype: "torch.dtype") -> dict[str, "torch.dtype"]: + """ + returns dtypes for modules that are not quantized - used for the computation of the device_map in case one + passes a str as a device_map. The method will use the `modules_to_not_convert` that is modified in + `_process_model_before_weight_loading`. `diffusers` models don't have any `modules_to_not_convert` attributes + yet but this can change soon in the future. + + Args: + model (`~diffusers.models.modeling_utils.ModelMixin`): + The model to quantize + torch_dtype (`torch.dtype`): + The dtype passed in `from_pretrained` method. + """ + + return { + name: torch_dtype + for name, _ in model.named_parameters() + if any(m in name for m in self.modules_to_not_convert) + } + + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + """adjust max_memory argument for infer_auto_device_map() if extra memory is needed for quantization""" + return max_memory + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ) -> bool: + """ + checks if a loaded state_dict component is part of quantized param + some validation; only defined for + quantization methods that require to create a new parameters for quantization. + """ + return False + + def create_quantized_param(self, *args, **kwargs) -> "torch.nn.Parameter": + """ + takes needed components from state_dict and creates quantized param. + """ + return + + def check_quantized_param_shape(self, *args, **kwargs): + """ + checks if the quantized param has expected shape. + """ + return True + + def validate_environment(self, *args, **kwargs): + """ + This method is used to potentially check for potential conflicts with arguments that are passed in + `from_pretrained`. You need to define it for all future quantizers that are integrated with diffusers. If no + explicit check are needed, simply return nothing. + """ + return + + def maybe_update_loaded_keys(self, loaded_keys: list[str], checkpoint_files: list[str]) -> list[str]: + return loaded_keys + + def maybe_update_state_dict(self, state_dict: dict[str, Any]) -> dict[str, Any]: + return state_dict + + @property + def supports_parallel_loading(self) -> bool: + return True + + def get_state_dict_and_metadata( + self, state_dict: dict[str, Any], safe_serialization: bool = False + ) -> tuple[dict[str, Any], dict[str, Any]]: + return state_dict, {} + + @property + def supports_safetensors_serialization(self) -> bool: + return True + + def preprocess_model(self, model: "ModelMixin", **kwargs): + """ + Setting model attributes and/or converting model before weights loading. At this point the model should be + initialized on the meta device so you can freely manipulate the skeleton of the model in order to replace + modules in-place. Make sure to override the abstract method `_process_model_before_weight_loading`. + + Args: + model (`~diffusers.models.modeling_utils.ModelMixin`): + The model to quantize + kwargs (`dict`, *optional*): + The keyword arguments that are passed along `_process_model_before_weight_loading`. + """ + model.is_quantized = True + model.quantization_method = self.quantization_config.quant_method + return self._process_model_before_weight_loading(model, **kwargs) + + def postprocess_model(self, model: "ModelMixin", **kwargs): + """ + Post-process the model post weights loading. Make sure to override the abstract method + `_process_model_after_weight_loading`. + + Args: + model (`~diffusers.models.modeling_utils.ModelMixin`): + The model to quantize + kwargs (`dict`, *optional*): + The keyword arguments that are passed along `_process_model_after_weight_loading`. + """ + return self._process_model_after_weight_loading(model, **kwargs) + + def dequantize(self, model): + """ + Potentially dequantize the model to retrieve the original model, with some loss in accuracy / performance. Note + not all quantization schemes support this. + """ + model = self._dequantize(model) + + # Delete quantizer and quantization config + del model.hf_quantizer + model.is_quantized = False + + return model + + def get_cuda_warm_up_factor(self): + """ + The factor to be used in `caching_allocator_warmup` to get the number of bytes to pre-allocate to warm up cuda. + A factor of 2 means we allocate all bytes in the empty model (since we allocate in fp16), a factor of 4 means + we allocate half the memory of the weights residing in the empty model, etc... + """ + # By default we return 4, i.e. half the model size (this corresponds to the case where the model is not + # really pre-processed, i.e. we do not have the info that weights are going to be 8 bits before actual + # weight loading) + return 4 + + def _dequantize(self, model): + raise NotImplementedError( + f"{self.quantization_config.quant_method} has no implementation of `dequantize`, please raise an issue on GitHub." + ) + + @abstractmethod + def _process_model_before_weight_loading(self, model, **kwargs): ... + + @abstractmethod + def _process_model_after_weight_loading(self, model, **kwargs): ... + + @property + @abstractmethod + def is_serializable(self): ... + + @property + @abstractmethod + def is_trainable(self): ... + + @property + def is_compileable(self) -> bool: + """Flag indicating whether the quantized model can be compiled""" + return False diff --git a/diffusers/quantizers/bitsandbytes/__init__.py b/diffusers/quantizers/bitsandbytes/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..9e745bc810faad76608b00ab9c493369466d0970 --- /dev/null +++ b/diffusers/quantizers/bitsandbytes/__init__.py @@ -0,0 +1,2 @@ +from .bnb_quantizer import BnB4BitDiffusersQuantizer, BnB8BitDiffusersQuantizer +from .utils import dequantize_and_replace, dequantize_bnb_weight, replace_with_bnb_linear diff --git a/diffusers/quantizers/bitsandbytes/bnb_quantizer.py b/diffusers/quantizers/bitsandbytes/bnb_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..7c5dcbc73d5f4c4449427f9a528b5f574a36e4b1 --- /dev/null +++ b/diffusers/quantizers/bitsandbytes/bnb_quantizer.py @@ -0,0 +1,569 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Adapted from +https://github.com/huggingface/transformers/blob/c409cd81777fb27aadc043ed3d8339dbc020fb3b/src/transformers/quantizers/quantizer_bnb_4bit.py +""" + +from typing import TYPE_CHECKING, Any + +from ...utils import get_module_from_name +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + +from ...utils import ( + is_accelerate_available, + is_accelerate_version, + is_bitsandbytes_available, + is_bitsandbytes_version, + is_torch_available, + logging, +) + + +if is_torch_available(): + import torch + +logger = logging.get_logger(__name__) + + +class BnB4BitDiffusersQuantizer(DiffusersQuantizer): + """ + 4-bit quantization from bitsandbytes.py quantization method: + before loading: converts transformer layers into Linear4bit during loading: load 16bit weight and pass to the + layer object after: quantizes individual weights in Linear4bit into 4bit at the first .cuda() call saving: + from state dict, as usual; saves weights and `quant_state` components + loading: + need to locate `quant_state` components and pass to Param4bit constructor + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + if self.quantization_config.llm_int8_skip_modules is not None: + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + def validate_environment(self, *args, **kwargs): + if not (torch.cuda.is_available() or torch.xpu.is_available() or torch.mps.is_available()): + raise RuntimeError("No GPU found. A GPU is needed for quantization.") + if not is_accelerate_available() or is_accelerate_version("<", "0.26.0"): + raise ImportError( + "Using `bitsandbytes` 4-bit quantization requires Accelerate: `pip install 'accelerate>=0.26.0'`" + ) + if not is_bitsandbytes_available() or is_bitsandbytes_version("<", "0.43.3"): + raise ImportError( + "Using `bitsandbytes` 4-bit quantization requires the latest version of bitsandbytes: `pip install -U bitsandbytes`" + ) + + device_map = kwargs.get("device_map", None) + if ( + device_map is not None + and isinstance(device_map, dict) + and not self.quantization_config.llm_int8_enable_fp32_cpu_offload + ): + device_map_without_no_convert = { + key: device_map[key] for key in device_map.keys() if key not in self.modules_to_not_convert + } + if "cpu" in device_map_without_no_convert.values() or "disk" in device_map_without_no_convert.values(): + raise ValueError( + "Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the " + "quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules " + "in 32-bit, you need to set `load_in_8bit_fp32_cpu_offload=True` and pass a custom `device_map` to " + "`from_pretrained`. Check " + "https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu " + "for more details. " + ) + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if target_dtype != torch.int8: + from accelerate.utils import CustomDtype + + logger.info("target_dtype {target_dtype} is replaced by `CustomDtype.INT4` for 4-bit BnB quantization") + return CustomDtype.INT4 + else: + raise ValueError(f"Wrong `target_dtype` ({target_dtype}) provided.") + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ) -> bool: + import bitsandbytes as bnb + + module, tensor_name = get_module_from_name(model, param_name) + if isinstance(module._parameters.get(tensor_name, None), bnb.nn.Params4bit): + # Add here check for loaded components' dtypes once serialization is implemented + return True + elif isinstance(module, bnb.nn.Linear4bit) and tensor_name == "bias": + # bias could be loaded by regular set_module_tensor_to_device() from accelerate, + # but it would wrongly use uninitialized weight there. + return True + else: + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: dict[str, Any], + unexpected_keys: list[str] | None = None, + **kwargs, + ): + import bitsandbytes as bnb + + module, tensor_name = get_module_from_name(model, param_name) + + if tensor_name not in module._parameters: + raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.") + + old_value = getattr(module, tensor_name) + + if tensor_name == "bias": + if param_value is None: + new_value = old_value.to(target_device) + else: + new_value = param_value.to(target_device) + + new_value = torch.nn.Parameter(new_value, requires_grad=old_value.requires_grad) + module._parameters[tensor_name] = new_value + return + + if not isinstance(module._parameters[tensor_name], bnb.nn.Params4bit): + raise ValueError("this function only loads `Linear4bit components`") + if ( + old_value.device == torch.device("meta") + and target_device not in ["meta", torch.device("meta")] + and param_value is None + ): + raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {target_device}.") + + # construct `new_value` for the module._parameters[tensor_name]: + if self.pre_quantized: + # 4bit loading. Collecting components for restoring quantized weight + # This can be expanded to make a universal call for any quantized weight loading + + if not self.is_serializable: + raise ValueError( + "Detected int4 weights but the version of bitsandbytes is not compatible with int4 serialization. " + "Make sure to download the latest `bitsandbytes` version. `pip install --upgrade bitsandbytes`." + ) + + if (param_name + ".quant_state.bitsandbytes__fp4" not in state_dict) and ( + param_name + ".quant_state.bitsandbytes__nf4" not in state_dict + ): + raise ValueError( + f"Supplied state dict for {param_name} does not contain `bitsandbytes__*` and possibly other `quantized_stats` components." + ) + + quantized_stats = {} + for k, v in state_dict.items(): + # `startswith` to counter for edge cases where `param_name` + # substring can be present in multiple places in the `state_dict` + if param_name + "." in k and k.startswith(param_name): + quantized_stats[k] = v + if unexpected_keys is not None and k in unexpected_keys: + unexpected_keys.remove(k) + + new_value = bnb.nn.Params4bit.from_prequantized( + data=param_value, + quantized_stats=quantized_stats, + requires_grad=False, + device=target_device, + ) + else: + new_value = param_value.to("cpu") + kwargs = old_value.__dict__ + new_value = bnb.nn.Params4bit(new_value, requires_grad=False, **kwargs).to(target_device) + + module._parameters[tensor_name] = new_value + + def check_quantized_param_shape(self, param_name, current_param, loaded_param): + current_param_shape = current_param.shape + loaded_param_shape = loaded_param.shape + + n = current_param_shape.numel() + inferred_shape = (n,) if "bias" in param_name else ((n + 1) // 2, 1) + if loaded_param_shape != inferred_shape: + raise ValueError( + f"Expected the flattened shape of the current param ({param_name}) to be {loaded_param_shape} but is {inferred_shape}." + ) + else: + return True + + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + # need more space for buffers that are created during quantization + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + if torch_dtype is None: + # We force the `dtype` to be float16, this is a requirement from `bitsandbytes` + logger.info( + "Overriding torch_dtype=%s with `torch_dtype=torch.float16` due to " + "requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. " + "Pass your own torch_dtype to specify the dtype of the remaining non-linear layers or pass" + " torch_dtype=torch.float16 to remove this warning.", + torch_dtype, + ) + torch_dtype = torch.float16 + return torch_dtype + + def update_device_map(self, device_map): + if device_map is None: + if torch.xpu.is_available(): + current_device = f"xpu:{torch.xpu.current_device()}" + elif torch.mps.is_available(): + current_device = "mps" + else: + current_device = f"cuda:{torch.cuda.current_device()}" + device_map = {"": current_device} + logger.info( + "The device_map was not initialized. " + "Setting device_map to {" + ": {current_device}}. " + "If you want to use the model for inference, please set device_map ='auto' " + ) + return device_map + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + from .utils import replace_with_bnb_linear + + load_in_8bit_fp32_cpu_offload = self.quantization_config.llm_int8_enable_fp32_cpu_offload + + # We may keep some modules such as the `proj_out` in their original dtype for numerical stability reasons + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk` + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]] + + if len(keys_on_cpu) > 0 and not load_in_8bit_fp32_cpu_offload: + raise ValueError( + "If you want to offload some keys to `cpu` or `disk`, you need to set " + "`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be " + " converted to 8-bit but kept in 32-bit." + ) + self.modules_to_not_convert.extend(keys_on_cpu) + + # Purge `None`. + # Unlike `transformers`, we don't know if we should always keep certain modules in FP32 + # in case of diffusion transformer models. For language models and others alike, `lm_head` + # and tied modules are usually kept in FP32. + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + model = replace_with_bnb_linear( + model, modules_to_not_convert=self.modules_to_not_convert, quantization_config=self.quantization_config + ) + model.config.quantization_config = self.quantization_config + model.is_loaded_in_4bit = True + + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + model.is_4bit_serializable = self.is_serializable + return model + + @property + def is_serializable(self): + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + @property + def is_trainable(self) -> bool: + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + def _dequantize(self, model): + from .utils import dequantize_and_replace + + is_model_on_cpu = model.device.type == "cpu" + if is_model_on_cpu: + logger.info( + "Model was found to be on CPU (could happen as a result of `enable_model_cpu_offload()`). So, moving it to GPU. After dequantization, will move the model back to CPU again to preserve the previous device." + ) + if torch.xpu.is_available(): + model.to(torch.xpu.current_device()) + else: + model.to(torch.cuda.current_device()) + + model = dequantize_and_replace( + model, self.modules_to_not_convert, quantization_config=self.quantization_config + ) + if is_model_on_cpu: + model.to("cpu") + return model + + +class BnB8BitDiffusersQuantizer(DiffusersQuantizer): + """ + 8-bit quantization from bitsandbytes quantization method: + before loading: converts transformer layers into Linear8bitLt during loading: load 16bit weight and pass to the + layer object after: quantizes individual weights in Linear8bitLt into 8bit at fitst .cuda() call + saving: + from state dict, as usual; saves weights and 'SCB' component + loading: + need to locate SCB component and pass to the Linear8bitLt object + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + if self.quantization_config.llm_int8_skip_modules is not None: + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + def validate_environment(self, *args, **kwargs): + if not (torch.cuda.is_available() or torch.xpu.is_available()): + raise RuntimeError("No GPU found. A GPU is needed for quantization.") + if not is_accelerate_available() or is_accelerate_version("<", "0.26.0"): + raise ImportError( + "Using `bitsandbytes` 8-bit quantization requires Accelerate: `pip install 'accelerate>=0.26.0'`" + ) + if not is_bitsandbytes_available() or is_bitsandbytes_version("<", "0.43.3"): + raise ImportError( + "Using `bitsandbytes` 8-bit quantization requires the latest version of bitsandbytes: `pip install -U bitsandbytes`" + ) + + device_map = kwargs.get("device_map", None) + if ( + device_map is not None + and isinstance(device_map, dict) + and not self.quantization_config.llm_int8_enable_fp32_cpu_offload + ): + device_map_without_no_convert = { + key: device_map[key] for key in device_map.keys() if key not in self.modules_to_not_convert + } + if "cpu" in device_map_without_no_convert.values() or "disk" in device_map_without_no_convert.values(): + raise ValueError( + "Some modules are dispatched on the CPU or the disk. Make sure you have enough GPU RAM to fit the " + "quantized model. If you want to dispatch the model on the CPU or the disk while keeping these modules " + "in 32-bit, you need to set `load_in_8bit_fp32_cpu_offload=True` and pass a custom `device_map` to " + "`from_pretrained`. Check " + "https://huggingface.co/docs/transformers/main/en/main_classes/quantization#offload-between-cpu-and-gpu " + "for more details. " + ) + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.adjust_max_memory + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + # need more space for buffers that are created during quantization + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.update_torch_dtype + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + if torch_dtype is None: + # We force the `dtype` to be float16, this is a requirement from `bitsandbytes` + logger.info( + "Overriding torch_dtype=%s with `torch_dtype=torch.float16` due to " + "requirements of `bitsandbytes` to enable model loading in 8-bit or 4-bit. " + "Pass your own torch_dtype to specify the dtype of the remaining non-linear layers or pass" + " torch_dtype=torch.float16 to remove this warning.", + torch_dtype, + ) + torch_dtype = torch.float16 + return torch_dtype + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.update_device_map + def update_device_map(self, device_map): + if device_map is None: + if torch.xpu.is_available(): + current_device = f"xpu:{torch.xpu.current_device()}" + elif torch.mps.is_available(): + current_device = "mps" + else: + current_device = f"cuda:{torch.cuda.current_device()}" + device_map = {"": current_device} + logger.info( + "The device_map was not initialized. " + "Setting device_map to {" + ": {current_device}}. " + "If you want to use the model for inference, please set device_map ='auto' " + ) + return device_map + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if target_dtype != torch.int8: + logger.info("target_dtype {target_dtype} is replaced by `torch.int8` for 8-bit BnB quantization") + return torch.int8 + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ): + import bitsandbytes as bnb + + module, tensor_name = get_module_from_name(model, param_name) + if isinstance(module._parameters.get(tensor_name, None), bnb.nn.Int8Params): + if self.pre_quantized: + if param_name.replace("weight", "SCB") not in state_dict.keys(): + raise ValueError("Missing quantization component `SCB`") + if param_value.dtype != torch.int8: + raise ValueError( + f"Incompatible dtype `{param_value.dtype}` when loading 8-bit prequantized weight. Expected `torch.int8`." + ) + return True + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: dict[str, Any], + unexpected_keys: list[str] | None = None, + **kwargs, + ): + import bitsandbytes as bnb + + fp16_statistics_key = param_name.replace("weight", "SCB") + fp16_weights_format_key = param_name.replace("weight", "weight_format") + + fp16_statistics = state_dict.get(fp16_statistics_key, None) + fp16_weights_format = state_dict.get(fp16_weights_format_key, None) + + module, tensor_name = get_module_from_name(model, param_name) + if tensor_name not in module._parameters: + raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.") + + old_value = getattr(module, tensor_name) + + if not isinstance(module._parameters[tensor_name], bnb.nn.Int8Params): + raise ValueError(f"Parameter `{tensor_name}` should only be a `bnb.nn.Int8Params` instance.") + if ( + old_value.device == torch.device("meta") + and target_device not in ["meta", torch.device("meta")] + and param_value is None + ): + raise ValueError(f"{tensor_name} is on the meta device, we need a `value` to put in on {target_device}.") + + new_value = param_value.to("cpu") + if self.pre_quantized and not self.is_serializable: + raise ValueError( + "Detected int8 weights but the version of bitsandbytes is not compatible with int8 serialization. " + "Make sure to download the latest `bitsandbytes` version. `pip install --upgrade bitsandbytes`." + ) + + kwargs = old_value.__dict__ + new_value = bnb.nn.Int8Params(new_value, requires_grad=False, **kwargs).to(target_device) + + module._parameters[tensor_name] = new_value + if fp16_statistics is not None: + setattr(module.weight, "SCB", fp16_statistics.to(target_device)) + if unexpected_keys is not None: + unexpected_keys.remove(fp16_statistics_key) + + # We just need to pop the `weight_format` keys from the state dict to remove unneeded + # messages. The correct format is correctly retrieved during the first forward pass. + if fp16_weights_format is not None and unexpected_keys is not None: + unexpected_keys.remove(fp16_weights_format_key) + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer._process_model_after_weight_loading with 4bit->8bit + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + model.is_8bit_serializable = self.is_serializable + return model + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer._process_model_before_weight_loading with 4bit->8bit + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + from .utils import replace_with_bnb_linear + + load_in_8bit_fp32_cpu_offload = self.quantization_config.llm_int8_enable_fp32_cpu_offload + + # We may keep some modules such as the `proj_out` in their original dtype for numerical stability reasons + self.modules_to_not_convert = self.quantization_config.llm_int8_skip_modules + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk` + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]] + + if len(keys_on_cpu) > 0 and not load_in_8bit_fp32_cpu_offload: + raise ValueError( + "If you want to offload some keys to `cpu` or `disk`, you need to set " + "`llm_int8_enable_fp32_cpu_offload=True`. Note that these modules will not be " + " converted to 8-bit but kept in 32-bit." + ) + self.modules_to_not_convert.extend(keys_on_cpu) + + # Purge `None`. + # Unlike `transformers`, we don't know if we should always keep certain modules in FP32 + # in case of diffusion transformer models. For language models and others alike, `lm_head` + # and tied modules are usually kept in FP32. + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + model = replace_with_bnb_linear( + model, modules_to_not_convert=self.modules_to_not_convert, quantization_config=self.quantization_config + ) + model.config.quantization_config = self.quantization_config + model.is_loaded_in_8bit = True + + @property + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.is_serializable + def is_serializable(self): + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + @property + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.is_serializable + def is_trainable(self) -> bool: + # Because we're mandating `bitsandbytes` 0.43.3. + return True + + @property + def is_compileable(self) -> bool: + return True + + def _dequantize(self, model): + from .utils import dequantize_and_replace + + model = dequantize_and_replace( + model, self.modules_to_not_convert, quantization_config=self.quantization_config + ) + return model diff --git a/diffusers/quantizers/bitsandbytes/utils.py b/diffusers/quantizers/bitsandbytes/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..aea4e7dda57f297c2cbf615b92d444ee4b7b1e92 --- /dev/null +++ b/diffusers/quantizers/bitsandbytes/utils.py @@ -0,0 +1,318 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Adapted from +https://github.com/huggingface/transformers/blob/c409cd81777fb27aadc043ed3d8339dbc020fb3b/src/transformers/integrations/bitsandbytes.py +""" + +import inspect +from inspect import signature + +from ...utils import is_accelerate_available, is_bitsandbytes_available, is_torch_available, logging +from ..quantization_config import QuantizationMethod + + +if is_torch_available(): + import torch + import torch.nn as nn + +if is_bitsandbytes_available(): + import bitsandbytes as bnb + +if is_accelerate_available(): + import accelerate + from accelerate import init_empty_weights + from accelerate.hooks import add_hook_to_module, remove_hook_from_module + +logger = logging.get_logger(__name__) + + +def _replace_with_bnb_linear( + model, + modules_to_not_convert=None, + current_key_name=None, + quantization_config=None, + has_been_replaced=False, +): + """ + Private method that wraps the recursion for module replacement. + + Returns the converted model and a boolean that indicates if the conversion has been successful or not. + """ + for name, module in model.named_children(): + if current_key_name is None: + current_key_name = [] + current_key_name.append(name) + + if isinstance(module, nn.Linear) and name not in modules_to_not_convert: + # Check if the current key is not in the `modules_to_not_convert` + current_key_name_str = ".".join(current_key_name) + if not any( + (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert + ): + with init_empty_weights(): + in_features = module.in_features + out_features = module.out_features + + if quantization_config.quantization_method() == "llm_int8": + model._modules[name] = bnb.nn.Linear8bitLt( + in_features, + out_features, + module.bias is not None, + has_fp16_weights=quantization_config.llm_int8_has_fp16_weight, + threshold=quantization_config.llm_int8_threshold, + ) + has_been_replaced = True + else: + if ( + quantization_config.llm_int8_skip_modules is not None + and name in quantization_config.llm_int8_skip_modules + ): + pass + else: + extra_kwargs = ( + {"quant_storage": quantization_config.bnb_4bit_quant_storage} + if "quant_storage" in list(signature(bnb.nn.Linear4bit).parameters) + else {} + ) + model._modules[name] = bnb.nn.Linear4bit( + in_features, + out_features, + module.bias is not None, + quantization_config.bnb_4bit_compute_dtype, + compress_statistics=quantization_config.bnb_4bit_use_double_quant, + quant_type=quantization_config.bnb_4bit_quant_type, + **extra_kwargs, + ) + has_been_replaced = True + # Store the module class in case we need to transpose the weight later + model._modules[name].source_cls = type(module) + # Force requires grad to False to avoid unexpected errors + model._modules[name].requires_grad_(False) + if len(list(module.children())) > 0: + _, has_been_replaced = _replace_with_bnb_linear( + module, + modules_to_not_convert, + current_key_name, + quantization_config, + has_been_replaced=has_been_replaced, + ) + # Remove the last key for recursion + current_key_name.pop(-1) + return model, has_been_replaced + + +def replace_with_bnb_linear(model, modules_to_not_convert=None, current_key_name=None, quantization_config=None): + """ + Helper function to replace the `nn.Linear` layers within `model` with either `bnb.nn.Linear8bit` or + `bnb.nn.Linear4bit` using the `bitsandbytes` library. + + References: + * `bnb.nn.Linear8bit`: [LLM.int8(): 8-bit Matrix Multiplication for Transformers at + Scale](https://huggingface.co/papers/2208.07339) + * `bnb.nn.Linear4bit`: [QLoRA: Efficient Finetuning of Quantized + LLMs](https://huggingface.co/papers/2305.14314) + + Parameters: + model (`torch.nn.Module`): + Input model or `torch.nn.Module` as the function is run recursively. + modules_to_not_convert (`list[`str`]`, *optional*, defaults to `[]`): + Names of the modules to not convert in `Linear8bitLt`. In practice we keep the `modules_to_not_convert` in + full precision for numerical stability reasons. + current_key_name (`list[`str`]`, *optional*): + An array to track the current key of the recursion. This is used to check whether the current key (part of + it) is not in the list of modules to not convert (for instances modules that are offloaded to `cpu` or + `disk`). + quantization_config ('transformers.utils.quantization_config.BitsAndBytesConfig'): + To configure and manage settings related to quantization, a technique used to compress neural network + models by reducing the precision of the weights and activations, thus making models more efficient in terms + of both storage and computation. + """ + model, _ = _replace_with_bnb_linear(model, modules_to_not_convert, current_key_name, quantization_config) + + has_been_replaced = any( + isinstance(replaced_module, (bnb.nn.Linear4bit, bnb.nn.Linear8bitLt)) + for _, replaced_module in model.named_modules() + ) + if not has_been_replaced: + logger.warning( + "You are loading your model in 8bit or 4bit but no linear modules were found in your model." + " Please double check your model architecture, or submit an issue on github if you think this is" + " a bug." + ) + + return model + + +# Adapted from PEFT: https://github.com/huggingface/peft/blob/6d458b300fc2ed82e19f796b53af4c97d03ea604/src/peft/utils/integrations.py#L81 +def dequantize_bnb_weight(weight: "torch.nn.Parameter", state=None, dtype: "torch.dtype" = None): + """ + Helper function to dequantize 4bit or 8bit bnb weights. + + If the weight is not a bnb quantized weight, it will be returned as is. + """ + if not isinstance(weight, torch.nn.Parameter): + raise TypeError(f"Input weight should be of type nn.Parameter, got {type(weight)} instead") + + cls_name = weight.__class__.__name__ + if cls_name not in ("Params4bit", "Int8Params"): + return weight + + if cls_name == "Params4bit": + output_tensor = bnb.functional.dequantize_4bit(weight.data, weight.quant_state) + msg = f"The model is going to be dequantized in {output_tensor.dtype} - if you want to upcast it to another dtype, make sure to pass the desired dtype when quantizing the model through `bnb_4bit_quant_type` argument of `BitsAndBytesConfig`" + if dtype: + msg = f"The model is going to be first dequantized in {output_tensor.dtype} and type-casted to {dtype}" + output_tensor = output_tensor.to(dtype) + logger.warning_once(msg) + return output_tensor + + if state.SCB is None: + state.SCB = weight.SCB + + if hasattr(bnb.functional, "int8_vectorwise_dequant"): + # Use bitsandbytes API if available (requires v0.45.0+) + dequantized = bnb.functional.int8_vectorwise_dequant(weight.data, state.SCB) + else: + # Multiply by (scale/127) to dequantize. + dequantized = weight.data * state.SCB.view(-1, 1) * 7.874015718698502e-3 + + if dtype: + dequantized = dequantized.to(dtype) + return dequantized + + +def _create_accelerate_new_hook(old_hook): + r""" + Creates a new hook based on the old hook. Use it only if you know what you are doing ! This method is a copy of: + https://github.com/huggingface/peft/blob/748f7968f3a31ec06a1c2b0328993319ad9a150a/src/peft/utils/other.py#L245 with + some changes + """ + old_hook_cls = getattr(accelerate.hooks, old_hook.__class__.__name__) + old_hook_attr = old_hook.__dict__ + filtered_old_hook_attr = {} + old_hook_init_signature = inspect.signature(old_hook_cls.__init__) + for k in old_hook_attr.keys(): + if k in old_hook_init_signature.parameters: + filtered_old_hook_attr[k] = old_hook_attr[k] + new_hook = old_hook_cls(**filtered_old_hook_attr) + return new_hook + + +def _dequantize_and_replace( + model, + dtype, + modules_to_not_convert=None, + current_key_name=None, + quantization_config=None, + has_been_replaced=False, +): + """ + Converts a quantized model into its dequantized original version. The newly converted model will have some + performance drop compared to the original model before quantization - use it only for specific usecases such as + QLoRA adapters merging. + + Returns the converted model and a boolean that indicates if the conversion has been successful or not. + """ + quant_method = quantization_config.quantization_method() + + target_cls = bnb.nn.Linear8bitLt if quant_method == "llm_int8" else bnb.nn.Linear4bit + + for name, module in model.named_children(): + if current_key_name is None: + current_key_name = [] + current_key_name.append(name) + + if isinstance(module, target_cls) and name not in modules_to_not_convert: + # Check if the current key is not in the `modules_to_not_convert` + current_key_name_str = ".".join(current_key_name) + + if not any( + (key + "." in current_key_name_str) or (key == current_key_name_str) for key in modules_to_not_convert + ): + bias = getattr(module, "bias", None) + + device = module.weight.device + with init_empty_weights(): + new_module = torch.nn.Linear(module.in_features, module.out_features, bias=bias is not None) + + if quant_method == "llm_int8": + state = module.state + else: + state = None + + new_module.weight = torch.nn.Parameter(dequantize_bnb_weight(module.weight, state, dtype)) + + if bias is not None: + new_module.bias = bias + + # Create a new hook and attach it in case we use accelerate + if hasattr(module, "_hf_hook"): + old_hook = module._hf_hook + new_hook = _create_accelerate_new_hook(old_hook) + + remove_hook_from_module(module) + add_hook_to_module(new_module, new_hook) + + new_module.to(device) + model._modules[name] = new_module + has_been_replaced = True + if len(list(module.children())) > 0: + _, has_been_replaced = _dequantize_and_replace( + module, + dtype=dtype, + modules_to_not_convert=modules_to_not_convert, + current_key_name=current_key_name, + quantization_config=quantization_config, + has_been_replaced=has_been_replaced, + ) + # Remove the last key for recursion + current_key_name.pop(-1) + return model, has_been_replaced + + +def dequantize_and_replace( + model, + modules_to_not_convert=None, + quantization_config=None, +): + model, _ = _dequantize_and_replace( + model, + dtype=model.dtype, + modules_to_not_convert=modules_to_not_convert, + quantization_config=quantization_config, + ) + has_been_replaced = any( + isinstance(replaced_module, torch.nn.Linear) for _, replaced_module in model.named_modules() + ) + if not has_been_replaced: + logger.warning( + "Some linear modules were not dequantized. This could lead to unexpected behaviour. Please check your model." + ) + + return model + + +def _check_bnb_status(module) -> bool | bool: + is_loaded_in_4bit_bnb = ( + hasattr(module, "is_loaded_in_4bit") + and module.is_loaded_in_4bit + and getattr(module, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES + ) + is_loaded_in_8bit_bnb = ( + hasattr(module, "is_loaded_in_8bit") + and module.is_loaded_in_8bit + and getattr(module, "quantization_method", None) == QuantizationMethod.BITS_AND_BYTES + ) + return is_loaded_in_4bit_bnb or is_loaded_in_8bit_bnb, is_loaded_in_4bit_bnb, is_loaded_in_8bit_bnb diff --git a/diffusers/quantizers/gguf/__init__.py b/diffusers/quantizers/gguf/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..b3d9082ac803c8da97ba7bb37021445307b96f7c --- /dev/null +++ b/diffusers/quantizers/gguf/__init__.py @@ -0,0 +1 @@ +from .gguf_quantizer import GGUFQuantizer diff --git a/diffusers/quantizers/gguf/gguf_quantizer.py b/diffusers/quantizers/gguf/gguf_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..42ea3982c91239294d31c6431262d1efe718d90a --- /dev/null +++ b/diffusers/quantizers/gguf/gguf_quantizer.py @@ -0,0 +1,190 @@ +from __future__ import annotations + +from typing import TYPE_CHECKING, Any + +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +from ...utils import ( + get_module_from_name, + is_accelerate_available, + is_accelerate_version, + is_gguf_available, + is_gguf_version, + is_torch_available, + logging, +) + + +if is_torch_available() and is_gguf_available(): + import torch + + from .utils import ( + GGML_QUANT_SIZES, + GGUFParameter, + _dequantize_gguf_and_restore_linear, + _quant_shape_from_byte_shape, + _replace_with_gguf_linear, + dequantize_gguf_tensor, + ) + + +logger = logging.get_logger(__name__) + + +class GGUFQuantizer(DiffusersQuantizer): + use_keep_in_fp32_modules = True + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + self.compute_dtype = quantization_config.compute_dtype + self.pre_quantized = quantization_config.pre_quantized + self.modules_to_not_convert = quantization_config.modules_to_not_convert or [] + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + def validate_environment(self, *args, **kwargs): + if not is_accelerate_available() or is_accelerate_version("<", "0.26.0"): + raise ImportError( + "Loading GGUF Parameters requires `accelerate` installed in your environment: `pip install 'accelerate>=0.26.0'`" + ) + if not is_gguf_available() or is_gguf_version("<", "0.10.0"): + raise ImportError( + "To load GGUF format files you must have `gguf` installed in your environment: `pip install gguf>=0.10.0`" + ) + + # Copied from diffusers.quantizers.bitsandbytes.bnb_quantizer.BnB4BitDiffusersQuantizer.adjust_max_memory + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + # need more space for buffers that are created during quantization + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if target_dtype != torch.uint8: + logger.info(f"target_dtype {target_dtype} is replaced by `torch.uint8` for GGUF quantization") + return torch.uint8 + + def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype": + if torch_dtype is None: + torch_dtype = self.compute_dtype + return torch_dtype + + def check_quantized_param_shape(self, param_name, current_param, loaded_param): + loaded_param_shape = loaded_param.shape + current_param_shape = current_param.shape + quant_type = loaded_param.quant_type + + block_size, type_size = GGML_QUANT_SIZES[quant_type] + + inferred_shape = _quant_shape_from_byte_shape(loaded_param_shape, type_size, block_size) + if inferred_shape != current_param_shape: + raise ValueError( + f"{param_name} has an expected shape of: {current_param_shape}, but the loaded GGUF weight decodes " + f"to shape: {inferred_shape}" + ) + + return True + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "GGUFParameter" | "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ) -> bool: + if isinstance(param_value, GGUFParameter): + return True + + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "GGUFParameter" | "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: dict[str, Any] | None = None, + unexpected_keys: list[str] | None = None, + **kwargs, + ): + module, tensor_name = get_module_from_name(model, param_name) + if tensor_name not in module._parameters and tensor_name not in module._buffers: + raise ValueError(f"{module} does not have a parameter or a buffer named {tensor_name}.") + + # If the GGUFParameter should not be quantized (for example, it is a submodule of any excluded module), + # dequantize it and set the (dequantized) parameter to the proper dtype. + if isinstance(param_value, GGUFParameter) and any( + m in param_name.split(".") for m in self.modules_to_not_convert + ): + keep_in_fp32 = getattr(self, "keep_in_fp32_modules", []) + param_should_be_fp32 = any(m in param_name.split(".") for m in keep_in_fp32) + target_dtype = torch.float32 if param_should_be_fp32 else self.compute_dtype + if param_should_be_fp32: + logger.warning(f"Quantized parameter {param_name} is required to remain in FP32, dequantizing now.") + else: + logger.warning( + f"Quantized parameter {param_name} is excluded by `modules_to_not_convert`, dequantizing now." + ) + param_value = dequantize_gguf_tensor(param_value).to(target_dtype) + + if tensor_name in module._parameters: + module._parameters[tensor_name] = param_value.to(target_device) + if tensor_name in module._buffers: + module._buffers[tensor_name] = param_value.to(target_device) + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + state_dict = kwargs.get("state_dict", None) + + self.keep_in_fp32_modules = [module for module in keep_in_fp32_modules if module is not None] + self.modules_to_not_convert.extend(self.keep_in_fp32_modules) + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + _replace_with_gguf_linear( + model, self.compute_dtype, state_dict, modules_to_not_convert=self.modules_to_not_convert + ) + + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + return model + + @property + def is_serializable(self): + return False + + @property + def is_trainable(self) -> bool: + return False + + @property + def is_compileable(self) -> bool: + return True + + def _dequantize(self, model): + is_model_on_cpu = model.device.type == "cpu" + if is_model_on_cpu: + logger.info( + "Model was found to be on CPU (could happen as a result of `enable_model_cpu_offload()`). So, moving it to accelerator. After dequantization, will move the model back to CPU again to preserve the previous device." + ) + device = ( + torch.accelerator.current_accelerator() + if hasattr(torch, "accelerator") + else torch.cuda.current_device() + ) + model.to(device) + + model = _dequantize_gguf_and_restore_linear(model, self.modules_to_not_convert) + if is_model_on_cpu: + model.to("cpu") + return model diff --git a/diffusers/quantizers/gguf/utils.py b/diffusers/quantizers/gguf/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..c7d9ec89bee65270c8b7e7411ef7554fb3e076aa --- /dev/null +++ b/diffusers/quantizers/gguf/utils.py @@ -0,0 +1,612 @@ +# Copyright 2025 The HuggingFace Team and City96. All rights reserved. +# # +# # Licensed under the Apache License, Version 2.0 (the "License"); +# # you may not use this file except in compliance with the License. +# # You may obtain a copy of the License at +# # +# # http://www.apache.org/licenses/LICENSE-2.0 +# # +# # Unless required by applicable law or agreed to in writing, software +# # distributed under the License is distributed on an "AS IS" BASIS, +# # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# # See the License for the specific language governing permissions and +# # limitations under the License. + +import inspect +import os +from contextlib import nullcontext + +import gguf +import torch +import torch.nn as nn + +from ...utils import is_accelerate_available, is_kernels_available + + +if is_accelerate_available(): + import accelerate + from accelerate import init_empty_weights + from accelerate.hooks import add_hook_to_module, remove_hook_from_module + + +can_use_cuda_kernels = ( + os.getenv("DIFFUSERS_GGUF_CUDA_KERNELS", "false").lower() in ["1", "true", "yes"] + and torch.cuda.is_available() + and torch.cuda.get_device_capability()[0] >= 7 +) +if can_use_cuda_kernels and is_kernels_available(): + from kernels import get_kernel + + ops = get_kernel("Isotr0py/ggml") +else: + ops = None + +UNQUANTIZED_TYPES = {gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16, gguf.GGMLQuantizationType.BF16} +STANDARD_QUANT_TYPES = { + gguf.GGMLQuantizationType.Q4_0, + gguf.GGMLQuantizationType.Q4_1, + gguf.GGMLQuantizationType.Q5_0, + gguf.GGMLQuantizationType.Q5_1, + gguf.GGMLQuantizationType.Q8_0, + gguf.GGMLQuantizationType.Q8_1, +} +KQUANT_TYPES = { + gguf.GGMLQuantizationType.Q2_K, + gguf.GGMLQuantizationType.Q3_K, + gguf.GGMLQuantizationType.Q4_K, + gguf.GGMLQuantizationType.Q5_K, + gguf.GGMLQuantizationType.Q6_K, +} +IMATRIX_QUANT_TYPES = { + gguf.GGMLQuantizationType.IQ1_M, + gguf.GGMLQuantizationType.IQ1_S, + gguf.GGMLQuantizationType.IQ2_XXS, + gguf.GGMLQuantizationType.IQ2_XS, + gguf.GGMLQuantizationType.IQ2_S, + gguf.GGMLQuantizationType.IQ3_XXS, + gguf.GGMLQuantizationType.IQ3_S, + gguf.GGMLQuantizationType.IQ4_XS, + gguf.GGMLQuantizationType.IQ4_NL, +} +# TODO(Isotr0py): Currently, we don't have MMQ kernel for I-Matrix quantization. +# Consolidate DEQUANT_TYPES, MMVQ_QUANT_TYPES and MMQ_QUANT_TYPES after we add +# MMQ kernel for I-Matrix quantization. +DEQUANT_TYPES = STANDARD_QUANT_TYPES | KQUANT_TYPES | IMATRIX_QUANT_TYPES +MMVQ_QUANT_TYPES = STANDARD_QUANT_TYPES | KQUANT_TYPES | IMATRIX_QUANT_TYPES +MMQ_QUANT_TYPES = STANDARD_QUANT_TYPES | KQUANT_TYPES + + +def _fused_mul_mat_gguf(x: torch.Tensor, qweight: torch.Tensor, qweight_type: int) -> torch.Tensor: + # there is no need to call any kernel for fp16/bf16 + if qweight_type in UNQUANTIZED_TYPES: + weight = dequantize_gguf_tensor(qweight) + return x @ weight.to(x.dtype).T + + # TODO(Isotr0py): GGUF's MMQ and MMVQ implementation are designed for + # contiguous batching and inefficient with diffusers' batching, + # so we disabled it now. + + # elif qweight_type in MMVQ_QUANT_TYPES: + # y = ops.ggml_mul_mat_vec_a8(qweight, x, qweight_type, qweight.shape[0]) + # elif qweight_type in MMQ_QUANT_TYPES: + # y = ops.ggml_mul_mat_a8(qweight, x, qweight_type, qweight.shape[0]) + + # If there is no available MMQ kernel, fallback to dequantize + if qweight_type in DEQUANT_TYPES: + block_size, type_size = gguf.GGML_QUANT_SIZES[qweight_type] + shape = (qweight.shape[0], qweight.shape[1] // type_size * block_size) + weight = ops.ggml_dequantize(qweight, qweight_type, *shape) + y = x @ weight.to(x.dtype).T + else: + # Raise an error if the quantization type is not supported. + # Might be useful if llama.cpp adds a new quantization type. + # Wrap to GGMLQuantizationType IntEnum to make sure it's a valid type. + qweight_type = gguf.GGMLQuantizationType(qweight_type) + raise NotImplementedError(f"Unsupported GGUF quantization type: {qweight_type}") + return y.as_tensor() + + +# Copied from diffusers.quantizers.bitsandbytes.utils._create_accelerate_new_hook +def _create_accelerate_new_hook(old_hook): + r""" + Creates a new hook based on the old hook. Use it only if you know what you are doing ! This method is a copy of: + https://github.com/huggingface/peft/blob/748f7968f3a31ec06a1c2b0328993319ad9a150a/src/peft/utils/other.py#L245 with + some changes + """ + old_hook_cls = getattr(accelerate.hooks, old_hook.__class__.__name__) + old_hook_attr = old_hook.__dict__ + filtered_old_hook_attr = {} + old_hook_init_signature = inspect.signature(old_hook_cls.__init__) + for k in old_hook_attr.keys(): + if k in old_hook_init_signature.parameters: + filtered_old_hook_attr[k] = old_hook_attr[k] + new_hook = old_hook_cls(**filtered_old_hook_attr) + return new_hook + + +def _replace_with_gguf_linear(model, compute_dtype, state_dict, prefix="", modules_to_not_convert=[]): + def _should_convert_to_gguf(state_dict, prefix): + weight_key = prefix + "weight" + return weight_key in state_dict and isinstance(state_dict[weight_key], GGUFParameter) + + has_children = list(model.children()) + if not has_children: + return + + for name, module in model.named_children(): + if name in modules_to_not_convert: + continue + module_prefix = prefix + name + "." + _replace_with_gguf_linear(module, compute_dtype, state_dict, module_prefix, modules_to_not_convert) + + if ( + isinstance(module, nn.Linear) + and _should_convert_to_gguf(state_dict, module_prefix) + and name not in modules_to_not_convert + ): + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + model._modules[name] = GGUFLinear( + module.in_features, + module.out_features, + module.bias is not None, + compute_dtype=compute_dtype, + ) + model._modules[name].source_cls = type(module) + # Force requires_grad to False to avoid unexpected errors + model._modules[name].requires_grad_(False) + + return model + + +def _dequantize_gguf_and_restore_linear(model, modules_to_not_convert=[]): + for name, module in model.named_children(): + if isinstance(module, GGUFLinear) and name not in modules_to_not_convert: + device = module.weight.device + bias = getattr(module, "bias", None) + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + new_module = nn.Linear( + module.in_features, + module.out_features, + module.bias is not None, + device=device, + ) + new_module.weight = nn.Parameter(dequantize_gguf_tensor(module.weight)) + if bias is not None: + new_module.bias = bias + + # Create a new hook and attach it in case we use accelerate + if hasattr(module, "_hf_hook"): + old_hook = module._hf_hook + new_hook = _create_accelerate_new_hook(old_hook) + + remove_hook_from_module(module) + add_hook_to_module(new_module, new_hook) + + new_module.to(device) + model._modules[name] = new_module + + has_children = list(module.children()) + if has_children: + _dequantize_gguf_and_restore_linear(module, modules_to_not_convert) + + return model + + +# dequantize operations based on torch ports of GGUF dequantize_functions +# from City96 +# more info: https://github.com/city96/ComfyUI-GGUF/blob/main/dequant.py + + +QK_K = 256 +K_SCALE_SIZE = 12 + + +def to_uint32(x): + x = x.view(torch.uint8).to(torch.int32) + return (x[:, 0] | x[:, 1] << 8 | x[:, 2] << 16 | x[:, 3] << 24).unsqueeze(1) + + +def split_block_dims(blocks, *args): + n_max = blocks.shape[1] + dims = list(args) + [n_max - sum(args)] + return torch.split(blocks, dims, dim=1) + + +def get_scale_min(scales): + n_blocks = scales.shape[0] + scales = scales.view(torch.uint8) + scales = scales.reshape((n_blocks, 3, 4)) + + d, m, m_d = torch.split(scales, scales.shape[-2] // 3, dim=-2) + + sc = torch.cat([d & 0x3F, (m_d & 0x0F) | ((d >> 2) & 0x30)], dim=-1) + min = torch.cat([m & 0x3F, (m_d >> 4) | ((m >> 2) & 0x30)], dim=-1) + + return (sc.reshape((n_blocks, 8)), min.reshape((n_blocks, 8))) + + +def dequantize_blocks_Q8_0(blocks, block_size, type_size, dtype=None): + d, x = split_block_dims(blocks, 2) + d = d.view(torch.float16).to(dtype) + x = x.view(torch.int8) + return d * x + + +def dequantize_blocks_Q5_1(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, m, qh, qs = split_block_dims(blocks, 2, 2, 4) + d = d.view(torch.float16).to(dtype) + m = m.view(torch.float16).to(dtype) + qh = to_uint32(qh) + + qh = qh.reshape((n_blocks, 1)) >> torch.arange(32, device=d.device, dtype=torch.int32).reshape(1, 32) + ql = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape(1, 1, 2, 1) + qh = (qh & 1).to(torch.uint8) + ql = (ql & 0x0F).reshape((n_blocks, -1)) + + qs = ql | (qh << 4) + return (d * qs) + m + + +def dequantize_blocks_Q5_0(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, qh, qs = split_block_dims(blocks, 2, 4) + d = d.view(torch.float16).to(dtype) + qh = to_uint32(qh) + + qh = qh.reshape(n_blocks, 1) >> torch.arange(32, device=d.device, dtype=torch.int32).reshape(1, 32) + ql = qs.reshape(n_blocks, -1, 1, block_size // 2) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape(1, 1, 2, 1) + + qh = (qh & 1).to(torch.uint8) + ql = (ql & 0x0F).reshape(n_blocks, -1) + + qs = (ql | (qh << 4)).to(torch.int8) - 16 + return d * qs + + +def dequantize_blocks_Q4_1(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, m, qs = split_block_dims(blocks, 2, 2) + d = d.view(torch.float16).to(dtype) + m = m.view(torch.float16).to(dtype) + + qs = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape(1, 1, 2, 1) + qs = (qs & 0x0F).reshape(n_blocks, -1) + + return (d * qs) + m + + +def dequantize_blocks_Q4_0(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, qs = split_block_dims(blocks, 2) + d = d.view(torch.float16).to(dtype) + + qs = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=d.device, dtype=torch.uint8 + ).reshape((1, 1, 2, 1)) + qs = (qs & 0x0F).reshape((n_blocks, -1)).to(torch.int8) - 8 + return d * qs + + +def dequantize_blocks_Q6_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + ( + ql, + qh, + scales, + d, + ) = split_block_dims(blocks, QK_K // 2, QK_K // 4, QK_K // 16) + + scales = scales.view(torch.int8).to(dtype) + d = d.view(torch.float16).to(dtype) + d = (d * scales).reshape((n_blocks, QK_K // 16, 1)) + + ql = ql.reshape((n_blocks, -1, 1, 64)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 2, 1) + ) + ql = (ql & 0x0F).reshape((n_blocks, -1, 32)) + qh = qh.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 2, 4, 6], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 4, 1) + ) + qh = (qh & 0x03).reshape((n_blocks, -1, 32)) + q = (ql | (qh << 4)).to(torch.int8) - 32 + q = q.reshape((n_blocks, QK_K // 16, -1)) + + return (d * q).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q5_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, dmin, scales, qh, qs = split_block_dims(blocks, 2, 2, K_SCALE_SIZE, QK_K // 8) + + d = d.view(torch.float16).to(dtype) + dmin = dmin.view(torch.float16).to(dtype) + + sc, m = get_scale_min(scales) + + d = (d * sc).reshape((n_blocks, -1, 1)) + dm = (dmin * m).reshape((n_blocks, -1, 1)) + + ql = qs.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 2, 1) + ) + qh = qh.reshape((n_blocks, -1, 1, 32)) >> torch.arange(0, 8, device=d.device, dtype=torch.uint8).reshape( + (1, 1, 8, 1) + ) + ql = (ql & 0x0F).reshape((n_blocks, -1, 32)) + qh = (qh & 0x01).reshape((n_blocks, -1, 32)) + q = ql | (qh << 4) + + return (d * q - dm).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q4_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + d, dmin, scales, qs = split_block_dims(blocks, 2, 2, K_SCALE_SIZE) + d = d.view(torch.float16).to(dtype) + dmin = dmin.view(torch.float16).to(dtype) + + sc, m = get_scale_min(scales) + + d = (d * sc).reshape((n_blocks, -1, 1)) + dm = (dmin * m).reshape((n_blocks, -1, 1)) + + qs = qs.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 2, 1) + ) + qs = (qs & 0x0F).reshape((n_blocks, -1, 32)) + + return (d * qs - dm).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q3_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + hmask, qs, scales, d = split_block_dims(blocks, QK_K // 8, QK_K // 4, 12) + d = d.view(torch.float16).to(dtype) + + lscales, hscales = scales[:, :8], scales[:, 8:] + lscales = lscales.reshape((n_blocks, 1, 8)) >> torch.tensor([0, 4], device=d.device, dtype=torch.uint8).reshape( + (1, 2, 1) + ) + lscales = lscales.reshape((n_blocks, 16)) + hscales = hscales.reshape((n_blocks, 1, 4)) >> torch.tensor( + [0, 2, 4, 6], device=d.device, dtype=torch.uint8 + ).reshape((1, 4, 1)) + hscales = hscales.reshape((n_blocks, 16)) + scales = (lscales & 0x0F) | ((hscales & 0x03) << 4) + scales = scales.to(torch.int8) - 32 + + dl = (d * scales).reshape((n_blocks, 16, 1)) + + ql = qs.reshape((n_blocks, -1, 1, 32)) >> torch.tensor([0, 2, 4, 6], device=d.device, dtype=torch.uint8).reshape( + (1, 1, 4, 1) + ) + qh = hmask.reshape(n_blocks, -1, 1, 32) >> torch.arange(0, 8, device=d.device, dtype=torch.uint8).reshape( + (1, 1, 8, 1) + ) + ql = ql.reshape((n_blocks, 16, QK_K // 16)) & 3 + qh = (qh.reshape((n_blocks, 16, QK_K // 16)) & 1) ^ 1 + q = ql.to(torch.int8) - (qh << 2).to(torch.int8) + + return (dl * q).reshape((n_blocks, QK_K)) + + +def dequantize_blocks_Q2_K(blocks, block_size, type_size, dtype=None): + n_blocks = blocks.shape[0] + + scales, qs, d, dmin = split_block_dims(blocks, QK_K // 16, QK_K // 4, 2) + d = d.view(torch.float16).to(dtype) + dmin = dmin.view(torch.float16).to(dtype) + + # (n_blocks, 16, 1) + dl = (d * (scales & 0xF)).reshape((n_blocks, QK_K // 16, 1)) + ml = (dmin * (scales >> 4)).reshape((n_blocks, QK_K // 16, 1)) + + shift = torch.tensor([0, 2, 4, 6], device=d.device, dtype=torch.uint8).reshape((1, 1, 4, 1)) + + qs = (qs.reshape((n_blocks, -1, 1, 32)) >> shift) & 3 + qs = qs.reshape((n_blocks, QK_K // 16, 16)) + qs = dl * qs - ml + + return qs.reshape((n_blocks, -1)) + + +def dequantize_blocks_BF16(blocks, block_size, type_size, dtype=None): + return (blocks.view(torch.int16).to(torch.int32) << 16).view(torch.float32) + + +# this part from calcuis (gguf.org) +# more info: https://github.com/calcuis/gguf-connector/blob/main/src/gguf_connector/quant2c.py + + +def dequantize_blocks_IQ4_NL(blocks, block_size, type_size, dtype=None): + kvalues = torch.tensor( + [-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113], + dtype=torch.float32, + device=blocks.device, + ) + n_blocks = blocks.shape[0] + d, qs = split_block_dims(blocks, 2) + d = d.view(torch.float16).to(dtype) + qs = qs.reshape((n_blocks, -1, 1, block_size // 2)) >> torch.tensor( + [0, 4], device=blocks.device, dtype=torch.uint8 + ).reshape((1, 1, 2, 1)) + qs = (qs & 15).reshape((n_blocks, -1)).to(torch.int64) + kvalues = kvalues.view(1, 1, 16) + qs = qs.unsqueeze(-1) + qs = torch.gather(kvalues.expand(qs.shape[0], qs.shape[1], 16), 2, qs) + qs = qs.squeeze(-1).to(dtype) + return d * qs + + +def dequantize_blocks_IQ4_XS(blocks, block_size, type_size, dtype=None): + kvalues = torch.tensor( + [-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113], + dtype=torch.float32, + device=blocks.device, + ) + n_blocks = blocks.shape[0] + d, scales_h, scales_l, qs = split_block_dims(blocks, 2, 2, QK_K // 64) + d = d.view(torch.float16).to(dtype) + scales_h = scales_h.view(torch.int16) + scales_l = scales_l.reshape((n_blocks, -1, 1)) >> torch.tensor( + [0, 4], device=blocks.device, dtype=torch.uint8 + ).reshape((1, 1, 2)) + scales_h = scales_h.reshape((n_blocks, 1, -1)) >> torch.tensor( + [2 * i for i in range(QK_K // 32)], device=blocks.device, dtype=torch.uint8 + ).reshape((1, -1, 1)) + scales_l = scales_l.reshape((n_blocks, -1)) & 0x0F + scales_h = scales_h.reshape((n_blocks, -1)) & 0x03 + scales = (scales_l | (scales_h << 4)) - 32 + dl = (d * scales.to(dtype)).reshape((n_blocks, -1, 1)) + shifts_q = torch.tensor([0, 4], device=blocks.device, dtype=torch.uint8).reshape(1, 1, 2, 1) + qs = qs.reshape((n_blocks, -1, 1, 16)) >> shifts_q + qs = (qs & 15).reshape((n_blocks, -1, 32)).to(torch.int64) + kvalues = kvalues.view(1, 1, 1, 16) + qs = qs.unsqueeze(-1) + qs = torch.gather(kvalues.expand(qs.shape[0], qs.shape[1], qs.shape[2], 16), 3, qs) + qs = qs.squeeze(-1).to(dtype) + return (dl * qs).reshape(n_blocks, -1) + + +GGML_QUANT_SIZES = gguf.GGML_QUANT_SIZES +dequantize_functions = { + gguf.GGMLQuantizationType.IQ4_NL: dequantize_blocks_IQ4_NL, + gguf.GGMLQuantizationType.IQ4_XS: dequantize_blocks_IQ4_XS, + gguf.GGMLQuantizationType.BF16: dequantize_blocks_BF16, + gguf.GGMLQuantizationType.Q8_0: dequantize_blocks_Q8_0, + gguf.GGMLQuantizationType.Q5_1: dequantize_blocks_Q5_1, + gguf.GGMLQuantizationType.Q5_0: dequantize_blocks_Q5_0, + gguf.GGMLQuantizationType.Q4_1: dequantize_blocks_Q4_1, + gguf.GGMLQuantizationType.Q4_0: dequantize_blocks_Q4_0, + gguf.GGMLQuantizationType.Q6_K: dequantize_blocks_Q6_K, + gguf.GGMLQuantizationType.Q5_K: dequantize_blocks_Q5_K, + gguf.GGMLQuantizationType.Q4_K: dequantize_blocks_Q4_K, + gguf.GGMLQuantizationType.Q3_K: dequantize_blocks_Q3_K, + gguf.GGMLQuantizationType.Q2_K: dequantize_blocks_Q2_K, +} +SUPPORTED_GGUF_QUANT_TYPES = list(dequantize_functions.keys()) + + +def _quant_shape_from_byte_shape(shape, type_size, block_size): + return (*shape[:-1], shape[-1] // type_size * block_size) + + +def dequantize_gguf_tensor(tensor): + if not hasattr(tensor, "quant_type"): + return tensor + + quant_type = tensor.quant_type + dequant_fn = dequantize_functions[quant_type] + + block_size, type_size = GGML_QUANT_SIZES[quant_type] + + # Conver to plain tensor to avoid unnecessary __torch_function__ overhead. + tensor = tensor.as_tensor() + + tensor = tensor.view(torch.uint8) + shape = _quant_shape_from_byte_shape(tensor.shape, type_size, block_size) + + n_blocks = tensor.numel() // type_size + blocks = tensor.reshape((n_blocks, type_size)) + + dequant = dequant_fn(blocks, block_size, type_size) + dequant = dequant.reshape(shape) + + return dequant + + +class GGUFParameter(torch.nn.Parameter): + def __new__(cls, data, requires_grad=False, quant_type=None): + data = data if data is not None else torch.empty(0) + self = torch.Tensor._make_subclass(cls, data, requires_grad) + self.quant_type = quant_type + block_size, type_size = GGML_QUANT_SIZES[quant_type] + self.quant_shape = _quant_shape_from_byte_shape(self.shape, type_size, block_size) + + return self + + def as_tensor(self): + return torch.Tensor._make_subclass(torch.Tensor, self, self.requires_grad) + + @staticmethod + def _extract_quant_type(args): + # When converting from original format checkpoints we often use splits, cats etc on tensors + # this method ensures that the returned tensor type from those operations remains GGUFParameter + # so that we preserve quant_type information + for arg in args: + if isinstance(arg, list) and isinstance(arg[0], GGUFParameter): + return arg[0].quant_type + if isinstance(arg, GGUFParameter): + return arg.quant_type + return None + + @classmethod + def __torch_function__(cls, func, types, args=(), kwargs=None): + if kwargs is None: + kwargs = {} + + result = super().__torch_function__(func, types, args, kwargs) + + if isinstance(result, torch.Tensor): + quant_type = cls._extract_quant_type(args) + return cls(result, quant_type=quant_type) + # Handle tuples and lists + elif type(result) in (list, tuple): + # Preserve the original type (tuple or list) + quant_type = cls._extract_quant_type(args) + wrapped = [cls(x, quant_type=quant_type) if isinstance(x, torch.Tensor) else x for x in result] + return type(result)(wrapped) + else: + return result + + +class GGUFLinear(nn.Linear): + def __init__( + self, + in_features, + out_features, + bias=False, + compute_dtype=None, + device=None, + ) -> None: + super().__init__(in_features, out_features, bias, device) + self.compute_dtype = compute_dtype + self.device = device + + def forward(self, inputs: torch.Tensor): + if ops is not None and self.weight.is_cuda and inputs.is_cuda: + return self.forward_cuda(inputs) + return self.forward_native(inputs) + + def forward_native(self, inputs: torch.Tensor): + weight = dequantize_gguf_tensor(self.weight) + weight = weight.to(self.compute_dtype) + bias = self.bias.to(self.compute_dtype) if self.bias is not None else None + + output = torch.nn.functional.linear(inputs, weight, bias) + return output + + def forward_cuda(self, inputs: torch.Tensor): + quant_type = self.weight.quant_type + output = _fused_mul_mat_gguf(inputs.to(self.compute_dtype), self.weight, quant_type) + if self.bias is not None: + output += self.bias.to(self.compute_dtype) + return output diff --git a/diffusers/quantizers/modelopt/__init__.py b/diffusers/quantizers/modelopt/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..ae0951cb30d140011f6b3595036f3ca225b423f4 --- /dev/null +++ b/diffusers/quantizers/modelopt/__init__.py @@ -0,0 +1 @@ +from .modelopt_quantizer import NVIDIAModelOptQuantizer diff --git a/diffusers/quantizers/modelopt/modelopt_quantizer.py b/diffusers/quantizers/modelopt/modelopt_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..b7cadb50e54b6d64ad1993ccf319e9a450bc10b7 --- /dev/null +++ b/diffusers/quantizers/modelopt/modelopt_quantizer.py @@ -0,0 +1,196 @@ +from typing import TYPE_CHECKING, Any + +from ...utils import ( + get_module_from_name, + is_accelerate_available, + is_nvidia_modelopt_available, + is_torch_available, + logging, +) +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +if is_torch_available(): + import torch + import torch.nn as nn + +if is_accelerate_available(): + from accelerate.utils import set_module_tensor_to_device + + +logger = logging.get_logger(__name__) + + +class NVIDIAModelOptQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for Nvidia-Model Optimizer + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + required_packages = ["nvidia_modelopt"] + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + def validate_environment(self, *args, **kwargs): + if not is_nvidia_modelopt_available(): + raise ImportError( + "Loading an nvidia-modelopt quantized model requires nvidia-modelopt library (`pip install nvidia-modelopt`)" + ) + + self.offload = False + + device_map = kwargs.get("device_map", None) + if isinstance(device_map, dict): + if "cpu" in device_map.values() or "disk" in device_map.values(): + if self.pre_quantized: + raise ValueError( + "You are attempting to perform cpu/disk offload with a pre-quantized modelopt model " + "This is not supported yet. Please remove the CPU or disk device from the `device_map` argument." + ) + else: + self.offload = True + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ): + # ModelOpt imports diffusers internally. This is here to prevent circular imports + from modelopt.torch.quantization.utils import is_quantized + + module, tensor_name = get_module_from_name(model, param_name) + if self.pre_quantized: + # ModelOpt restoration recreates quantizer state such as `_amax` and + # `_scale` as buffers. Let the regular low-memory loader materialize + # those buffers on their target device. Only parameters need the + # wrapper-aware assignment below. + return tensor_name in module._parameters + elif is_quantized(module) and "weight" in tensor_name: + return True + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + *args, + **kwargs, + ): + """ + Create the quantized parameter by calling .calibrate() after setting it to the module. + """ + # ModelOpt imports diffusers internally. This is here to prevent circular imports + import modelopt.torch.quantization as mtq + + dtype = kwargs.get("dtype", torch.float32) + module, tensor_name = get_module_from_name(model, param_name) + if self.pre_quantized: + if tensor_name not in module._parameters: + raise ValueError(f"Expected {param_name} to be a parameter in the restored ModelOpt graph.") + module._parameters[tensor_name] = torch.nn.Parameter(param_value.to(device=target_device)) + else: + set_module_tensor_to_device(model, param_name, target_device, param_value, dtype) + mtq.calibrate( + module, self.quantization_config.modelopt_config["algorithm"], self.quantization_config.forward_loop + ) + mtq.compress(module) + module.weight.requires_grad = False + + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if self.quantization_config.quant_type == "FP8": + target_dtype = torch.float8_e4m3fn + return target_dtype + + def update_torch_dtype(self, torch_dtype: "torch.dtype" = None) -> "torch.dtype": + if torch_dtype is None: + logger.info("You did not specify `torch_dtype` in `from_pretrained`. Setting it to `torch.float32`.") + torch_dtype = torch.float32 + return torch_dtype + + def get_conv_param_names(self, model: "ModelMixin") -> list[str]: + """ + Get parameter names for all convolutional layers in a HuggingFace ModelMixin. Includes Conv1d/2d/3d and + ConvTranspose1d/2d/3d. + """ + conv_types = ( + nn.Conv1d, + nn.Conv2d, + nn.Conv3d, + nn.ConvTranspose1d, + nn.ConvTranspose2d, + nn.ConvTranspose3d, + ) + + conv_param_names = [] + for name, module in model.named_modules(): + if isinstance(module, conv_types): + for param_name, _ in module.named_parameters(recurse=False): + conv_param_names.append(f"{name}.{param_name}") + + return conv_param_names + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + # ModelOpt imports diffusers internally. This is here to prevent circular imports + import modelopt.torch.opt as mto + + if self.pre_quantized: + return + + modules_to_not_convert = self.quantization_config.modules_to_not_convert + + if modules_to_not_convert is None: + modules_to_not_convert = [] + if isinstance(modules_to_not_convert, str): + modules_to_not_convert = [modules_to_not_convert] + modules_to_not_convert.extend(keep_in_fp32_modules) + if self.quantization_config.disable_conv_quantization: + modules_to_not_convert.extend(self.get_conv_param_names(model)) + + for module in modules_to_not_convert: + self.quantization_config.modelopt_config["quant_cfg"]["*" + module + "*"] = {"enable": False} + self.quantization_config.modules_to_not_convert = modules_to_not_convert + mto.apply_mode(model, mode=[("quantize", self.quantization_config.modelopt_config)]) + model.config.quantization_config = self.quantization_config + + def _process_model_after_weight_loading(self, model, **kwargs): + # ModelOpt imports diffusers internally. This is here to prevent circular imports + from modelopt.torch.opt import ModeloptStateManager + + if self.pre_quantized: + return model + + for _, m in model.named_modules(): + if hasattr(m, ModeloptStateManager._state_key) and m is not model: + ModeloptStateManager.remove_state(m) + + return model + + @property + def is_trainable(self): + return True + + @property + def is_serializable(self): + self.quantization_config.check_model_patching(operation="saving") + return True diff --git a/diffusers/quantizers/nunchaku/__init__.py b/diffusers/quantizers/nunchaku/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..807b053ff56c1ab5bb23dad20cecbe991bd0147b --- /dev/null +++ b/diffusers/quantizers/nunchaku/__init__.py @@ -0,0 +1 @@ +from .nunchaku_quantizer import NunchakuLiteQuantizer diff --git a/diffusers/quantizers/nunchaku/nunchaku_quantizer.py b/diffusers/quantizers/nunchaku/nunchaku_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..b8f20d7ddba3de16f393da3c3f47e51da9116423 --- /dev/null +++ b/diffusers/quantizers/nunchaku/nunchaku_quantizer.py @@ -0,0 +1,89 @@ +from __future__ import annotations + +from typing import TYPE_CHECKING, Any + +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +from ...utils import is_kernels_available, logging + + +logger = logging.get_logger(__name__) + + +class NunchakuLiteQuantizer(DiffusersQuantizer): + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + self.compute_dtype = quantization_config.compute_dtype + self.pre_quantized = quantization_config.pre_quantized + + def validate_environment(self, *args, **kwargs): + if not is_kernels_available(): + raise ImportError( + "Loading Nunchaku checkpoints requires the Hugging Face `kernels` package. " + "Install it with `pip install kernels`." + ) + import torch + + cuda_available = torch.cuda.is_available() + if not cuda_available: + raise ValueError("Loading Nunchaku checkpoints requires a CUDA-capable NVIDIA GPU.") + + device_capability = torch.cuda.get_device_capability() + + if device_capability[0] == 9: + raise ValueError("Loading Nunchaku checkpoints is not supported on Hopper NVIDIA GPUs.") + + has_nvfp4_config = ( + self.quantization_config.svdq_w4a4 is not None + and self.quantization_config.svdq_w4a4["precision"] == "nvfp4" + ) + has_int4_config = any( + config is not None and config["precision"] == "int4" + for config in (self.quantization_config.svdq_w4a4, self.quantization_config.awq_w4a16) + ) + if has_nvfp4_config and device_capability < (10, 0): + raise ValueError("Loading Nunchaku NVFP4 checkpoints requires a Blackwell or newer NVIDIA GPU.") + if has_int4_config and device_capability < (7, 5): + raise ValueError("Loading Nunchaku INT4 checkpoints on CUDA requires a Turing or newer NVIDIA GPU.") + + def update_torch_dtype(self, torch_dtype): + if torch_dtype is None: + torch_dtype = self.compute_dtype + else: + self.compute_dtype = torch_dtype + return torch_dtype + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + state_dict: dict[str, Any] | None = None, + **kwargs, + ): + from .utils import check_strict_state_dict_match, replace_with_nunchaku_linear + + quantization_config = self.quantization_config.to_dict() + num_replaced = replace_with_nunchaku_linear(model, quantization_config, self.compute_dtype) + + if state_dict is not None: + check_strict_state_dict_match(model, state_dict) + logger.info(f"Applied Nunchaku quantization config with {num_replaced} targets.") + + def _process_model_after_weight_loading(self, model: "ModelMixin", **kwargs): + return model + + @property + def is_serializable(self): + return False + + @property + def is_trainable(self) -> bool: + return False + + @property + def is_compileable(self) -> bool: + return True diff --git a/diffusers/quantizers/nunchaku/utils.py b/diffusers/quantizers/nunchaku/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..eda8024377656052232311c53c8f42576dbea271 --- /dev/null +++ b/diffusers/quantizers/nunchaku/utils.py @@ -0,0 +1,349 @@ +from __future__ import annotations + +import itertools +import math +from contextlib import nullcontext +from typing import Any + +import torch +import torch.nn as nn + +from ...utils import is_accelerate_available, is_kernels_available + + +if is_accelerate_available(): + from accelerate import init_empty_weights + + +_HF_KERNEL_REPO = "rootonchair/nunchaku-lite-kernels" +_HF_KERNEL_VERSION = 2 + + +if is_kernels_available(): + from kernels import get_kernel + + ops = get_kernel(_HF_KERNEL_REPO, version=_HF_KERNEL_VERSION, trust_remote_code=True).ops +else: + raise ImportError( + "Loading Nunchaku checkpoints requires the Hugging Face `kernels` package. " + "Install it with `pip install kernels`." + ) + + +def _gemm_w4a4( + act: torch.Tensor, + wgt: torch.Tensor, + out: torch.Tensor, + ascales: torch.Tensor, + wscales: torch.Tensor, + lora_act_in: torch.Tensor, + lora_up: torch.Tensor, + bias: torch.Tensor | None, + act_unsigned: bool, + lora_scales: list[float], + nvfp4: bool, + alpha: torch.Tensor | None, + wcscales: torch.Tensor | None, +) -> None: + ops.gemm_w4a4( + act, + wgt, + out, + None, + ascales, + wscales, + None, + None, + lora_act_in, + lora_up, + None, + None, + None, + None, + None, + bias, + None, + None, + None, + act_unsigned, + lora_scales, + False, + nvfp4, + alpha, + wcscales, + None, + None, + None, + 0, + ) + + +def replace_with_nunchaku_linear( + model: nn.Module, quantization_config: dict[str, Any], compute_dtype: torch.dtype +) -> int: + num_replaced = 0 + svdq_config = quantization_config.get("svdq_w4a4") + awq_config = quantization_config.get("awq_w4a16") + + if svdq_config is not None: + num_replaced += _replace_quantize_targets(model, "svdq_w4a4", svdq_config, compute_dtype) + if awq_config is not None: + num_replaced += _replace_quantize_targets(model, "awq_w4a16", awq_config, compute_dtype) + if num_replaced == 0: + raise ValueError( + "Nunchaku compact quantization config must include `svdq_w4a4.targets` or `awq_w4a16.targets`." + ) + + return num_replaced + + +class SVDQW4A4Linear(nn.Module): + def __init__( + self, + in_features: int, + out_features: int, + rank: int = 32, + bias: bool = True, + precision: str = "int4", + group_size: int = 64, + torch_dtype: torch.dtype = torch.bfloat16, + device: str | torch.device | None = None, + act_unsigned: bool = False, + ): + super().__init__() + + self.in_features = in_features + self.out_features = out_features + self.rank = rank + self.precision = precision + self.group_size = group_size + self.torch_dtype = torch_dtype + self.act_unsigned = act_unsigned + + self.qweight = nn.Parameter( + torch.empty(out_features, in_features // 2, dtype=torch.int8, device=device), requires_grad=False + ) + self.bias = ( + nn.Parameter(torch.empty(out_features, dtype=torch_dtype, device=device), requires_grad=False) + if bias + else None + ) + self.wscales = nn.Parameter( + torch.empty( + in_features // group_size, + out_features, + dtype=torch_dtype if precision == "int4" else torch.float8_e4m3fn, + device=device, + ), + requires_grad=False, + ) + self.smooth_factor = nn.Parameter( + torch.empty(in_features, dtype=torch_dtype, device=device), requires_grad=False + ) + self.proj_down = nn.Parameter( + torch.empty(in_features, rank, dtype=torch_dtype, device=device), requires_grad=False + ) + self.proj_up = nn.Parameter( + torch.empty(out_features, rank, dtype=torch_dtype, device=device), requires_grad=False + ) + + if precision == "nvfp4": + self.wcscales = nn.Parameter( + torch.ones(out_features, dtype=torch_dtype, device=device), requires_grad=False + ) + self.wtscale = nn.Parameter(torch.ones(1, dtype=torch_dtype, device=device), requires_grad=False) + else: + self.wcscales = None + self.wtscale = None + + def forward(self, x: torch.Tensor, output: torch.Tensor | None = None) -> torch.Tensor: + original_shape = x.shape + channels = x.shape[-1] + x = x.reshape(-1, channels) + rows = x.shape[0] + if output is None: + output = torch.empty(rows, self.out_features, dtype=self.torch_dtype, device=x.device) + + pad_size = 256 + batch_size_pad = math.ceil(x.shape[0] / pad_size) * pad_size + quantized_x = torch.empty(batch_size_pad, channels // 2, dtype=torch.uint8, device=x.device) + if self.precision == "nvfp4": + ascales = torch.empty(channels // 16, batch_size_pad, dtype=torch.float8_e4m3fn, device=x.device) + else: + ascales = torch.empty(channels // 64, batch_size_pad, dtype=x.dtype, device=x.device) + lora_act = torch.empty(batch_size_pad, self.rank, dtype=torch.float32, device=x.device) + + ops.quantize_w4a4_act_fuse_lora( + x, + quantized_x, + ascales, + self.proj_down, + lora_act, + self.smooth_factor, + False, + self.precision == "nvfp4", + ) + lora_scales = [1.0] * math.ceil(self.rank / 16) + _gemm_w4a4( + quantized_x, + self.qweight, + output, + ascales, + self.wscales, + lora_act, + self.proj_up, + self.bias, + self.act_unsigned, + lora_scales, + self.precision == "nvfp4", + self.wtscale, + self.wcscales, + ) + return output.reshape(*original_shape[:-1], self.out_features) + + +class AWQW4A16Linear(nn.Module): + def __init__( + self, + in_features: int, + out_features: int, + bias: bool = True, + group_size: int = 64, + torch_dtype: torch.dtype = torch.bfloat16, + device: str | torch.device | None = None, + ): + super().__init__() + + self.in_features = in_features + self.out_features = out_features + self.group_size = group_size + + self.qweight = nn.Parameter( + torch.empty(out_features // 4, in_features // 2, dtype=torch.int32, device=device), requires_grad=False + ) + self.bias = ( + nn.Parameter(torch.empty(out_features, dtype=torch_dtype, device=device), requires_grad=False) + if bias + else None + ) + self.wscales = nn.Parameter( + torch.empty(in_features // group_size, out_features, dtype=torch_dtype, device=device), requires_grad=False + ) + self.wzeros = nn.Parameter( + torch.empty(in_features // group_size, out_features, dtype=torch_dtype, device=device), requires_grad=False + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if x.shape[-1] != self.in_features: + raise ValueError( + f"AWQW4A16Linear expected input last dimension {self.in_features}, got shape {tuple(x.shape)}." + ) + + output_shape = (*x.shape[:-1], self.out_features) + x_flat = x.reshape(-1, self.in_features).contiguous() + if x_flat.shape[0] == 0: + output = x.new_empty(output_shape) + elif self._use_gemm(x_flat.shape[0]): + output = ops.awq_gemm_w4a16_g64_int32(x_flat, self.qweight, self.wscales, self.wzeros).reshape( + output_shape + ) + else: + output = self._forward_gemv_chunks(x_flat, ops.gemv_awq).reshape(output_shape) + + if self.bias is not None: + output = output + self.bias.view([1] * (output.ndim - 1) + [-1]) + return output + + def _use_gemm(self, rows: int) -> bool: + return rows >= 16 and self.in_features % 64 == 0 and self.out_features % 128 == 0 + + def _forward_gemv_chunks(self, x_flat: torch.Tensor, gemv) -> torch.Tensor: + outputs = [] + for start in range(0, x_flat.shape[0], 8): + chunk = x_flat[start : start + 8] + outputs.append( + gemv( + chunk, + self.qweight, + self.wscales, + self.wzeros, + chunk.shape[0], + self.out_features, + self.in_features, + 64, + ) + ) + return torch.cat(outputs, dim=0) + + +def _replace_quantize_targets(model: nn.Module, op: str, raw: Any, compute_dtype: torch.dtype) -> int: + precision = raw["precision"] + group_size = raw["group_size"] + targets = raw["targets"] + rank = raw["rank"] if op == "svdq_w4a4" else 0 + + for target in targets: + try: + module = model.get_submodule(target) + except AttributeError as exc: + raise ValueError(f"Nunchaku target {target!r} does not exist in the model.") from exc + + in_features = getattr(module, "in_features", None) + out_features = getattr(module, "out_features", None) + bias = getattr(module, "bias", None) + if not isinstance(in_features, int) or not isinstance(out_features, int): + raise TypeError(f"Nunchaku target {target!r} must expose integer in_features/out_features.") + + ctx = init_empty_weights if is_accelerate_available() else nullcontext + with ctx(): + if op == "svdq_w4a4": + replacement = SVDQW4A4Linear( + in_features, + out_features, + rank=rank, + bias=bias is not None, + precision=precision, + group_size=group_size, + torch_dtype=compute_dtype, + ) + elif op == "awq_w4a16": + replacement = AWQW4A16Linear( + in_features, + out_features, + bias=bias is not None, + group_size=group_size, + torch_dtype=compute_dtype, + ) + + _set_submodule(model, target, replacement) + + return len(targets) + + +def _set_submodule(model: nn.Module, path: str, module: nn.Module) -> None: + parent_path, _, child_name = path.rpartition(".") + parent = model.get_submodule(parent_path) if parent_path else model + if child_name.isdigit() and isinstance(parent, (nn.Sequential, nn.ModuleList)): + parent[int(child_name)] = module + else: + setattr(parent, child_name, module) + + +def check_strict_state_dict_match(model: nn.Module, state_dict: dict[str, Any]) -> None: + expected_keys = {n for n, _ in itertools.chain(model.named_parameters(), model.named_buffers())} + loaded_keys = set(state_dict.keys()) + missing_keys = sorted(expected_keys - loaded_keys) + unexpected_keys = sorted(loaded_keys - expected_keys) + if missing_keys or unexpected_keys: + message = "Nunchaku checkpoint keys must exactly match the patched model state dict." + if missing_keys: + message += f" Missing keys: {missing_keys[:10]}" + if len(missing_keys) > 10: + message += f" and {len(missing_keys) - 10} more" + message += "." + if unexpected_keys: + message += f" Unexpected keys: {unexpected_keys[:10]}" + if len(unexpected_keys) > 10: + message += f" and {len(unexpected_keys) - 10} more" + message += "." + raise ValueError(message) diff --git a/diffusers/quantizers/pipe_quant_config.py b/diffusers/quantizers/pipe_quant_config.py new file mode 100644 index 0000000000000000000000000000000000000000..cfb5803ee0ca4ee748d9bc6c14075cc6039d65b0 --- /dev/null +++ b/diffusers/quantizers/pipe_quant_config.py @@ -0,0 +1,206 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import inspect + +from ..utils import is_transformers_available, logging +from .quantization_config import QuantizationConfigMixin as DiffQuantConfigMixin + + +try: + from transformers.utils.quantization_config import QuantizationConfigMixin as TransformersQuantConfigMixin +except ImportError: + + class TransformersQuantConfigMixin: + pass + + +logger = logging.get_logger(__name__) + + +class PipelineQuantizationConfig: + """ + Configuration class to be used when applying quantization on-the-fly to [`~DiffusionPipeline.from_pretrained`]. + + Args: + quant_backend (`str`): Quantization backend to be used. When using this option, we assume that the backend + is available to both `diffusers` and `transformers`. + quant_kwargs (`dict`): Params to initialize the quantization backend class. + components_to_quantize (`list`): Components of a pipeline to be quantized. + quant_mapping (`dict`): Mapping defining the quantization specs to be used for the pipeline + components. When using this argument, users are not expected to provide `quant_backend`, `quant_kawargs`, + and `components_to_quantize`. + """ + + def __init__( + self, + quant_backend: str = None, + quant_kwargs: dict[str, str | float | int | dict] = None, + components_to_quantize: list[str] | str | None = None, + quant_mapping: dict[str, DiffQuantConfigMixin | "TransformersQuantConfigMixin"] = None, + ): + self.quant_backend = quant_backend + # Initialize kwargs to be {} to set to the defaults. + self.quant_kwargs = quant_kwargs or {} + if components_to_quantize: + if isinstance(components_to_quantize, str): + components_to_quantize = [components_to_quantize] + self.components_to_quantize = components_to_quantize + self.quant_mapping = quant_mapping + self.config_mapping = {} # book-keeping Example: `{module_name: quant_config}` + self.post_init() + + def post_init(self): + quant_mapping = self.quant_mapping + self.is_granular = True if quant_mapping is not None else False + + self._validate_init_args() + + def _validate_init_args(self): + if self.quant_backend and self.quant_mapping: + raise ValueError("Both `quant_backend` and `quant_mapping` cannot be specified at the same time.") + + if not self.quant_mapping and not self.quant_backend: + raise ValueError("Must provide a `quant_backend` when not providing a `quant_mapping`.") + + if not self.quant_kwargs and not self.quant_mapping: + raise ValueError("Both `quant_kwargs` and `quant_mapping` cannot be None.") + + if self.quant_backend is not None: + self._validate_init_kwargs_in_backends() + + if self.quant_mapping is not None: + self._validate_quant_mapping_args() + + def _validate_init_kwargs_in_backends(self): + quant_backend = self.quant_backend + + self._check_backend_availability(quant_backend) + + quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list() + + if quant_config_mapping_transformers is not None: + init_kwargs_transformers = inspect.signature(quant_config_mapping_transformers[quant_backend].__init__) + init_kwargs_transformers = {name for name in init_kwargs_transformers.parameters if name != "self"} + else: + init_kwargs_transformers = None + + init_kwargs_diffusers = inspect.signature(quant_config_mapping_diffusers[quant_backend].__init__) + init_kwargs_diffusers = {name for name in init_kwargs_diffusers.parameters if name != "self"} + + if init_kwargs_transformers != init_kwargs_diffusers: + raise ValueError( + "The signatures of the __init__ methods of the quantization config classes in `diffusers` and `transformers` don't match. " + f"Please provide a `quant_mapping` instead, in the {self.__class__.__name__} class. Refer to [the docs](https://huggingface.co/docs/diffusers/main/en/quantization/overview#pipeline-level-quantization) to learn more about how " + "this mapping would look like." + ) + + def _validate_quant_mapping_args(self): + quant_mapping = self.quant_mapping + transformers_map, diffusers_map = self._get_quant_config_list() + + available_transformers = list(transformers_map.values()) if transformers_map else None + available_diffusers = list(diffusers_map.values()) + + for module_name, config in quant_mapping.items(): + if any(isinstance(config, cfg) for cfg in available_diffusers): + continue + + if available_transformers and any(isinstance(config, cfg) for cfg in available_transformers): + continue + + if available_transformers: + raise ValueError( + f"Provided config for module_name={module_name} could not be found. " + f"Available diffusers configs: {available_diffusers}; " + f"Available transformers configs: {available_transformers}." + ) + else: + raise ValueError( + f"Provided config for module_name={module_name} could not be found. " + f"Available diffusers configs: {available_diffusers}." + ) + + def _check_backend_availability(self, quant_backend: str): + quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list() + + available_backends_transformers = ( + list(quant_config_mapping_transformers.keys()) if quant_config_mapping_transformers else None + ) + available_backends_diffusers = list(quant_config_mapping_diffusers.keys()) + + if ( + available_backends_transformers and quant_backend not in available_backends_transformers + ) or quant_backend not in quant_config_mapping_diffusers: + error_message = f"Provided quant_backend={quant_backend} was not found." + if available_backends_transformers: + error_message += f"\nAvailable ones (transformers): {available_backends_transformers}." + error_message += f"\nAvailable ones (diffusers): {available_backends_diffusers}." + raise ValueError(error_message) + + def _resolve_quant_config(self, is_diffusers: bool = True, module_name: str = None): + quant_config_mapping_transformers, quant_config_mapping_diffusers = self._get_quant_config_list() + + quant_mapping = self.quant_mapping + components_to_quantize = self.components_to_quantize + + # Granular case + if self.is_granular and module_name in quant_mapping: + logger.debug(f"Initializing quantization config class for {module_name}.") + config = quant_mapping[module_name] + self.config_mapping.update({module_name: config}) + return config + + # Global config case + else: + should_quantize = False + # Only quantize the modules requested for. + if components_to_quantize and module_name in components_to_quantize: + should_quantize = True + # No specification for `components_to_quantize` means all modules should be quantized. + elif not self.is_granular and not components_to_quantize: + should_quantize = True + + if should_quantize: + logger.debug(f"Initializing quantization config class for {module_name}.") + mapping_to_use = quant_config_mapping_diffusers if is_diffusers else quant_config_mapping_transformers + quant_config_cls = mapping_to_use[self.quant_backend] + quant_kwargs = self.quant_kwargs + quant_obj = quant_config_cls(**quant_kwargs) + self.config_mapping.update({module_name: quant_obj}) + return quant_obj + + # Fallback: no applicable configuration found. + return None + + def _get_quant_config_list(self): + if is_transformers_available(): + from transformers.quantizers.auto import ( + AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_transformers, + ) + else: + quant_config_mapping_transformers = None + + from ..quantizers.auto import AUTO_QUANTIZATION_CONFIG_MAPPING as quant_config_mapping_diffusers + + return quant_config_mapping_transformers, quant_config_mapping_diffusers + + def __repr__(self): + out = "" + config_mapping = dict(sorted(self.config_mapping.copy().items())) + for module_name, config in config_mapping.items(): + out += f"{module_name} {config}" + return out diff --git a/diffusers/quantizers/quantization_config.py b/diffusers/quantizers/quantization_config.py new file mode 100644 index 0000000000000000000000000000000000000000..ea78b5f7ff53014dc16be90541f06ac520d18ba1 --- /dev/null +++ b/diffusers/quantizers/quantization_config.py @@ -0,0 +1,989 @@ +#!/usr/bin/env python +# coding=utf-8 + +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Adapted from +https://github.com/huggingface/transformers/blob/52cb4034ada381fe1ffe8d428a1076e5411a8026/src/transformers/utils/quantization_config.py +""" + +from __future__ import annotations + +import copy +import importlib.metadata +import json +import os +import warnings +from dataclasses import dataclass +from enum import Enum +from typing import Any, Callable + +from packaging import version + +from ..utils import deprecate, is_torch_available, is_torchao_version, logging + + +if is_torch_available(): + import torch + +logger = logging.get_logger(__name__) + + +class QuantizationMethod(str, Enum): + BITS_AND_BYTES = "bitsandbytes" + GGUF = "gguf" + NUNCHAKU_LITE = "nunchaku_lite" + TORCHAO = "torchao" + QUANTO = "quanto" + MODELOPT = "modelopt" + AUTOROUND = "auto-round" + SDNQ = "sdnq" + + +@dataclass +class QuantizationConfigMixin: + """ + Mixin class for quantization config + """ + + quant_method: QuantizationMethod + _exclude_attributes_at_init = [] + + @classmethod + def from_dict(cls, config_dict, return_unused_kwargs=False, **kwargs): + """ + Instantiates a [`QuantizationConfigMixin`] from a Python dictionary of parameters. + + Args: + config_dict (`dict[str, Any]`): + Dictionary that will be used to instantiate the configuration object. + return_unused_kwargs (`bool`, *optional*, defaults to `False`): + Whether or not to return a list of unused keyword arguments. Used for `from_pretrained` method in + `PreTrainedModel`. + kwargs (`dict[str, Any]`): + Additional parameters from which to initialize the configuration object. + + Returns: + [`QuantizationConfigMixin`]: The configuration object instantiated from those parameters. + """ + + config = cls(**config_dict) + + to_remove = [] + for key, value in kwargs.items(): + if hasattr(config, key): + setattr(config, key, value) + to_remove.append(key) + for key in to_remove: + kwargs.pop(key, None) + + if return_unused_kwargs: + return config, kwargs + else: + return config + + def to_json_file(self, json_file_path: str | os.PathLike): + """ + Save this instance to a JSON file. + + Args: + json_file_path (`str` or `os.PathLike`): + Path to the JSON file in which this configuration instance's parameters will be saved. + use_diff (`bool`, *optional*, defaults to `True`): + If set to `True`, only the difference between the config instance and the default + `QuantizationConfig()` is serialized to JSON file. + """ + with open(json_file_path, "w", encoding="utf-8") as writer: + config_dict = self.to_dict() + json_string = json.dumps(config_dict, indent=2, sort_keys=True) + "\n" + + writer.write(json_string) + + def to_dict(self) -> dict[str, Any]: + """ + Serializes this instance to a Python dictionary. Returns: + `dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance. + """ + return copy.deepcopy(self.__dict__) + + def __iter__(self): + """allows `dict(obj)` for situations where obj may be a dict or QuantizationConfigMixin""" + for attr, value in copy.deepcopy(self.__dict__).items(): + yield attr, value + + def __repr__(self): + return f"{self.__class__.__name__} {self.to_json_string()}" + + def to_json_string(self, use_diff: bool = True) -> str: + """ + Serializes this instance to a JSON string. + + Args: + use_diff (`bool`, *optional*, defaults to `True`): + If set to `True`, only the difference between the config instance and the default `PretrainedConfig()` + is serialized to JSON string. + + Returns: + `str`: String containing all the attributes that make up this configuration instance in JSON format. + """ + if use_diff is True: + config_dict = self.to_diff_dict() + else: + config_dict = self.to_dict() + return json.dumps(config_dict, indent=2, sort_keys=True) + "\n" + + def update(self, **kwargs): + """ + Updates attributes of this class instance with attributes from `kwargs` if they match existing attributes, + returning all the unused kwargs. + + Args: + kwargs (`dict[str, Any]`): + Dictionary of attributes to tentatively update this class. + + Returns: + `dict[str, Any]`: Dictionary containing all the key-value pairs that were not used to update the instance. + """ + to_remove = [] + for key, value in kwargs.items(): + if hasattr(self, key): + setattr(self, key, value) + to_remove.append(key) + + # Remove all the attributes that were updated, without modifying the input dict + unused_kwargs = {key: value for key, value in kwargs.items() if key not in to_remove} + return unused_kwargs + + +@dataclass +class BitsAndBytesConfig(QuantizationConfigMixin): + """ + This is a wrapper class about all possible attributes and features that you can play with a model that has been + loaded using `bitsandbytes`. + + This replaces `load_in_8bit` or `load_in_4bit` therefore both options are mutually exclusive. + + Currently only supports `LLM.int8()`, `FP4`, and `NF4` quantization. If more methods are added to `bitsandbytes`, + then more arguments will be added to this class. + + Args: + load_in_8bit (`bool`, *optional*, defaults to `False`): + This flag is used to enable 8-bit quantization with LLM.int8(). + load_in_4bit (`bool`, *optional*, defaults to `False`): + This flag is used to enable 4-bit quantization by replacing the Linear layers with FP4/NF4 layers from + `bitsandbytes`. + llm_int8_threshold (`float`, *optional*, defaults to 6.0): + This corresponds to the outlier threshold for outlier detection as described in `LLM.int8() : 8-bit Matrix + Multiplication for Transformers at Scale` paper: https://huggingface.co/papers/2208.07339 Any hidden states + value that is above this threshold will be considered an outlier and the operation on those values will be + done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], + but there are some exceptional systematic outliers that are very differently distributed for large models. + These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of + magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, + but a lower threshold might be needed for more unstable models (small models, fine-tuning). + llm_int8_skip_modules (`list[str]`, *optional*): + An explicit list of the modules that we do not want to convert in 8-bit. This is useful for models such as + Jukebox that has several heads in different places and not necessarily at the last position. For example + for `CausalLM` models, the last `lm_head` is typically kept in its original `dtype`. + llm_int8_enable_fp32_cpu_offload (`bool`, *optional*, defaults to `False`): + This flag is used for advanced use cases and users that are aware of this feature. If you want to split + your model in different parts and run some parts in int8 on GPU and some parts in fp32 on CPU, you can use + this flag. This is useful for offloading large models such as `google/flan-t5-xxl`. Note that the int8 + operations will not be run on CPU. + llm_int8_has_fp16_weight (`bool`, *optional*, defaults to `False`): + This flag runs LLM.int8() with 16-bit main weights. This is useful for fine-tuning as the weights do not + have to be converted back and forth for the backward pass. + bnb_4bit_compute_dtype (`torch.dtype` or str, *optional*, defaults to `torch.float32`): + This sets the computational type which might be different than the input type. For example, inputs might be + fp32, but computation can be set to bf16 for speedups. + bnb_4bit_quant_type (`str`, *optional*, defaults to `"fp4"`): + This sets the quantization data type in the bnb.nn.Linear4Bit layers. Options are FP4 and NF4 data types + which are specified by `fp4` or `nf4`. + bnb_4bit_use_double_quant (`bool`, *optional*, defaults to `False`): + This flag is used for nested quantization where the quantization constants from the first quantization are + quantized again. + bnb_4bit_quant_storage (`torch.dtype` or str, *optional*, defaults to `torch.uint8`): + This sets the storage type to pack the quanitzed 4-bit prarams. + kwargs (`dict[str, Any]`, *optional*): + Additional parameters from which to initialize the configuration object. + """ + + _exclude_attributes_at_init = ["_load_in_4bit", "_load_in_8bit", "quant_method"] + + def __init__( + self, + load_in_8bit=False, + load_in_4bit=False, + llm_int8_threshold=6.0, + llm_int8_skip_modules=None, + llm_int8_enable_fp32_cpu_offload=False, + llm_int8_has_fp16_weight=False, + bnb_4bit_compute_dtype=None, + bnb_4bit_quant_type="fp4", + bnb_4bit_use_double_quant=False, + bnb_4bit_quant_storage=None, + **kwargs, + ): + self.quant_method = QuantizationMethod.BITS_AND_BYTES + + if load_in_4bit and load_in_8bit: + raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time") + + self._load_in_8bit = load_in_8bit + self._load_in_4bit = load_in_4bit + self.llm_int8_threshold = llm_int8_threshold + self.llm_int8_skip_modules = llm_int8_skip_modules + self.llm_int8_enable_fp32_cpu_offload = llm_int8_enable_fp32_cpu_offload + self.llm_int8_has_fp16_weight = llm_int8_has_fp16_weight + self.bnb_4bit_quant_type = bnb_4bit_quant_type + self.bnb_4bit_use_double_quant = bnb_4bit_use_double_quant + + if bnb_4bit_compute_dtype is None: + self.bnb_4bit_compute_dtype = torch.float32 + elif isinstance(bnb_4bit_compute_dtype, str): + self.bnb_4bit_compute_dtype = getattr(torch, bnb_4bit_compute_dtype) + elif isinstance(bnb_4bit_compute_dtype, torch.dtype): + self.bnb_4bit_compute_dtype = bnb_4bit_compute_dtype + else: + raise ValueError("bnb_4bit_compute_dtype must be a string or a torch.dtype") + + if bnb_4bit_quant_storage is None: + self.bnb_4bit_quant_storage = torch.uint8 + elif isinstance(bnb_4bit_quant_storage, str): + if bnb_4bit_quant_storage not in [ + "float16", + "float32", + "int8", + "uint8", + "float64", + "bfloat16", + ]: + raise ValueError( + "`bnb_4bit_quant_storage` must be a valid string (one of 'float16', 'float32', 'int8', 'uint8', 'float64', 'bfloat16') " + ) + self.bnb_4bit_quant_storage = getattr(torch, bnb_4bit_quant_storage) + elif isinstance(bnb_4bit_quant_storage, torch.dtype): + self.bnb_4bit_quant_storage = bnb_4bit_quant_storage + else: + raise ValueError("bnb_4bit_quant_storage must be a string or a torch.dtype") + + if kwargs and not all(k in self._exclude_attributes_at_init for k in kwargs): + logger.warning(f"Unused kwargs: {list(kwargs.keys())}. These kwargs are not used in {self.__class__}.") + + self.post_init() + + @property + def load_in_4bit(self): + return self._load_in_4bit + + @load_in_4bit.setter + def load_in_4bit(self, value: bool): + if not isinstance(value, bool): + raise TypeError("load_in_4bit must be a boolean") + + if self.load_in_8bit and value: + raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time") + self._load_in_4bit = value + + @property + def load_in_8bit(self): + return self._load_in_8bit + + @load_in_8bit.setter + def load_in_8bit(self, value: bool): + if not isinstance(value, bool): + raise TypeError("load_in_8bit must be a boolean") + + if self.load_in_4bit and value: + raise ValueError("load_in_4bit and load_in_8bit are both True, but only one can be used at the same time") + self._load_in_8bit = value + + def post_init(self): + r""" + Safety checker that arguments are correct - also replaces some NoneType arguments with their default values. + """ + if not isinstance(self.load_in_4bit, bool): + raise TypeError("load_in_4bit must be a boolean") + + if not isinstance(self.load_in_8bit, bool): + raise TypeError("load_in_8bit must be a boolean") + + if not isinstance(self.llm_int8_threshold, float): + raise TypeError("llm_int8_threshold must be a float") + + if self.llm_int8_skip_modules is not None and not isinstance(self.llm_int8_skip_modules, list): + raise TypeError("llm_int8_skip_modules must be a list of strings") + if not isinstance(self.llm_int8_enable_fp32_cpu_offload, bool): + raise TypeError("llm_int8_enable_fp32_cpu_offload must be a boolean") + + if not isinstance(self.llm_int8_has_fp16_weight, bool): + raise TypeError("llm_int8_has_fp16_weight must be a boolean") + + if self.bnb_4bit_compute_dtype is not None and not isinstance(self.bnb_4bit_compute_dtype, torch.dtype): + raise TypeError("bnb_4bit_compute_dtype must be torch.dtype") + + if not isinstance(self.bnb_4bit_quant_type, str): + raise TypeError("bnb_4bit_quant_type must be a string") + + if not isinstance(self.bnb_4bit_use_double_quant, bool): + raise TypeError("bnb_4bit_use_double_quant must be a boolean") + + if self.load_in_4bit and not version.parse(importlib.metadata.version("bitsandbytes")) >= version.parse( + "0.39.0" + ): + raise ValueError( + "4 bit quantization requires bitsandbytes>=0.39.0 - please upgrade your bitsandbytes version" + ) + + def is_quantizable(self): + r""" + Returns `True` if the model is quantizable, `False` otherwise. + """ + return self.load_in_8bit or self.load_in_4bit + + def quantization_method(self): + r""" + This method returns the quantization method used for the model. If the model is not quantizable, it returns + `None`. + """ + if self.load_in_8bit: + return "llm_int8" + elif self.load_in_4bit and self.bnb_4bit_quant_type == "fp4": + return "fp4" + elif self.load_in_4bit and self.bnb_4bit_quant_type == "nf4": + return "nf4" + else: + return None + + def to_dict(self) -> dict[str, Any]: + """ + Serializes this instance to a Python dictionary. Returns: + `dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance. + """ + output = copy.deepcopy(self.__dict__) + output["bnb_4bit_compute_dtype"] = str(output["bnb_4bit_compute_dtype"]).split(".")[1] + output["bnb_4bit_quant_storage"] = str(output["bnb_4bit_quant_storage"]).split(".")[1] + output["load_in_4bit"] = self.load_in_4bit + output["load_in_8bit"] = self.load_in_8bit + + return output + + def __repr__(self): + config_dict = self.to_dict() + return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n" + + def to_diff_dict(self) -> dict[str, Any]: + """ + Removes all attributes from config which correspond to the default config attributes for better readability and + serializes to a Python dictionary. + + Returns: + `dict[str, Any]`: Dictionary of all the attributes that make up this configuration instance, + """ + config_dict = self.to_dict() + + # get the default config dict + default_config_dict = BitsAndBytesConfig().to_dict() + + serializable_config_dict = {} + + # only serialize values that differ from the default config + for key, value in config_dict.items(): + if value != default_config_dict[key]: + serializable_config_dict[key] = value + + return serializable_config_dict + + +@dataclass +class GGUFQuantizationConfig(QuantizationConfigMixin): + """This is a config class for GGUF Quantization techniques. + + Args: + compute_dtype: (`torch.dtype`, defaults to `torch.float32`): + This sets the computational type which might be different than the input type. For example, inputs might be + fp32, but computation can be set to bf16 for speedups. + + """ + + def __init__(self, compute_dtype: "torch.dtype" | None = None): + self.quant_method = QuantizationMethod.GGUF + self.compute_dtype = compute_dtype + self.pre_quantized = True + + # TODO: (Dhruv) Add this as an init argument when we can support loading unquantized checkpoints. + self.modules_to_not_convert = None + + if self.compute_dtype is None: + self.compute_dtype = torch.float32 + + +@dataclass +class NunchakuLiteQuantizationConfig(QuantizationConfigMixin): + """Configuration for loading Nunchaku Lite checkpoints. + + Nunchaku Lite support in Diffusers loads prequantized checkpoints. To create a compatible checkpoint, use + [`diffuse-compressor`](https://github.com/rootonchair/diffuse-compressor) to choose or adapt a target configuration + for the model architecture, quantize and export the transformer, and package it as a Diffusers pipeline with the + compact `quantization_config` stored in `config.json`. + + The exported state dict must match the target Diffusers model architecture exactly. Checkpoints quantized with + fused QKV projections won't load into a model config that expects separate Q, K, and V projection modules. + + Example compact `config.json` config: + + ```json + { + "_class_name": "ErnieImageTransformer2DModel", + "quantization_config": { + "quant_method": "nunchaku_lite", + "compute_dtype": "bfloat16", + "svdq_w4a4": { + "precision": "nvfp4", + "group_size": 16, + "rank": 32, + "targets": ["layers.0.self_attention.to_q"] + }, + "awq_w4a16": { + "precision": "int4", + "group_size": 64, + "targets": ["final_linear"] + } + } + } + ``` + + Args: + compute_dtype (`torch.dtype`, defaults to `torch.bfloat16`): + Runtime dtype used by the floating-point buffers in the quantized modules. + svdq_w4a4 (`dict`, *optional*): + Explicit SVDQ W4A4 target configuration with `precision`, `group_size`, `rank`, and `targets`. + awq_w4a16 (`dict`, *optional*): + Explicit AWQ W4A16 target configuration with `precision`, `group_size`, and `targets`. + """ + + def __init__( + self, + compute_dtype: "torch.dtype" | str | None = None, + svdq_w4a4: dict[str, Any] | None = None, + awq_w4a16: dict[str, Any] | None = None, + **kwargs, + ): + self.quant_method = QuantizationMethod.NUNCHAKU_LITE + if compute_dtype is None: + compute_dtype = torch.bfloat16 + if isinstance(compute_dtype, str): + if not hasattr(torch, compute_dtype): + raise ValueError(f"Unsupported Nunchaku compute dtype: {compute_dtype!r}.") + compute_dtype = getattr(torch, compute_dtype) + if not isinstance(compute_dtype, torch.dtype): + raise ValueError("Nunchaku compute_dtype must be a string or a torch.dtype.") + self.compute_dtype = compute_dtype + self.pre_quantized = True + self.svdq_w4a4 = svdq_w4a4 + self.awq_w4a16 = awq_w4a16 + + self.post_init() + + def post_init(self): + if self.svdq_w4a4 is None and self.awq_w4a16 is None: + raise ValueError( + "Nunchaku compact quantization config must include `svdq_w4a4.targets` or `awq_w4a16.targets`." + ) + + for op, raw in (("svdq_w4a4", self.svdq_w4a4), ("awq_w4a16", self.awq_w4a16)): + if raw is None: + continue + if not isinstance(raw, dict): + raise ValueError(f"Nunchaku compact config section {op!r} must be a JSON object.") + + for key, expected_type in (("precision", str), ("group_size", int), ("targets", list)): + if key not in raw: + raise ValueError(f"Nunchaku compact config section {op!r} is missing required field {key!r}.") + if not isinstance(raw[key], expected_type): + raise ValueError( + f"Nunchaku compact config section {op!r} field {key!r} must be {expected_type.__name__}." + ) + + precision = raw["precision"] + group_size = raw["group_size"] + targets = raw["targets"] + if precision not in ("int4", "nvfp4"): + raise ValueError(f"Unsupported Nunchaku precision {precision!r} for {op!r}.") + if group_size <= 0: + raise ValueError(f"Nunchaku compact config section {op!r} must have positive group_size.") + if not targets: + raise ValueError(f"Nunchaku compact config section {op!r} must contain at least one target.") + if not all(isinstance(target, str) for target in targets): + raise ValueError(f"Nunchaku compact config section {op!r} targets must be strings.") + + if op == "svdq_w4a4": + if "rank" not in raw: + raise ValueError(f"Nunchaku compact config section {op!r} is missing required field 'rank'.") + if not isinstance(raw["rank"], int): + raise ValueError(f"Nunchaku compact config section {op!r} field 'rank' must be int.") + if raw["rank"] < 0: + raise ValueError(f"Nunchaku compact config section {op!r} must have non-negative rank.") + expected_group_size = 16 if precision == "nvfp4" else 64 + if group_size != expected_group_size: + raise ValueError( + f"Nunchaku SVDQ config with precision={precision!r} requires " + f"group_size={expected_group_size}, got {group_size}." + ) + elif precision != "int4": + raise ValueError("Nunchaku AWQ target requires precision='int4'.") + + def to_dict(self) -> dict[str, Any]: + output = super().to_dict() + output["compute_dtype"] = str(output["compute_dtype"]).split(".")[1] + return output + + +@dataclass +class TorchAoConfig(QuantizationConfigMixin): + """This is a config class for torchao quantization/sparsity techniques. + + Args: + quant_type (`AOBaseConfig`): + An `AOBaseConfig` subclass instance specifying the quantization type. See the [torchao + documentation](https://docs.pytorch.org/ao/main/api_ref_quantization.html#inference-apis-for-quantize) for + available config classes (e.g. `Int4WeightOnlyConfig`, `Int8WeightOnlyConfig`, `Float8WeightOnlyConfig`, + `Float8DynamicActivationFloat8WeightConfig`, etc.). + modules_to_not_convert (`list[str]`, *optional*, default to `None`): + The list of modules to not quantize, useful for quantizing models that explicitly require to have some + modules left in their original precision. + + Example: + ```python + from diffusers import FluxTransformer2DModel, TorchAoConfig + from torchao.quantization import Int8WeightOnlyConfig + + quantization_config = TorchAoConfig(Int8WeightOnlyConfig()) + transformer = FluxTransformer2DModel.from_pretrained( + "black-forest-labs/Flux.1-Dev", + subfolder="transformer", + quantization_config=quantization_config, + torch_dtype=torch.bfloat16, + ) + ``` + """ + + def __init__( + self, + quant_type: "AOBaseConfig", # noqa: F821 + modules_to_not_convert: list[str] | None = None, + **kwargs, + ) -> None: + self.quant_method = QuantizationMethod.TORCHAO + self.quant_type = quant_type + self.modules_to_not_convert = modules_to_not_convert + + self.post_init() + + def post_init(self): + if is_torchao_version("<", "0.15.0"): + raise ValueError("TorchAoConfig requires torchao >= 0.15.0. Please upgrade with `pip install -U torchao`.") + + from torchao.quantization.quant_api import AOBaseConfig + + if not isinstance(self.quant_type, AOBaseConfig): + raise TypeError(f"quant_type must be an AOBaseConfig instance, got {type(self.quant_type).__name__}") + + def to_dict(self): + """Convert configuration to a dictionary.""" + d = super().to_dict() + + # Handle AOBaseConfig serialization + from torchao.core.config import config_to_dict + + # For now we assume there is 1 config per Transformer, however in the future + # we may want to support a config per fqn. + # See: https://docs.pytorch.org/ao/stable/api_reference/generated/torchao.quantization.quantize_.html + d["quant_type"] = {"default": config_to_dict(self.quant_type)} + + return d + + @classmethod + def from_dict(cls, config_dict, return_unused_kwargs=False, **kwargs): + """Create configuration from a dictionary.""" + if not is_torchao_version(">=", "0.15.0"): + raise NotImplementedError("TorchAoConfig requires torchao >= 0.15.0 for construction from dict") + config_dict = config_dict.copy() + quant_type = config_dict.pop("quant_type") + + # Check if we only have one key which is "default" + # In the future we may update this + assert len(quant_type) == 1 and "default" in quant_type, ( + "Expected only one key 'default' in quant_type dictionary" + ) + quant_type = quant_type["default"] + + # Deserialize quant_type if needed + from torchao.core.config import config_from_dict + + quant_type = config_from_dict(quant_type) + + return cls(quant_type=quant_type, **config_dict) + + def get_apply_tensor_subclass(self): + """Create the appropriate quantization method based on configuration.""" + return self.quant_type + + def __repr__(self): + config_dict = self.to_dict() + return f"{self.__class__.__name__} {json.dumps(config_dict, indent=2, sort_keys=True)}\n" + + +@dataclass +class QuantoConfig(QuantizationConfigMixin): + """ + This is a wrapper class about all possible attributes and features that you can play with a model that has been + loaded using `quanto`. + + Args: + weights_dtype (`str`, *optional*, defaults to `"int8"`): + The target dtype for the weights after quantization. Supported values are ("float8","int8","int4","int2") + modules_to_not_convert (`list`, *optional*, default to `None`): + The list of modules to not quantize, useful for quantizing models that explicitly require to have some + modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers). + """ + + def __init__( + self, + weights_dtype: str = "int8", + modules_to_not_convert: list[str] | None = None, + **kwargs, + ): + deprecation_message = "`QuantoConfig` is deprecated and will be removed in version 1.0.0." + deprecate("QuantoConfig", "1.0.0", deprecation_message) + self.quant_method = QuantizationMethod.QUANTO + self.weights_dtype = weights_dtype + self.modules_to_not_convert = modules_to_not_convert + + self.post_init() + + def post_init(self): + r""" + Safety checker that arguments are correct + """ + accepted_weights = ["float8", "int8", "int4", "int2"] + if self.weights_dtype not in accepted_weights: + raise ValueError(f"Only support weights in {accepted_weights} but found {self.weights_dtype}") + + +@dataclass +class NVIDIAModelOptConfig(QuantizationConfigMixin): + """This is a config class to use nvidia modelopt for quantization. + + Args: + quant_type (`str`): + The type of quantization we want to use, following is how to use: + **weightquant_activationquant ==> FP8_FP8** In the above example we have use FP8 for both weight and + activation quantization. Following are the all the options: + - FP8 + - INT8 + - INT4 + - NF4 + - NVFP4 + modules_to_not_convert (`list[str]`, *optional*, default to `None`): + The list of modules to not quantize, useful for quantizing models that explicitly require to have some + weight_only (`bool`, *optional*, default to `False`): + If set to `True`, the quantization will be applied only to the weights of the model. + channel_quantize (`int`, *optional*, default to `None`): + The channel quantization axis, useful for quantizing models across different axes. + block_quantize (`int`, *optional*, default to `None`): + The block size, useful to further quantize each channel/axes into blocks. + scale_channel_quantize (`int`, *optional*, default to `None`): + The scale channel quantization axis, useful for quantizing calculated scale across different axes. + scale_block_quantize (`int`, *optional*, default to `None`): + The scale block size, useful for quantizing each scale channel/axes into blocks. + algorithm (`str`, *optional*, default to `"max"`): + The algorithm to use for quantization, currently only supports `"max"`. + forward_loop (`Callable`, *optional*, default to `None`): + The forward loop function to use for calibration during quantization. + modelopt_config (`dict`, *optional*, default to `None`): + The modelopt config, useful for passing custom configs to modelopt. + disable_conv_quantization (`bool`, *optional*, default to `False`): + If set to `True`, the quantization will be disabled for convolutional layers. + kwargs (`dict[str, Any]`, *optional*): + Additional parameters which are to be used for calibration. + """ + + quanttype_to_numbits = { + "FP8": (4, 3), + "INT8": 8, + "INT4": 4, + "NF4": 4, + "NVFP4": (2, 1), + } + quanttype_to_scalingbits = { + "NF4": 8, + "NVFP4": (4, 3), + } + + def __init__( + self, + quant_type: str, + modules_to_not_convert: list[str] | None = None, + weight_only: bool = True, + channel_quantize: int | None = None, + block_quantize: int | None = None, + scale_channel_quantize: int | None = None, + scale_block_quantize: int | None = None, + algorithm: str = "max", + forward_loop: Callable | None = None, + modelopt_config: dict | None = None, + disable_conv_quantization: bool = False, + **kwargs, + ) -> None: + self.quant_method = QuantizationMethod.MODELOPT + self._normalize_quant_type(quant_type) + self.modules_to_not_convert = modules_to_not_convert + self.weight_only = weight_only + self.channel_quantize = channel_quantize + self.block_quantize = block_quantize + self.calib_cfg = { + "method": algorithm, + # add more options here if needed + } + self.forward_loop = forward_loop + self.scale_channel_quantize = scale_channel_quantize + self.scale_block_quantize = scale_block_quantize + self.modelopt_config = self.get_config_from_quant_type() if not modelopt_config else modelopt_config + self.disable_conv_quantization = disable_conv_quantization + + def check_model_patching(self, operation: str = "loading"): + # ModelOpt imports diffusers internally. This is here to prevent circular imports + from modelopt.torch.opt.plugins.huggingface import _PATCHED_CLASSES + + if len(_PATCHED_CLASSES) == 0: + warning_msg = ( + f"Not {operation} weights in modelopt format. This might cause unreliable behavior." + "Please make sure to run the following code before loading/saving model weights:\n\n" + " from modelopt.torch.opt import enable_huggingface_checkpointing\n" + " enable_huggingface_checkpointing()\n" + ) + warnings.warn(warning_msg) + + def _normalize_quant_type(self, quant_type: str) -> str: + """ + Validates and normalizes the quantization type string. + + Splits the quant_type into weight and activation components, verifies them against supported types, and + replaces unsupported values with safe defaults. + + Args: + quant_type (str): The input quantization type string (e.g., 'FP8_INT8'). + + Returns: + str: A valid quantization type string (e.g., 'FP8_INT8' or 'FP8'). + """ + parts = quant_type.split("_") + w_type = parts[0] + act_type = parts[1] if len(parts) > 1 else None + if len(parts) > 2: + logger.warning(f"Quantization type {quant_type} is not supported. Picking FP8_INT8 as default") + w_type = "FP8" + act_type = None + else: + if w_type not in NVIDIAModelOptConfig.quanttype_to_numbits: + logger.warning(f"Weight Quantization type {w_type} is not supported. Picking FP8 as default") + w_type = "FP8" + if act_type is not None and act_type not in NVIDIAModelOptConfig.quanttype_to_numbits: + logger.warning(f"Activation Quantization type {act_type} is not supported. Picking INT8 as default") + act_type = None + self.quant_type = w_type + ("_" + act_type if act_type is not None else "") + + def get_config_from_quant_type(self) -> dict[str, Any]: + """ + Get the config from the quantization type. + """ + import modelopt.torch.quantization as mtq + + BASE_CONFIG = { + "quant_cfg": { + "*weight_quantizer": {"fake_quant": False}, + "*input_quantizer": {}, + "*output_quantizer": {"enable": False}, + "*q_bmm_quantizer": {}, + "*k_bmm_quantizer": {}, + "*v_bmm_quantizer": {}, + "*softmax_quantizer": {}, + **mtq.config._default_disabled_quantizer_cfg, + }, + "algorithm": self.calib_cfg, + } + + quant_cfg = BASE_CONFIG["quant_cfg"] + if self.weight_only: + for k in quant_cfg: + if "*weight_quantizer" not in k and not quant_cfg[k]: + quant_cfg[k]["enable"] = False + + parts = self.quant_type.split("_") + w_type = parts[0] + act_type = parts[1].replace("A", "") if len(parts) > 1 else None + for k in quant_cfg: + if k not in mtq.config._default_disabled_quantizer_cfg and "enable" not in quant_cfg[k]: + if k == "*input_quantizer": + if act_type is not None: + quant_cfg[k]["num_bits"] = NVIDIAModelOptConfig.quanttype_to_numbits[act_type] + continue + quant_cfg[k]["num_bits"] = NVIDIAModelOptConfig.quanttype_to_numbits[w_type] + + if self.block_quantize is not None and self.channel_quantize is not None: + quant_cfg["*weight_quantizer"]["block_sizes"] = {self.channel_quantize: self.block_quantize} + quant_cfg["*input_quantizer"]["block_sizes"] = { + self.channel_quantize: self.block_quantize, + "type": "dynamic", + } + elif self.channel_quantize is not None: + quant_cfg["*weight_quantizer"]["axis"] = self.channel_quantize + quant_cfg["*input_quantizer"]["axis"] = self.channel_quantize + quant_cfg["*input_quantizer"]["type"] = "dynamic" + + # Only fixed scaling sizes are supported for now in modelopt + if self.scale_channel_quantize is not None and self.scale_block_quantize is not None: + if w_type in NVIDIAModelOptConfig.quanttype_to_scalingbits: + quant_cfg["*weight_quantizer"]["block_sizes"].update( + { + "scale_bits": NVIDIAModelOptConfig.quanttype_to_scalingbits[w_type], + "scale_block_sizes": {self.scale_channel_quantize: self.scale_block_quantize}, + } + ) + if act_type and act_type in NVIDIAModelOptConfig.quanttype_to_scalingbits: + quant_cfg["*input_quantizer"]["block_sizes"].update( + { + "scale_bits": NVIDIAModelOptConfig.quanttype_to_scalingbits[act_type], + "scale_block_sizes": {self.scale_channel_quantize: self.scale_block_quantize}, + } + ) + + return BASE_CONFIG + + +@dataclass +class AutoRoundConfig(QuantizationConfigMixin): + """Configuration class for AutoRound quantization. + + AutoRound is a weight-only quantization algorithm that uses sign gradient descent to jointly optimize weight + rounding and min-max values. This config targets the W4A16 (4-bit weights, 16-bit activations) setting. + + Reference: https://github.com/intel/auto-round + + Args: + bits (`int`, *optional*, defaults to `4`): + The number of bits to quantize weights to. For W4A16 this should be 4. + group_size (`int`, *optional*, defaults to `128`): + The group size for weight quantization. Weights in each group share the same scale and zero-point. Common + choices: 32, 64, 128, -1 (per-channel). + sym (`bool`, *optional*, defaults to `True`): + Whether to use symmetric quantization (zero-point fixed at 0) or asymmetric quantization (zero-point is + learned). + backend (`str`, *optional*, defaults to `"auto"`): + The backend kernel to use for quantized inference. Available backends: + - `"auto"`: Automatically select the best available backend for the current device. + - `"torch"`: Pure PyTorch kernel — works on CPU and CUDA. + - `"tritonv2"`: Triton-based kernel — requires CUDA. + - `"exllamav2"`: Exllamav2 kernel via GPTQModel — requires CUDA and `gptqmodel>=5.8.0`. Offers good CUDA + inference performance. + - `"marlin"`: Marlin kernel via GPTQModel — requires CUDA and `gptqmodel>=5.8.0`. Offers the best CUDA + inference performance. + kwargs (`dict[str, Any]`, *optional*): + Additional keyword arguments forwarded to AutoRound (e.g. `iters`, `seqlen`, `batch_size`, `lr`, + `minmax_lr` for calibration when quantizing from scratch). + """ + + VALID_BACKENDS = ["auto", "torch", "tritonv2", "exllamav2", "marlin"] + + def __init__( + self, + bits: int = 4, + group_size: int = 128, + sym: bool = True, + backend: str = "auto", + **kwargs, + ) -> None: + self.quant_method = QuantizationMethod.AUTOROUND + self._validate_backend(backend) + self.bits = bits + self.group_size = group_size + self.sym = sym + self.backend = backend + for k, v in kwargs.items(): + setattr(self, k, v) + + def _validate_backend(self, backend): + if backend not in self.VALID_BACKENDS: + raise ValueError(f"Invalid backend '{backend}'. Valid options are: {self.VALID_BACKENDS}") + + def to_dict(self) -> dict: + """Serialize the config to a JSON-compatible dict. + + Output: A dict containing all config fields. The `quant_method` is stored as its string value so it can be + round-tripped through JSON. + """ + output = super().to_dict() + output["quant_method"] = output["quant_method"].value + return output + + @classmethod + def from_dict(cls, config_dict: dict, return_unused_kwargs: bool = False, **kwargs): + """Instantiate an AutoRoundConfig from a dictionary. + + Input: config_dict with keys like bits, group_size, sym, etc. Output: An AutoRoundConfig instance (and + optionally unused kwargs). + """ + # Filter out keys that are not constructor parameters + # (e.g. quant_method is set automatically) + config_dict = {k: v for k, v in config_dict.items() if k != "quant_method"} + return super().from_dict(config_dict, return_unused_kwargs=return_unused_kwargs, **kwargs) + + +class SDNQConfig(QuantizationConfigMixin): + """Configuration class for SDNQ (SD.Next Quantization). + + The `sdnq` library ships its own diffusers-compatible config and quantizer; this class is a thin factory that + defers to them so that `quant_method="sdnq"` checkpoints load natively with diffusers. All arguments are forwarded + to `sdnq.SDNQConfig`. Requires the `sdnq` library: `pip install sdnq`. + + Reference: https://github.com/Disty0/sdnq + + Args: + weights_dtype (`str`, *optional*, defaults to `"int8"`): + The target dtype for the weights after quantization, e.g. `"int8"`, `"uint4"`, `"float8_e4m3fn"`. See + `sdnq.common.accepted_weight_dtypes` for all supported values. + group_size (`int`, *optional*, defaults to `0`): + How many elements of a tensor share the same quantization group. `0` auto-selects based on `weights_dtype`, + `-1` disables grouping and uses row-wise quantization. + use_svd (`bool`, *optional*, defaults to `False`): + Whether to apply the SVDQuant algorithm on top of SDNQ quantization. + use_quantized_matmul (`bool`, *optional*, defaults to `False`): + Whether to use quantized INT8 / FP8 / FP16 matmul on the forward pass instead of BF16 / FP16. + modules_to_not_convert (`list`, *optional*, defaults to `None`): + The list of modules to skip during quantization. + kwargs (`dict[str, Any]`, *optional*): + Additional keyword arguments forwarded to `sdnq.SDNQConfig` (e.g. `quantized_matmul_dtype`, `svd_rank`, + `use_hadamard`, `quant_conv`, `quant_embedding`, `modules_dtype_dict`). + """ + + def __new__(cls, *args, **kwargs): + from .sdnq.sdnq_quantizer import _check_sdnq_requirement + + _check_sdnq_requirement() + from sdnq import SDNQConfig as SDNQLibConfig + + return SDNQLibConfig(*args, **kwargs) diff --git a/diffusers/quantizers/quanto/__init__.py b/diffusers/quantizers/quanto/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a4e8a1f41a1e4bf3bbb53c5684398277eb671a83 --- /dev/null +++ b/diffusers/quantizers/quanto/__init__.py @@ -0,0 +1 @@ +from .quanto_quantizer import QuantoQuantizer diff --git a/diffusers/quantizers/quanto/quanto_quantizer.py b/diffusers/quantizers/quanto/quanto_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..9a04291c883a206b983626662a2a6484e0f967b9 --- /dev/null +++ b/diffusers/quantizers/quanto/quanto_quantizer.py @@ -0,0 +1,185 @@ +from typing import TYPE_CHECKING, Any + +from diffusers.utils.import_utils import is_optimum_quanto_version + +from ...utils import ( + deprecate, + get_module_from_name, + is_accelerate_available, + is_accelerate_version, + is_optimum_quanto_available, + is_torch_available, + logging, +) +from ..base import DiffusersQuantizer + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + + +if is_torch_available(): + import torch + +if is_accelerate_available(): + from accelerate.utils import CustomDtype, set_module_tensor_to_device + +if is_optimum_quanto_available(): + from .utils import _replace_with_quanto_layers + +logger = logging.get_logger(__name__) + + +class QuantoQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for Optimum Quanto + """ + + use_keep_in_fp32_modules = True + requires_calibration = False + required_packages = ["quanto", "accelerate"] + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + def validate_environment(self, *args, **kwargs): + deprecation_message = "The Quanto quantizer is deprecated and will be removed in version 1.0.0." + deprecate("QuantoQuantizer", "1.0.0", deprecation_message) + + if not is_optimum_quanto_available(): + raise ImportError( + "Loading an optimum-quanto quantized model requires optimum-quanto library (`pip install optimum-quanto`)" + ) + if not is_optimum_quanto_version(">=", "0.2.6"): + raise ImportError( + "Loading an optimum-quanto quantized model requires `optimum-quanto>=0.2.6`. " + "Please upgrade your installation with `pip install --upgrade optimum-quanto" + ) + + if not is_accelerate_available(): + raise ImportError( + "Loading an optimum-quanto quantized model requires accelerate library (`pip install accelerate`)" + ) + + device_map = kwargs.get("device_map", None) + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + raise ValueError( + "`device_map` for multi-GPU inference or CPU/disk offload is currently not supported with Diffusers and the Quanto backend" + ) + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ): + # Quanto imports diffusers internally. This is here to prevent circular imports + from optimum.quanto import QModuleMixin, QTensor + from optimum.quanto.tensor.packed import PackedTensor + + module, tensor_name = get_module_from_name(model, param_name) + if self.pre_quantized and any(isinstance(module, t) for t in [QTensor, PackedTensor]): + return True + elif isinstance(module, QModuleMixin) and "weight" in tensor_name: + return not module.frozen + + return False + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + *args, + **kwargs, + ): + """ + Create the quantized parameter by calling .freeze() after setting it to the module. + """ + + dtype = kwargs.get("dtype", torch.float32) + module, tensor_name = get_module_from_name(model, param_name) + if self.pre_quantized: + setattr(module, tensor_name, param_value) + else: + set_module_tensor_to_device(model, param_name, target_device, param_value, dtype) + module.freeze() + module.weight.requires_grad = False + + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + max_memory = {key: val * 0.90 for key, val in max_memory.items()} + return max_memory + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + if is_accelerate_version(">=", "0.27.0"): + mapping = { + "int8": torch.int8, + "float8": CustomDtype.FP8, + "int4": CustomDtype.INT4, + "int2": CustomDtype.INT2, + } + target_dtype = mapping[self.quantization_config.weights_dtype] + + return target_dtype + + def update_torch_dtype(self, torch_dtype: "torch.dtype" = None) -> "torch.dtype": + if torch_dtype is None: + logger.info("You did not specify `torch_dtype` in `from_pretrained`. Setting it to `torch.float32`.") + torch_dtype = torch.float32 + return torch_dtype + + def update_missing_keys(self, model, missing_keys: list[str], prefix: str) -> list[str]: + # Quanto imports diffusers internally. This is here to prevent circular imports + from optimum.quanto import QModuleMixin + + not_missing_keys = [] + for name, module in model.named_modules(): + if isinstance(module, QModuleMixin): + for missing in missing_keys: + if ( + (name in missing or name in f"{prefix}.{missing}") + and not missing.endswith(".weight") + and not missing.endswith(".bias") + ): + not_missing_keys.append(missing) + return [k for k in missing_keys if k not in not_missing_keys] + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + self.modules_to_not_convert = self.quantization_config.modules_to_not_convert + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + model = _replace_with_quanto_layers( + model, + modules_to_not_convert=self.modules_to_not_convert, + quantization_config=self.quantization_config, + pre_quantized=self.pre_quantized, + ) + model.config.quantization_config = self.quantization_config + + def _process_model_after_weight_loading(self, model, **kwargs): + return model + + @property + def is_trainable(self): + return True + + @property + def is_serializable(self): + return True + + @property + def is_compileable(self) -> bool: + return True diff --git a/diffusers/quantizers/quanto/utils.py b/diffusers/quantizers/quanto/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..6f41fd36b43a7c222de17a297523939d27f34537 --- /dev/null +++ b/diffusers/quantizers/quanto/utils.py @@ -0,0 +1,60 @@ +import torch.nn as nn + +from ...utils import is_accelerate_available, logging + + +logger = logging.get_logger(__name__) + +if is_accelerate_available(): + from accelerate import init_empty_weights + + +def _replace_with_quanto_layers(model, quantization_config, modules_to_not_convert: list, pre_quantized=False): + # Quanto imports diffusers internally. These are placed here to avoid circular imports + from optimum.quanto import QLinear, freeze, qfloat8, qint2, qint4, qint8 + + def _get_weight_type(dtype: str): + return {"float8": qfloat8, "int8": qint8, "int4": qint4, "int2": qint2}[dtype] + + def _replace_layers(model, quantization_config, modules_to_not_convert): + has_children = list(model.children()) + if not has_children: + return model + + for name, module in model.named_children(): + _replace_layers(module, quantization_config, modules_to_not_convert) + + if name in modules_to_not_convert: + continue + + if isinstance(module, nn.Linear): + with init_empty_weights(): + qlinear = QLinear( + in_features=module.in_features, + out_features=module.out_features, + bias=module.bias is not None, + dtype=module.weight.dtype, + weights=_get_weight_type(quantization_config.weights_dtype), + ) + model._modules[name] = qlinear + model._modules[name].source_cls = type(module) + model._modules[name].requires_grad_(False) + + return model + + model = _replace_layers(model, quantization_config, modules_to_not_convert) + has_been_replaced = any(isinstance(replaced_module, QLinear) for _, replaced_module in model.named_modules()) + + if not has_been_replaced: + logger.warning( + f"{model.__class__.__name__} does not appear to have any `nn.Linear` modules. Quantization will not be applied." + " Please check your model architecture, or submit an issue on Github if you think this is a bug." + " https://github.com/huggingface/diffusers/issues/new" + ) + + # We need to freeze the pre_quantized model in order for the loaded state_dict and model state dict + # to match when trying to load weights with load_model_dict_into_meta + if pre_quantized: + freeze(model) + + return model diff --git a/diffusers/quantizers/sdnq/__init__.py b/diffusers/quantizers/sdnq/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..a55441900e3bbeef8fc1f014d69fb5e68a314033 --- /dev/null +++ b/diffusers/quantizers/sdnq/__init__.py @@ -0,0 +1 @@ +from .sdnq_quantizer import SDNQQuantizer diff --git a/diffusers/quantizers/sdnq/sdnq_quantizer.py b/diffusers/quantizers/sdnq/sdnq_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..7d040e899399e0158ebc807ebc74f524bf4075ea --- /dev/null +++ b/diffusers/quantizers/sdnq/sdnq_quantizer.py @@ -0,0 +1,72 @@ +# Copyright 2026 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import os +from functools import lru_cache + +from ...utils import is_sdnq_available, is_sdnq_version, logging +from ..base import DiffusersQuantizer + + +logger = logging.get_logger(__name__) + +# Recommended sdnq version. Older versions load/save most SDNQ checkpoints fine; only some of the +# latest ones require this version, so a mismatch is a warning, not a hard error. +_RECOMMENDED_SDNQ_VERSION = "0.2.2" + + +def _check_sdnq_requirement(): + if not is_sdnq_available(): + raise ImportError( + f"Loading or creating an SDNQ quantized model requires the sdnq library: " + f"`pip install 'sdnq>={_RECOMMENDED_SDNQ_VERSION}'`" + ) + + if is_sdnq_version("<", _RECOMMENDED_SDNQ_VERSION) and not os.getenv("DIFFUSERS_NO_ADVISORY_WARNINGS"): + logger.warning( + f"Your installed sdnq is older than {_RECOMMENDED_SDNQ_VERSION}. Most SDNQ checkpoints will still " + f"load and save fine, but some of the latest ones require sdnq>={_RECOMMENDED_SDNQ_VERSION}. " + f"If you hit issues, upgrade with `pip install -U sdnq`." + ) + + +class SDNQQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for SDNQ (https://github.com/Disty0/sdnq). + + The `sdnq` library ships its own `DiffusersQuantizer` subclass; this class is a thin factory that defers to it. It + only exists so that `quant_method="sdnq"` checkpoints load natively through `DiffusersAutoQuantizer` without + requiring `import sdnq` beforehand. `sdnq` cannot be imported at module level because it imports + `diffusers.quantizers.auto` at import time. + """ + + def __new__(cls, quantization_config, **kwargs): + _check_sdnq_requirement() + from sdnq import SDNQQuantizer as SDNQLibQuantizer + + return SDNQLibQuantizer(quantization_config, **kwargs) + + +@lru_cache +def _ensure_sdnq_registered(): + """ + Import sdnq once so it registers itself with transformers (import side effect). transformers loads text encoders + through its own quantizer registry and only learns about SDNQ once sdnq has been imported; without this it silently + skips the quantization config of a prequantized SDNQ component and mis-loads the weights. + """ + logger.info( + "DIFFUSERS_SDNQ_TRANSFORMERS is set; importing sdnq so it registers with transformers " + "for loading SDNQ-quantized transformers models (e.g. text encoders or LLMs)." + ) + import sdnq # noqa: F401 diff --git a/diffusers/quantizers/torchao/__init__.py b/diffusers/quantizers/torchao/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c56bf54c25158fb9e7ef9ff3bc1eb651743764fb --- /dev/null +++ b/diffusers/quantizers/torchao/__init__.py @@ -0,0 +1,15 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from .torchao_quantizer import TorchAoHfQuantizer diff --git a/diffusers/quantizers/torchao/torchao_quantizer.py b/diffusers/quantizers/torchao/torchao_quantizer.py new file mode 100644 index 0000000000000000000000000000000000000000..503e452c20b12b9936a0a0f83be2fe511e16b64e --- /dev/null +++ b/diffusers/quantizers/torchao/torchao_quantizer.py @@ -0,0 +1,484 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Adapted from +https://github.com/huggingface/transformers/blob/3a8eb74668e9c2cc563b2f5c62fac174797063e0/src/transformers/quantizers/quantizer_torchao.py +""" + +import importlib +import json +import re +import types +from typing import TYPE_CHECKING, Any + +from packaging import version + +from ...utils import ( + get_module_from_name, + is_safetensors_available, + is_torch_available, + is_torch_version, + is_torchao_available, + is_torchao_version, + logging, +) +from ..base import DiffusersQuantizer + + +logger = logging.get_logger(__name__) + + +if TYPE_CHECKING: + from ...models.modeling_utils import ModelMixin + +if is_safetensors_available(): + from safetensors import safe_open + + +if is_torch_available(): + import torch + import torch.nn as nn + + if is_torch_version(">=", "2.5"): + SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION = ( + # At the moment, only int8 is supported for integer quantization dtypes. + # In Torch 2.6, int1-int7 will be introduced, so this can be visited in the future + # to support more quantization methods, such as intx_weight_only. + torch.int8, + torch.float8_e4m3fn, + torch.float8_e5m2, + torch.uint1, + torch.uint2, + torch.uint3, + torch.uint4, + torch.uint5, + torch.uint6, + torch.uint7, + ) + else: + SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION = ( + torch.int8, + torch.float8_e4m3fn, + torch.float8_e5m2, + ) + +if is_torchao_available(): + from torchao.quantization import quantize_ + + if is_torchao_version(">=", "0.16.0"): + from torchao.prototype.safetensors.safetensors_support import ( + flatten_tensor_state_dict, + unflatten_tensor_state_dict, + ) + from torchao.prototype.safetensors.safetensors_utils import is_metadata_torchao + + +def _update_torch_safe_globals(): + safe_globals = [ + (torch.uint1, "torch.uint1"), + (torch.uint2, "torch.uint2"), + (torch.uint3, "torch.uint3"), + (torch.uint4, "torch.uint4"), + (torch.uint5, "torch.uint5"), + (torch.uint6, "torch.uint6"), + (torch.uint7, "torch.uint7"), + ] + try: + from torchao.dtypes import NF4Tensor + from torchao.dtypes.uintx.uintx_layout import UintxAQTTensorImpl, UintxTensor + + safe_globals.extend([UintxTensor, UintxAQTTensorImpl, NF4Tensor]) + + # note: is_torchao_version(">=", "0.16.0") does not work correctly + # with torchao nightly, so using a ">" check which does work correctly + if is_torchao_version(">", "0.15.0"): + pass + else: + from torchao.dtypes.floatx.float8_layout import Float8AQTTensorImpl + from torchao.dtypes.uintx.uint4_layout import UInt4Tensor + + safe_globals.extend([UInt4Tensor, Float8AQTTensorImpl]) + + except (ImportError, ModuleNotFoundError) as e: + logger.warning( + "Unable to import `torchao` Tensor objects. This may affect loading checkpoints serialized with `torchao`" + ) + logger.debug(e) + + finally: + torch.serialization.add_safe_globals(safe_globals=safe_globals) + + +if ( + is_torch_available() + and is_torch_version(">=", "2.6.0") + and is_torchao_available() + and is_torchao_version(">=", "0.15.0") +): + _update_torch_safe_globals() + + +def fuzzy_match_size(config_name: str) -> str | None: + """ + Extract the size digit from strings like "4weight", "8weight". Returns the digit as an integer if found, otherwise + None. + """ + config_name = config_name.lower() + + str_match = re.search(r"(\d)weight", config_name) + + if str_match: + return str_match.group(1) + + return None + + +def _linear_extra_repr(self): + from torchao.utils import TorchAOBaseTensor + + weight = self.weight.__class__.__name__ if isinstance(self.weight, TorchAOBaseTensor) else None + if weight is None: + return f"in_features={self.weight.shape[1]}, out_features={self.weight.shape[0]}, weight=None" + else: + return f"in_features={self.weight.shape[1]}, out_features={self.weight.shape[0]}, weight={weight}" + + +class TorchAoHfQuantizer(DiffusersQuantizer): + r""" + Diffusers Quantizer for TorchAO: https://github.com/pytorch/ao/. + """ + + requires_calibration = False + required_packages = ["torchao"] + use_keep_in_fp32_modules = True + + def __init__(self, quantization_config, **kwargs): + super().__init__(quantization_config, **kwargs) + + self._metadata = {} + self._pending_flattened_state_dict = {} + + def validate_environment(self, *args, **kwargs): + if not is_torchao_available(): + raise ImportError( + "Loading a TorchAO quantized model requires the torchao library. Please install with `pip install torchao`" + ) + torchao_version = version.parse(importlib.metadata.version("torchao")) + if torchao_version < version.parse("0.15.0"): + raise RuntimeError( + f"The minimum required version of `torchao` is 0.15.0, but the current version is {torchao_version}. Please upgrade with `pip install -U torchao`." + ) + + self.offload = False + + device_map = kwargs.get("device_map", None) + if isinstance(device_map, dict): + if "cpu" in device_map.values() or "disk" in device_map.values(): + if self.pre_quantized: + raise ValueError( + "You are attempting to perform cpu/disk offload with a pre-quantized torchao model " + "This is not supported yet. Please remove the CPU or disk device from the `device_map` argument." + ) + else: + self.offload = True + + if self.pre_quantized: + weights_only = kwargs.get("weights_only", None) + if weights_only: + torch_version = version.parse(importlib.metadata.version("torch")) + if torch_version < version.parse("2.5.0"): + # TODO(aryan): TorchAO is compatible with Pytorch >= 2.2 for certain quantization types. Try to see if we can support it in future + raise RuntimeError( + f"In order to use TorchAO pre-quantized model, you need to have torch>=2.5.0. However, the current version is {torch_version}." + ) + + def update_torch_dtype(self, torch_dtype): + config_name = self.quantization_config.quant_type.__class__.__name__ + is_int_quant = config_name.startswith("Int") or config_name.startswith("Uint") + if is_int_quant and torch_dtype is not None and torch_dtype != torch.bfloat16: + logger.warning( + f"You are trying to set torch_dtype to {torch_dtype} for integer quantization, but " + f"only bfloat16 is supported right now. Please set `torch_dtype=torch.bfloat16`." + ) + + if torch_dtype is None: + # We need to set the torch_dtype, otherwise we have dtype mismatch when performing the quantized linear op + logger.warning( + "Overriding `torch_dtype` with `torch_dtype=torch.bfloat16` due to requirements of `torchao` " + "to enable model loading in different precisions. Pass your own `torch_dtype` to specify the " + "dtype of the remaining non-linear layers, or pass torch_dtype=torch.bfloat16, to remove this warning." + ) + torch_dtype = torch.bfloat16 + + return torch_dtype + + def adjust_target_dtype(self, target_dtype: "torch.dtype") -> "torch.dtype": + from accelerate.utils import CustomDtype + + quant_type = self.quantization_config.quant_type + config_name = quant_type.__class__.__name__ + size_digit = fuzzy_match_size(config_name) + + if size_digit == "4": + return CustomDtype.INT4 + else: + return torch.int8 + + if isinstance(target_dtype, SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION): + return target_dtype + + # We need one of the supported dtypes to be selected in order for accelerate to determine + # the total size of modules/parameters for auto device placement. + possible_device_maps = ["auto", "balanced", "balanced_low_0", "sequential"] + raise ValueError( + f"You have set `device_map` as one of {possible_device_maps} on a TorchAO quantized model but a suitable target dtype " + f"could not be inferred. The supported target_dtypes are: {SUPPORTED_TORCH_DTYPES_FOR_QUANTIZATION}. If you think the " + f"dtype you are using should be supported, please open an issue at https://github.com/huggingface/diffusers/issues." + ) + + def adjust_max_memory(self, max_memory: dict[str, int | str]) -> dict[str, int | str]: + max_memory = {key: val * 0.9 for key, val in max_memory.items()} + return max_memory + + def get_state_dict_and_metadata(self, state_dict: dict[str, Any], safe_serialization: bool = False): + """ + We flatten the state dict of tensor subclasses so that it is compatible with the safetensors format. + """ + if not safe_serialization or not is_torchao_available() or not is_torchao_version(">=", "0.16.0"): + return state_dict, {} + + flattened_state_dict = flatten_tensor_state_dict(state_dict) + if isinstance(flattened_state_dict, tuple): + return flattened_state_dict + + return flattened_state_dict, {} + + def maybe_update_loaded_keys(self, loaded_keys: list[str], checkpoint_files: list[str]) -> list[str]: + self.set_metadata(checkpoint_files) + if self._metadata: + return list(self.get_weight_names()) + return loaded_keys + + def set_metadata(self, checkpoint_files: list[str]): + self._metadata = {} + self._pending_flattened_state_dict = {} + + if not is_safetensors_available() or not is_torchao_version(">=", "0.16.0"): + return + + if len(checkpoint_files) == 0: + return + + if not all( + isinstance(checkpoint, str) and checkpoint.endswith(".safetensors") for checkpoint in checkpoint_files + ): + return + + metadata = {} + for checkpoint in checkpoint_files: + with safe_open(checkpoint, framework="pt") as f: + metadata.update(f.metadata() or {}) + + self._metadata = metadata if is_metadata_torchao(metadata) else {} + + @property + def metadata(self): + return self._metadata + + def maybe_update_state_dict(self, state_dict: dict[str, Any]) -> dict[str, Any]: + if not self._metadata or not is_torchao_version(">=", "0.16.0") or not is_metadata_torchao(self._metadata): + return state_dict + + merged_state_dict = {**self._pending_flattened_state_dict, **state_dict} + reconstructed_state_dict, self._pending_flattened_state_dict = unflatten_tensor_state_dict( + merged_state_dict, self._metadata + ) + + return reconstructed_state_dict + + @property + def supports_parallel_loading(self) -> bool: + # Safetensors reconstruction can carry leftover flattened tensor pieces from one shard to the next. + return not self._metadata + + def get_weight_names(self): + if not self._metadata: + return set() + return set(json.loads(self._metadata["tensor_names"])) + + def check_if_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + state_dict: dict[str, Any], + **kwargs, + ) -> bool: + param_device = kwargs.pop("param_device", None) + # Check if the param_name is not in self.modules_to_not_convert + if any((key + "." in param_name) or (key == param_name) for key in self.modules_to_not_convert): + return False + elif param_device == "cpu" and self.offload: + # We don't quantize weights that we offload + return False + else: + # We only quantize the weight of nn.Linear + module, tensor_name = get_module_from_name(model, param_name) + return isinstance(module, torch.nn.Linear) and (tensor_name == "weight") + + def create_quantized_param( + self, + model: "ModelMixin", + param_value: "torch.Tensor", + param_name: str, + target_device: "torch.device", + state_dict: dict[str, Any], + unexpected_keys: list[str], + **kwargs, + ): + r""" + Each nn.Linear layer that needs to be quantized is processed here. First, we set the value the weight tensor, + then we move it to the target device. Finally, we quantize the module. + """ + module, tensor_name = get_module_from_name(model, param_name) + + if self.pre_quantized: + # If we're loading pre-quantized weights, replace the repr of linear layers for pretty printing info + # about the quantized tensor type + module._parameters[tensor_name] = torch.nn.Parameter(param_value.to(device=target_device)) + if isinstance(module, nn.Linear): + module.extra_repr = types.MethodType(_linear_extra_repr, module) + else: + # As we perform quantization here, the repr of linear layers is set by TorchAO, so we don't have to do it ourselves + module._parameters[tensor_name] = torch.nn.Parameter(param_value).to(device=target_device) + quantize_(module, self.quantization_config.get_apply_tensor_subclass()) + + def get_cuda_warm_up_factor(self): + """ + This factor is used in caching_allocator_warmup to determine how many bytes to pre-allocate for CUDA warmup. + - A factor of 2 means we pre-allocate the full memory footprint of the model. + - A factor of 4 means we pre-allocate half of that, and so on + + However, when using TorchAO, calculating memory usage with param.numel() * param.element_size() doesn't give + the correct size for quantized weights (like int4 or int8) That's because TorchAO internally represents + quantized tensors using subtensors and metadata, and the reported element_size() still corresponds to the + torch_dtype not the actual bit-width of the quantized data. + + To correct for this: + - Use a division factor of 8 for int4 weights + - Use a division factor of 4 for int8 weights + """ + quant_type = self.quantization_config.quant_type + config_name = quant_type.__class__.__name__ + size_digit = fuzzy_match_size(config_name) + + if size_digit == "4": + return 8 + else: + return 4 + + def _process_model_before_weight_loading( + self, + model: "ModelMixin", + device_map, + keep_in_fp32_modules: list[str] = [], + **kwargs, + ): + self.modules_to_not_convert = self.quantization_config.modules_to_not_convert + + if not isinstance(self.modules_to_not_convert, list): + self.modules_to_not_convert = [self.modules_to_not_convert] + + self.modules_to_not_convert.extend(keep_in_fp32_modules) + + # Extend `self.modules_to_not_convert` to keys that are supposed to be offloaded to `cpu` or `disk` + if isinstance(device_map, dict) and len(device_map.keys()) > 1: + keys_on_cpu = [key for key, value in device_map.items() if value in ["disk", "cpu"]] + self.modules_to_not_convert.extend(keys_on_cpu) + + # Purge `None`. + # Unlike `transformers`, we don't know if we should always keep certain modules in FP32 + # in case of diffusion transformer models. For language models and others alike, `lm_head` + # and tied modules are usually kept in FP32. + self.modules_to_not_convert = [module for module in self.modules_to_not_convert if module is not None] + + model.config.quantization_config = self.quantization_config + + def _process_model_after_weight_loading(self, model: "ModelMixin"): + return model + + @property + def supports_safetensors_serialization(self): + if not is_torchao_version(">=", "0.16.0"): + logger.warning( + "TorchAO quantized model is not serializable with safe serialization without safetensors support " + "from the installed torchao version." + ) + return False + + return True + + @property + def is_serializable(self): + _is_torchao_serializable = version.parse(importlib.metadata.version("huggingface_hub")) >= version.parse( + "0.25.0" + ) + + if not _is_torchao_serializable: + logger.warning("torchao quantized model is only serializable after huggingface_hub >= 0.25.0 ") + + if self.offload and self.quantization_config.modules_to_not_convert is None: + logger.warning( + "The model contains offloaded modules and these modules are not quantized. We don't recommend saving the model as we won't be able to reload them." + "If you want to specify modules to not quantize, please specify modules_to_not_convert in the quantization_config." + ) + return False + + return _is_torchao_serializable + + _TRAINABLE_QUANTIZATION_CONFIGS = ( + "Int8WeightOnlyConfig", + "Int8DynamicActivationInt8WeightConfig", + "Int8StaticActivationInt8WeightConfig", + "Float8WeightOnlyConfig", + "Float8DynamicActivationFloat8WeightConfig", + ) + + @property + def is_trainable(self): + return self.quantization_config.quant_type.__class__.__name__ in self._TRAINABLE_QUANTIZATION_CONFIGS + + @property + def is_compileable(self) -> bool: + return True + + def _dequantize(self, model): + from torchao.utils import TorchAOBaseTensor + + for name, module in model.named_modules(): + if isinstance(module, nn.Linear) and isinstance(module.weight, TorchAOBaseTensor): + if not hasattr(module.weight, "dequantize"): + raise NotImplementedError( + f"Dequantization is not supported for {type(module.weight).__name__} " + f"(module: {name}). Please use a quantization type that supports dequantization." + ) + device = module.weight.device + dequantized_weight = module.weight.dequantize().to(device) + module.weight = nn.Parameter(dequantized_weight) + # Reset extra_repr if it was overridden + if hasattr(module.extra_repr, "__func__") and module.extra_repr.__func__ is not nn.Linear.extra_repr: + module.extra_repr = types.MethodType(nn.Linear.extra_repr, module) + + return model diff --git a/diffusers/schedulers/README.md b/diffusers/schedulers/README.md new file mode 100644 index 0000000000000000000000000000000000000000..31ad27793e34783faabc222adf98691fb396a0d8 --- /dev/null +++ b/diffusers/schedulers/README.md @@ -0,0 +1,3 @@ +# Schedulers + +For more information on the schedulers, please refer to the [docs](https://huggingface.co/docs/diffusers/api/schedulers/overview). \ No newline at end of file diff --git a/diffusers/schedulers/__init__.py b/diffusers/schedulers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..c0e46ef445df4ba86e8397270fab905954814aa7 --- /dev/null +++ b/diffusers/schedulers/__init__.py @@ -0,0 +1,193 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import TYPE_CHECKING + +from ..utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_scipy_available, + is_torch_available, + is_torchsde_available, +) + + +_dummy_modules = {} +_import_structure = {} + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_pt_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_pt_objects)) + +else: + _import_structure["deprecated"] = ["KarrasVeScheduler", "ScoreSdeVpScheduler"] + _import_structure["scheduling_amused"] = ["AmusedScheduler"] + _import_structure["scheduling_block_refinement"] = ["BlockRefinementScheduler", "BlockRefinementSchedulerOutput"] + _import_structure["scheduling_consistency_decoder"] = ["ConsistencyDecoderScheduler"] + _import_structure["scheduling_consistency_models"] = ["CMStochasticIterativeScheduler"] + _import_structure["scheduling_ddim"] = ["DDIMScheduler"] + _import_structure["scheduling_ddim_cogvideox"] = ["CogVideoXDDIMScheduler"] + _import_structure["scheduling_ddim_inverse"] = ["DDIMInverseScheduler"] + _import_structure["scheduling_ddim_parallel"] = ["DDIMParallelScheduler"] + _import_structure["scheduling_ddpm"] = ["DDPMScheduler"] + _import_structure["scheduling_ddpm_parallel"] = ["DDPMParallelScheduler"] + _import_structure["scheduling_ddpm_wuerstchen"] = ["DDPMWuerstchenScheduler"] + _import_structure["scheduling_deis_multistep"] = ["DEISMultistepScheduler"] + _import_structure["scheduling_discrete_ddim"] = ["DiscreteDDIMScheduler", "DiscreteDDIMSchedulerOutput"] + _import_structure["scheduling_dpm_cogvideox"] = ["CogVideoXDPMScheduler"] + _import_structure["scheduling_dpmsolver_multistep"] = ["DPMSolverMultistepScheduler"] + _import_structure["scheduling_dpmsolver_multistep_inverse"] = ["DPMSolverMultistepInverseScheduler"] + _import_structure["scheduling_dpmsolver_singlestep"] = ["DPMSolverSinglestepScheduler"] + _import_structure["scheduling_edm_dpmsolver_multistep"] = ["EDMDPMSolverMultistepScheduler"] + _import_structure["scheduling_edm_euler"] = ["EDMEulerScheduler"] + _import_structure["scheduling_entropy_bound"] = ["EntropyBoundScheduler", "EntropyBoundSchedulerOutput"] + _import_structure["scheduling_euler_ancestral_discrete"] = ["EulerAncestralDiscreteScheduler"] + _import_structure["scheduling_euler_discrete"] = ["EulerDiscreteScheduler"] + _import_structure["scheduling_flow_map_euler_discrete"] = ["FlowMapEulerDiscreteScheduler"] + _import_structure["scheduling_flow_match_euler_discrete"] = ["FlowMatchEulerDiscreteScheduler"] + _import_structure["scheduling_flow_match_heun_discrete"] = ["FlowMatchHeunDiscreteScheduler"] + _import_structure["scheduling_flow_match_lcm"] = ["FlowMatchLCMScheduler"] + _import_structure["scheduling_helios"] = ["HeliosScheduler"] + _import_structure["scheduling_helios_dmd"] = ["HeliosDMDScheduler"] + _import_structure["scheduling_heun_discrete"] = ["HeunDiscreteScheduler"] + _import_structure["scheduling_ipndm"] = ["IPNDMScheduler"] + _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"] + _import_structure["scheduling_k_dpm_2_discrete"] = ["KDPM2DiscreteScheduler"] + _import_structure["scheduling_lcm"] = ["LCMScheduler"] + _import_structure["scheduling_ltx_euler_ancestral_rf"] = ["LTXEulerAncestralRFScheduler"] + _import_structure["scheduling_minimax_h3"] = ["MiniMaxH3Scheduler"] + _import_structure["scheduling_pndm"] = ["PNDMScheduler"] + _import_structure["scheduling_repaint"] = ["RePaintScheduler"] + _import_structure["scheduling_sasolver"] = ["SASolverScheduler"] + _import_structure["scheduling_scm"] = ["SCMScheduler"] + _import_structure["scheduling_sde_ve"] = ["ScoreSdeVeScheduler"] + _import_structure["scheduling_tcd"] = ["TCDScheduler"] + _import_structure["scheduling_unclip"] = ["UnCLIPScheduler"] + _import_structure["scheduling_unipc_multistep"] = ["UniPCMultistepScheduler"] + _import_structure["scheduling_utils"] = ["AysSchedules", "KarrasDiffusionSchedulers", "SchedulerMixin"] + _import_structure["scheduling_vq_diffusion"] = ["VQDiffusionScheduler"] + +try: + if not (is_torch_available() and is_scipy_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_scipy_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_torch_and_scipy_objects)) + +else: + _import_structure["scheduling_lms_discrete"] = ["LMSDiscreteScheduler"] + +try: + if not (is_torch_available() and is_torchsde_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ..utils import dummy_torch_and_torchsde_objects # noqa F403 + + _dummy_modules.update(get_objects_from_module(dummy_torch_and_torchsde_objects)) + +else: + _import_structure["scheduling_cosine_dpmsolver_multistep"] = ["CosineDPMSolverMultistepScheduler"] + _import_structure["scheduling_dpmsolver_sde"] = ["DPMSolverSDEScheduler"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + from ..utils import ( + OptionalDependencyNotAvailable, + is_scipy_available, + is_torch_available, + is_torchsde_available, + ) + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_pt_objects import * # noqa F403 + else: + from .deprecated import KarrasVeScheduler, ScoreSdeVpScheduler + from .scheduling_amused import AmusedScheduler + from .scheduling_block_refinement import BlockRefinementScheduler, BlockRefinementSchedulerOutput + from .scheduling_consistency_decoder import ConsistencyDecoderScheduler + from .scheduling_consistency_models import CMStochasticIterativeScheduler + from .scheduling_ddim import DDIMScheduler + from .scheduling_ddim_cogvideox import CogVideoXDDIMScheduler + from .scheduling_ddim_inverse import DDIMInverseScheduler + from .scheduling_ddim_parallel import DDIMParallelScheduler + from .scheduling_ddpm import DDPMScheduler + from .scheduling_ddpm_parallel import DDPMParallelScheduler + from .scheduling_ddpm_wuerstchen import DDPMWuerstchenScheduler + from .scheduling_deis_multistep import DEISMultistepScheduler + from .scheduling_discrete_ddim import DiscreteDDIMScheduler, DiscreteDDIMSchedulerOutput + from .scheduling_dpm_cogvideox import CogVideoXDPMScheduler + from .scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler + from .scheduling_dpmsolver_multistep_inverse import DPMSolverMultistepInverseScheduler + from .scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler + from .scheduling_edm_dpmsolver_multistep import EDMDPMSolverMultistepScheduler + from .scheduling_edm_euler import EDMEulerScheduler + from .scheduling_entropy_bound import EntropyBoundScheduler, EntropyBoundSchedulerOutput + from .scheduling_euler_ancestral_discrete import EulerAncestralDiscreteScheduler + from .scheduling_euler_discrete import EulerDiscreteScheduler + from .scheduling_flow_map_euler_discrete import FlowMapEulerDiscreteScheduler + from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler + from .scheduling_flow_match_heun_discrete import FlowMatchHeunDiscreteScheduler + from .scheduling_flow_match_lcm import FlowMatchLCMScheduler + from .scheduling_helios import HeliosScheduler + from .scheduling_helios_dmd import HeliosDMDScheduler + from .scheduling_heun_discrete import HeunDiscreteScheduler + from .scheduling_ipndm import IPNDMScheduler + from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler + from .scheduling_k_dpm_2_discrete import KDPM2DiscreteScheduler + from .scheduling_lcm import LCMScheduler + from .scheduling_ltx_euler_ancestral_rf import LTXEulerAncestralRFScheduler + from .scheduling_minimax_h3 import MiniMaxH3Scheduler + from .scheduling_pndm import PNDMScheduler + from .scheduling_repaint import RePaintScheduler + from .scheduling_sasolver import SASolverScheduler + from .scheduling_scm import SCMScheduler + from .scheduling_sde_ve import ScoreSdeVeScheduler + from .scheduling_tcd import TCDScheduler + from .scheduling_unclip import UnCLIPScheduler + from .scheduling_unipc_multistep import UniPCMultistepScheduler + from .scheduling_utils import AysSchedules, KarrasDiffusionSchedulers, SchedulerMixin + from .scheduling_vq_diffusion import VQDiffusionScheduler + + try: + if not (is_torch_available() and is_scipy_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_scipy_objects import * # noqa F403 + else: + from .scheduling_lms_discrete import LMSDiscreteScheduler + + try: + if not (is_torch_available() and is_torchsde_available()): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + from ..utils.dummy_torch_and_torchsde_objects import * # noqa F403 + else: + from .scheduling_cosine_dpmsolver_multistep import CosineDPMSolverMultistepScheduler + from .scheduling_dpmsolver_sde import DPMSolverSDEScheduler + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) + for name, value in _dummy_modules.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/schedulers/deprecated/__init__.py b/diffusers/schedulers/deprecated/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..479cf9bd568be320b60e87a831bf80835e9943f2 --- /dev/null +++ b/diffusers/schedulers/deprecated/__init__.py @@ -0,0 +1,50 @@ +from typing import TYPE_CHECKING + +from ...utils import ( + DIFFUSERS_SLOW_IMPORT, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_torch_available, + is_transformers_available, +) + + +_dummy_objects = {} +_import_structure = {} + +try: + if not (is_transformers_available() and is_torch_available()): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + from ...utils import dummy_pt_objects # noqa F403 + + _dummy_objects.update(get_objects_from_module(dummy_pt_objects)) +else: + _import_structure["scheduling_karras_ve"] = ["KarrasVeScheduler"] + _import_structure["scheduling_sde_vp"] = ["ScoreSdeVpScheduler"] + +if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT: + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + + except OptionalDependencyNotAvailable: + from ...utils.dummy_pt_objects import * # noqa F403 + else: + from .scheduling_karras_ve import KarrasVeScheduler + from .scheduling_sde_vp import ScoreSdeVpScheduler + + +else: + import sys + + sys.modules[__name__] = _LazyModule( + __name__, + globals()["__file__"], + _import_structure, + module_spec=__spec__, + ) + + for name, value in _dummy_objects.items(): + setattr(sys.modules[__name__], name, value) diff --git a/diffusers/schedulers/deprecated/scheduling_karras_ve.py b/diffusers/schedulers/deprecated/scheduling_karras_ve.py new file mode 100644 index 0000000000000000000000000000000000000000..b6b905b0f08803352d575ffef29618ac744a9975 --- /dev/null +++ b/diffusers/schedulers/deprecated/scheduling_karras_ve.py @@ -0,0 +1,239 @@ +# Copyright 2025 NVIDIA and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from dataclasses import dataclass + +import numpy as np +import torch + +from ...configuration_utils import ConfigMixin, register_to_config +from ...utils import BaseOutput +from ...utils.torch_utils import randn_tensor +from ..scheduling_utils import SchedulerMixin + + +@dataclass +class KarrasVeOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + derivative (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Derivative of predicted original image sample (x_0). + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample (x_{0}) based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + derivative: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +class KarrasVeScheduler(SchedulerMixin, ConfigMixin): + """ + A stochastic scheduler tailored to variance-expanding models. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + > [!TIP] > For more details on the parameters, see [Appendix E](https://huggingface.co/papers/2206.00364). The grid + search > values used to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in + Table 5 of > the paper. + + Args: + sigma_min (`float`, defaults to 0.02): + The minimum noise magnitude. + sigma_max (`float`, defaults to 100): + The maximum noise magnitude. + s_noise (`float`, defaults to 1.007): + The amount of additional noise to counteract loss of detail during sampling. A reasonable range is [1.000, + 1.011]. + s_churn (`float`, defaults to 80): + The parameter controlling the overall amount of stochasticity. A reasonable range is [0, 100]. + s_min (`float`, defaults to 0.05): + The start value of the sigma range to add noise (enable stochasticity). A reasonable range is [0, 10]. + s_max (`float`, defaults to 50): + The end value of the sigma range to add noise. A reasonable range is [0.2, 80]. + """ + + order = 2 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.02, + sigma_max: float = 100, + s_noise: float = 1.007, + s_churn: float = 80, + s_min: float = 0.05, + s_max: float = 50, + ): + # standard deviation of the initial noise distribution + self.init_noise_sigma = sigma_max + + # setable values + self.num_inference_steps: int = None + self.timesteps: np.IntTensor = None + self.schedule: torch.Tensor = None # sigma(t_i) + + def scale_model_input(self, sample: torch.Tensor, timestep: int = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + timesteps = np.arange(0, self.num_inference_steps)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps).to(device) + schedule = [ + ( + self.config.sigma_max**2 + * (self.config.sigma_min**2 / self.config.sigma_max**2) ** (i / (num_inference_steps - 1)) + ) + for i in self.timesteps + ] + self.schedule = torch.tensor(schedule, dtype=torch.float32, device=device) + + def add_noise_to_input( + self, sample: torch.Tensor, sigma: float, generator: torch.Generator | None = None + ) -> tuple[torch.Tensor, float]: + """ + Explicit Langevin-like "churn" step of adding noise to the sample according to a `gamma_i ≥ 0` to reach a + higher noise level `sigma_hat = sigma_i + gamma_i*sigma_i`. + + Args: + sample (`torch.Tensor`): + The input sample. + sigma (`float`): + generator (`torch.Generator`, *optional*): + A random number generator. + """ + if self.config.s_min <= sigma <= self.config.s_max: + gamma = min(self.config.s_churn / self.num_inference_steps, 2**0.5 - 1) + else: + gamma = 0 + + # sample eps ~ N(0, S_noise^2 * I) + eps = self.config.s_noise * randn_tensor(sample.shape, generator=generator).to(sample.device) + sigma_hat = sigma + gamma * sigma + sample_hat = sample + ((sigma_hat**2 - sigma**2) ** 0.5 * eps) + + return sample_hat, sigma_hat + + def step( + self, + model_output: torch.Tensor, + sigma_hat: float, + sigma_prev: float, + sample_hat: torch.Tensor, + return_dict: bool = True, + ) -> KarrasVeOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + sigma_hat (`float`): + sigma_prev (`float`): + sample_hat (`torch.Tensor`): + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + + """ + + pred_original_sample = sample_hat + sigma_hat * model_output + derivative = (sample_hat - pred_original_sample) / sigma_hat + sample_prev = sample_hat + (sigma_prev - sigma_hat) * derivative + + if not return_dict: + return (sample_prev, derivative) + + return KarrasVeOutput( + prev_sample=sample_prev, derivative=derivative, pred_original_sample=pred_original_sample + ) + + def step_correct( + self, + model_output: torch.Tensor, + sigma_hat: float, + sigma_prev: float, + sample_hat: torch.Tensor, + sample_prev: torch.Tensor, + derivative: torch.Tensor, + return_dict: bool = True, + ) -> KarrasVeOutput | tuple: + """ + Corrects the predicted sample based on the `model_output` of the network. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + sigma_hat (`float`): TODO + sigma_prev (`float`): TODO + sample_hat (`torch.Tensor`): TODO + sample_prev (`torch.Tensor`): TODO + derivative (`torch.Tensor`): TODO + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`. + + Returns: + prev_sample (TODO): updated sample in the diffusion chain. derivative (TODO): TODO + + """ + pred_original_sample = sample_prev + sigma_prev * model_output + derivative_corr = (sample_prev - pred_original_sample) / sigma_prev + sample_prev = sample_hat + (sigma_prev - sigma_hat) * (0.5 * derivative + 0.5 * derivative_corr) + + if not return_dict: + return (sample_prev, derivative) + + return KarrasVeOutput( + prev_sample=sample_prev, derivative=derivative, pred_original_sample=pred_original_sample + ) + + def add_noise(self, original_samples, noise, timesteps): + raise NotImplementedError() diff --git a/diffusers/schedulers/deprecated/scheduling_sde_vp.py b/diffusers/schedulers/deprecated/scheduling_sde_vp.py new file mode 100644 index 0000000000000000000000000000000000000000..99fa408e70c0d5eb010b9dd72e8fa62ad461abd8 --- /dev/null +++ b/diffusers/schedulers/deprecated/scheduling_sde_vp.py @@ -0,0 +1,108 @@ +# Copyright 2025 Google Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch + +import math + +import torch + +from ...configuration_utils import ConfigMixin, register_to_config +from ...utils.torch_utils import randn_tensor +from ..scheduling_utils import SchedulerMixin + + +class ScoreSdeVpScheduler(SchedulerMixin, ConfigMixin): + """ + `ScoreSdeVpScheduler` is a variance preserving stochastic differential equation (SDE) scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 2000): + The number of diffusion steps to train the model. + beta_min (`int`, defaults to 0.1): + beta_max (`int`, defaults to 20): + sampling_eps (`int`, defaults to 1e-3): + The end value of sampling where timesteps decrease progressively from 1 to epsilon. + """ + + order = 1 + + @register_to_config + def __init__(self, num_train_timesteps=2000, beta_min=0.1, beta_max=20, sampling_eps=1e-3): + self.sigmas = None + self.discrete_sigmas = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps, device: str | torch.device = None): + """ + Sets the continuous timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.timesteps = torch.linspace(1, self.config.sampling_eps, num_inference_steps, device=device) + + def step_pred(self, score, x, t, generator=None): + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + score (): + x (): + t (): + generator (`torch.Generator`, *optional*): + A random number generator. + """ + if self.timesteps is None: + raise ValueError( + "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + # TODO(Patrick) better comments + non-PyTorch + # postprocess model score + log_mean_coeff = -0.25 * t**2 * (self.config.beta_max - self.config.beta_min) - 0.5 * t * self.config.beta_min + std = torch.sqrt(1.0 - torch.exp(2.0 * log_mean_coeff)) + std = std.flatten() + while len(std.shape) < len(score.shape): + std = std.unsqueeze(-1) + score = -score / std + + # compute + dt = -1.0 / len(self.timesteps) + + beta_t = self.config.beta_min + t * (self.config.beta_max - self.config.beta_min) + beta_t = beta_t.flatten() + while len(beta_t.shape) < len(x.shape): + beta_t = beta_t.unsqueeze(-1) + drift = -0.5 * beta_t * x + + diffusion = torch.sqrt(beta_t) + drift = drift - diffusion**2 * score + x_mean = x + drift * dt + + # add noise + noise = randn_tensor(x.shape, layout=x.layout, generator=generator, device=x.device, dtype=x.dtype) + x = x_mean + diffusion * math.sqrt(-dt) * noise + + return x, x_mean + + def __len__(self): + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_amused.py b/diffusers/schedulers/scheduling_amused.py new file mode 100644 index 0000000000000000000000000000000000000000..ef49d8db91b1c94e254b18a1f446ff1adb27b451 --- /dev/null +++ b/diffusers/schedulers/scheduling_amused.py @@ -0,0 +1,235 @@ +import math +from dataclasses import dataclass +from typing import Literal + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +def gumbel_noise(t: torch.Tensor, generator: torch.Generator | None = None) -> torch.Tensor: + """ + Generate Gumbel noise for sampling. + + Args: + t (`torch.Tensor`): + Input tensor to match the shape and dtype of the output noise. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible sampling. + + Returns: + `torch.Tensor`: + Gumbel-distributed noise with the same shape, dtype, and device as the input tensor. + """ + device = generator.device if generator is not None else t.device + noise = torch.zeros_like(t, device=device).uniform_(0, 1, generator=generator).to(t.device) + return -torch.log((-torch.log(noise.clamp(1e-20))).clamp(1e-20)) + + +def mask_by_random_topk( + mask_len: torch.Tensor, + probs: torch.Tensor, + temperature: float = 1.0, + generator: torch.Generator | None = None, +) -> torch.Tensor: + """ + Mask tokens by selecting the top-k lowest confidence scores with temperature-based randomness. + + Args: + mask_len (`torch.Tensor`): + Number of tokens to mask per sample in the batch. + probs (`torch.Tensor`): + Probability scores for each token. + temperature (`float`, *optional*, defaults to 1.0): + Temperature parameter for controlling randomness in the masking process. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible sampling. + + Returns: + `torch.Tensor`: + Boolean mask indicating which tokens should be masked. + """ + confidence = torch.log(probs.clamp(1e-20)) + temperature * gumbel_noise(probs, generator=generator) + sorted_confidence = torch.sort(confidence, dim=-1).values + cut_off = torch.gather(sorted_confidence, 1, mask_len.long()) + masking = confidence < cut_off + return masking + + +@dataclass +class AmusedSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.LongTensor` of shape `(batch_size, height, width)` or `(batch_size, sequence_length)`): + Computed sample `(x_{t-1})` of previous timestep with token IDs. `prev_sample` should be used as next model + input in the denoising loop. + pred_original_sample (`torch.LongTensor` of shape `(batch_size, height, width)` or `(batch_size, sequence_length)`, *optional*): + The predicted fully denoised sample `(x_{0})` with token IDs based on the model output from the current + timestep. `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Generator | None = None + + +class AmusedScheduler(SchedulerMixin, ConfigMixin): + """ + A scheduler for masked token generation as used in [`AmusedPipeline`]. + + This scheduler iteratively unmasks tokens based on their confidence scores, following either a cosine or linear + schedule. Unlike traditional diffusion schedulers that work with continuous pixel values, this scheduler operates + on discrete token IDs, making it suitable for autoregressive and non-autoregressive masked token generation models. + + This scheduler inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the + generic methods the library implements for all schedulers such as loading and saving. + + Args: + mask_token_id (`int`): + The token ID used to represent masked tokens in the sequence. + masking_schedule (`Literal["cosine", "linear"]`, *optional*, defaults to `"cosine"`): + The schedule type for determining the mask ratio at each timestep. Can be either `"cosine"` or `"linear"`. + """ + + order = 1 + + temperatures: torch.Generator | None + timesteps: torch.Generator | None + + @register_to_config + def __init__( + self, + mask_token_id: int, + masking_schedule: Literal["cosine", "linear"] = "cosine", + ): + self.temperatures = None + self.timesteps = None + + def set_timesteps( + self, + num_inference_steps: int, + temperature: int | tuple[int, int] | list[int] = (2, 0), + device: str | torch.device = None, + ): + self.timesteps = torch.arange(num_inference_steps, device=device).flip(0) + + if isinstance(temperature, (tuple, list)): + self.temperatures = torch.linspace(temperature[0], temperature[1], num_inference_steps, device=device) + else: + self.temperatures = torch.linspace(temperature, 0.01, num_inference_steps, device=device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.LongTensor, + starting_mask_ratio: float = 1.0, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> AmusedSchedulerOutput | tuple: + two_dim_input = sample.ndim == 3 and model_output.ndim == 4 + + if two_dim_input: + batch_size, codebook_size, height, width = model_output.shape + sample = sample.reshape(batch_size, height * width) + model_output = model_output.reshape(batch_size, codebook_size, height * width).permute(0, 2, 1) + + unknown_map = sample == self.config.mask_token_id + + probs = model_output.softmax(dim=-1) + + device = probs.device + probs_ = probs.to(generator.device) if generator is not None else probs # handles when generator is on CPU + if probs_.device.type == "cpu" and probs_.dtype != torch.float32: + probs_ = probs_.float() # multinomial is not implemented for cpu half precision + probs_ = probs_.reshape(-1, probs.size(-1)) + pred_original_sample = torch.multinomial(probs_, 1, generator=generator).to(device=device) + pred_original_sample = pred_original_sample[:, 0].view(*probs.shape[:-1]) + pred_original_sample = torch.where(unknown_map, pred_original_sample, sample) + + if timestep == 0: + prev_sample = pred_original_sample + else: + seq_len = sample.shape[1] + step_idx = (self.timesteps == timestep).nonzero() + ratio = (step_idx + 1) / len(self.timesteps) + + if self.config.masking_schedule == "cosine": + mask_ratio = torch.cos(ratio * math.pi / 2) + elif self.config.masking_schedule == "linear": + mask_ratio = 1 - ratio + else: + raise ValueError(f"unknown masking schedule {self.config.masking_schedule}") + + mask_ratio = starting_mask_ratio * mask_ratio + + mask_len = (seq_len * mask_ratio).floor() + # do not mask more than amount previously masked + mask_len = torch.min(unknown_map.sum(dim=-1, keepdim=True) - 1, mask_len) + # mask at least one + mask_len = torch.max(torch.tensor([1], device=model_output.device), mask_len) + + selected_probs = torch.gather(probs, -1, pred_original_sample[:, :, None])[:, :, 0] + # Ignores the tokens given in the input by overwriting their confidence. + selected_probs = torch.where(unknown_map, selected_probs, torch.finfo(selected_probs.dtype).max) + + masking = mask_by_random_topk(mask_len, selected_probs, self.temperatures[step_idx], generator) + + # Masks tokens with lower confidence. + prev_sample = torch.where(masking, self.config.mask_token_id, pred_original_sample) + + if two_dim_input: + prev_sample = prev_sample.reshape(batch_size, height, width) + pred_original_sample = pred_original_sample.reshape(batch_size, height, width) + + if not return_dict: + return (prev_sample, pred_original_sample) + + return AmusedSchedulerOutput(prev_sample, pred_original_sample) + + def add_noise( + self, + sample: torch.LongTensor, + timesteps: int, + generator: torch.Generator | None = None, + ) -> torch.LongTensor: + """ + Add noise to a sample by randomly masking tokens according to the masking schedule. + + Args: + sample (`torch.LongTensor`): + The input sample containing token IDs to be partially masked. + timesteps (`int`): + The timestep that determines how much masking to apply. Higher timesteps result in more masking. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible masking. + + Returns: + `torch.LongTensor`: + The sample with some tokens replaced by `mask_token_id` according to the masking schedule. + """ + step_idx = (self.timesteps == timesteps).nonzero() + ratio = (step_idx + 1) / len(self.timesteps) + + if self.config.masking_schedule == "cosine": + mask_ratio = torch.cos(ratio * math.pi / 2) + elif self.config.masking_schedule == "linear": + mask_ratio = 1 - ratio + else: + raise ValueError(f"unknown masking schedule {self.config.masking_schedule}") + + mask_indices = ( + torch.rand( + sample.shape, device=generator.device if generator is not None else sample.device, generator=generator + ).to(sample.device) + < mask_ratio + ) + + masked_sample = sample.clone() + + masked_sample[mask_indices] = self.config.mask_token_id + + return masked_sample diff --git a/diffusers/schedulers/scheduling_block_refinement.py b/diffusers/schedulers/scheduling_block_refinement.py new file mode 100644 index 0000000000000000000000000000000000000000..6ff7963748b082fb4f780163a49724833c42c830 --- /dev/null +++ b/diffusers/schedulers/scheduling_block_refinement.py @@ -0,0 +1,520 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from dataclasses import dataclass + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +@dataclass +class BlockRefinementSchedulerOutput(BaseOutput): + """ + Output class for block refinement scheduling. + + Args: + prev_sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Updated block tokens after the current refinement step. + transfer_index (`torch.BoolTensor` of shape `(batch_size, block_length)`): + Boolean mask indicating which tokens were committed (mask-filling). + editing_transfer_index (`torch.BoolTensor` of shape `(batch_size, block_length)`): + Boolean mask indicating which tokens were edited (non-mask replacement). + sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`): + Sampled token IDs from the model logits. + sampled_probs (`torch.Tensor` of shape `(batch_size, block_length)`): + Probabilities of the sampled tokens. + pred_logits (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + The denoiser logits, passed through for self-conditioning the next step. + """ + + prev_sample: torch.LongTensor + transfer_index: torch.BoolTensor + editing_transfer_index: torch.BoolTensor + sampled_tokens: torch.LongTensor + sampled_probs: torch.Tensor + pred_logits: torch.Tensor + + +class BlockRefinementScheduler(SchedulerMixin, ConfigMixin): + """ + Scheduler for block-wise iterative refinement (commit-by-confidence). + + At each step, the scheduler samples candidate tokens from model logits and commits those with the highest + confidence. The number of tokens to commit per step is determined by evenly distributing the block length across + the number of refinement steps. + + Optionally supports editing: after all mask tokens are resolved, tokens can be replaced if the model predicts a + different token with confidence above a positive `editing_threshold` (`None`, `0.0`, or negative disables editing). + """ + + order = 1 + + @register_to_config + def __init__( + self, + block_length: int = 32, + num_inference_steps: int = 32, + threshold: float = 0.95, + editing_threshold: float | None = None, + minimal_topk: int = 1, + ): + self.num_inference_steps = num_inference_steps + self.timesteps = torch.arange(self.num_inference_steps - 1, -1, -1, dtype=torch.long) + self._transfer_schedule: torch.LongTensor | None = None + # committed positions for the uniform corruption mode (no mask token); reset at the start of each block + self._committed: torch.BoolTensor | None = None + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device | None = None, + block_length: int | None = None, + ) -> None: + if num_inference_steps <= 0: + raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.") + if block_length is None: + block_length = self.config.block_length + elif block_length <= 0: + raise ValueError(f"`block_length` must be > 0, got {block_length}.") + self.num_inference_steps = num_inference_steps + self.timesteps = torch.arange(self.num_inference_steps - 1, -1, -1, device=device, dtype=torch.long) + self._transfer_schedule = self.get_num_transfer_tokens(block_length, self.num_inference_steps).to( + device=device if device is not None else "cpu" + ) + self._committed = None + + def get_num_transfer_tokens(self, block_length: int, num_inference_steps: int) -> torch.LongTensor: + """Evenly distribute `block_length` token commits across `num_inference_steps` steps.""" + if num_inference_steps <= 0: + return torch.zeros((0,), dtype=torch.long) + base = block_length // num_inference_steps + remainder = block_length % num_inference_steps + out = torch.full((num_inference_steps,), base, dtype=torch.long) + out[:remainder] += 1 + return out + + # --- SAR sampling utilities --- + + @staticmethod + def _top_p_filtering(logits: torch.Tensor, top_p: float | None) -> torch.Tensor: + """Nucleus (top-p) logit filtering.""" + if top_p is None or top_p >= 1.0: + return logits + if not (0.0 < top_p <= 1.0): + raise ValueError(f"`top_p` must be in (0, 1], got {top_p}.") + + sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1) + sorted_probs = torch.softmax(sorted_logits, dim=-1) + cumulative_probs = sorted_probs.cumsum(dim=-1) + + sorted_indices_to_remove = cumulative_probs > float(top_p) + sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone() + sorted_indices_to_remove[..., 0] = 0 + + sorted_logits = sorted_logits.masked_fill(sorted_indices_to_remove, torch.finfo(sorted_logits.dtype).min) + filtered = logits.scatter(-1, sorted_indices, sorted_logits) + return filtered + + @staticmethod + def _top_k_filtering(logits: torch.Tensor, top_k: int | None) -> torch.Tensor: + """Top-k logit filtering.""" + if top_k is None or top_k <= 0: + return logits + if top_k >= logits.shape[-1]: + return logits + values, _ = torch.topk(logits, k=top_k, dim=-1) + min_keep = values[..., -1, None] + return logits.masked_fill(logits < min_keep, torch.finfo(logits.dtype).min) + + @staticmethod + def _sample_from_logits( + logits: torch.Tensor, + *, + temperature: float, + top_k: int | None, + top_p: float | None, + generator: torch.Generator | None, + use_multinomial: bool, + ) -> tuple[torch.LongTensor, torch.Tensor]: + """Sample tokens from logits with temperature scaling, top-k, and top-p.""" + if temperature < 0: + raise ValueError(f"`temperature` must be >= 0, got {temperature}.") + + vocab_size = logits.shape[-1] + flat_logits = logits.reshape(-1, vocab_size) + + if temperature == 0.0 or not use_multinomial: + probs = torch.softmax(flat_logits.float(), dim=-1) + token = flat_logits.argmax(dim=-1, keepdim=True) + token_prob = torch.gather(probs, -1, token) + return token.view(*logits.shape[:-1]), token_prob.view(*logits.shape[:-1]) + + scaled = flat_logits + if temperature != 1.0: + scaled = flat_logits / temperature + + filtered = BlockRefinementScheduler._top_k_filtering(scaled, top_k=top_k) + filtered = BlockRefinementScheduler._top_p_filtering(filtered, top_p=top_p) + + probs = torch.softmax(filtered.float(), dim=-1) + token = torch.multinomial(probs, num_samples=1, generator=generator) + token_prob = torch.gather(probs, -1, token) + + return token.view(*logits.shape[:-1]), token_prob.view(*logits.shape[:-1]) + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.LongTensor, + *, + mask_token_id: int | None = None, + temperature: float = 0.0, + top_p: float | None = None, + top_k: int | None = None, + sampling_method: str = "auto", + threshold: float | None = None, + editing_threshold: float | None = None, + minimal_topk: int | None = None, + prompt_mask: torch.BoolTensor | None = None, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> ( + BlockRefinementSchedulerOutput + | tuple[torch.LongTensor, torch.BoolTensor, torch.BoolTensor, torch.LongTensor, torch.Tensor] + ): + """ + Perform a single refinement step: sample from logits, commit confident tokens, and optionally edit existing + ones. + + Args: + model_output (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + Raw logits from the model for the current block. + timestep (`int` or `torch.Tensor`): + Current step index within the block's refinement schedule. + sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Current block token IDs (contains mask tokens for uncommitted positions in the mask-based mode). + mask_token_id (`int`, *optional*): + Token ID used for masked positions. When `None`, the scheduler runs in uniform corruption mode: it + tracks committed positions internally (resetting at `timestep == 0`) and renoises the uncommitted ones + with uniformly random tokens, matching DiffusionGemma's block refinement sampler. + temperature (`float`): + Sampling temperature. + top_p (`float`, *optional*): + Nucleus sampling cutoff. + top_k (`int`, *optional*): + Top-k sampling cutoff. + sampling_method (`str`): + Sampling method (`auto`, `greedy`, `multinomial`). + threshold (`float`, *optional*): + Confidence threshold for committing tokens. Defaults to config value. + editing_threshold (`float`, *optional*): + Confidence threshold for editing non-mask tokens; must be positive to enable editing. Defaults to + config value. + minimal_topk (`int`, *optional*): + Minimum tokens to commit per step. Defaults to config value. + prompt_mask (`torch.BoolTensor`, *optional*): + Boolean mask of shape `(block_length,)` where `True` marks prompt (non-editable) positions. + generator (`torch.Generator`, *optional*): + RNG for sampling. + return_dict (`bool`): + Whether to return a `BlockRefinementSchedulerOutput` or a tuple. + """ + if threshold is None: + threshold = float(self.config.threshold) + if editing_threshold is None: + editing_threshold = self.config.editing_threshold + if minimal_topk is None: + minimal_topk = self.config.minimal_topk + + # Sample from logits + use_multinomial = sampling_method == "multinomial" or (sampling_method == "auto" and temperature != 0.0) + sampled_tokens, sampled_probs = self._sample_from_logits( + model_output, + temperature=temperature, + top_k=top_k, + top_p=top_p, + generator=generator, + use_multinomial=use_multinomial, + ) + + batch_size, block_length = sample.shape + + if isinstance(timestep, torch.Tensor): + step_index = int(timestep.item()) + else: + step_index = int(timestep) + + # --- Uniform corruption mode (DiffusionGemma): no mask token, committed positions tracked as state --- + if mask_token_id is None: + if step_index == 0 or self._committed is None or self._committed.shape != sample.shape: + self._committed = torch.zeros_like(sample, dtype=torch.bool) + committed = self._committed + confidence = sampled_probs.to(dtype=torch.float32) + + # Cumulative quota: evenly distribute the block across the steps, commit what is still owed + steps_done = step_index + 1 + target = (steps_done * block_length + self.num_inference_steps - 1) // self.num_inference_steps + needed = (target - committed.sum(dim=-1)).clamp(min=0) + + masked_confidence = confidence.masked_fill(committed, float("-inf")) + ranks = masked_confidence.argsort(dim=-1, descending=True).argsort(dim=-1) + transfer_index = ~committed & ((ranks < needed[:, None]) | (confidence > threshold)) + + editing_transfer_index = torch.zeros_like(transfer_index) + if editing_threshold is not None: + editing_transfer_index = ( + committed & (sampled_tokens != sample) & (confidence > float(editing_threshold)) + ) + + prev_sample = torch.where(transfer_index | editing_transfer_index, sampled_tokens, sample) + self._committed = committed | transfer_index + random_tokens = torch.randint( + low=0, high=model_output.shape[-1], size=sample.shape, device=sample.device, generator=generator + ) + prev_sample = torch.where(self._committed, prev_sample, random_tokens) + + if not return_dict: + return prev_sample, transfer_index, editing_transfer_index, sampled_tokens, sampled_probs, model_output + return BlockRefinementSchedulerOutput( + prev_sample=prev_sample, + transfer_index=transfer_index, + editing_transfer_index=editing_transfer_index, + sampled_tokens=sampled_tokens, + sampled_probs=sampled_probs, + pred_logits=model_output, + ) + + active_block = sample == mask_token_id + masks_remaining = active_block.any() + + # --- Mask-filling transfer --- + transfer_index = torch.zeros_like(sampled_tokens, dtype=torch.bool) + if masks_remaining and self._transfer_schedule is not None: + clamped_step = min(step_index, len(self._transfer_schedule) - 1) + num_to_transfer = int(self._transfer_schedule[clamped_step].item()) + + confidence = torch.where( + active_block, + sampled_probs.to(dtype=torch.float32), + torch.full_like(sampled_probs, -torch.inf, dtype=torch.float32), + ) + + for b in range(batch_size): + high_conf = confidence[b] > threshold + if high_conf.sum().item() >= num_to_transfer: + transfer_index[b] = high_conf + else: + k = min(num_to_transfer, int(active_block[b].sum().item())) + if k > 0: + _, idx = torch.topk(confidence[b], k=k) + transfer_index[b, idx] = True + + # --- Editing transfer (non-mask, non-prompt positions) --- + editing_enabled = editing_threshold is not None and editing_threshold > 0.0 + editing_transfer_index = torch.zeros_like(sampled_tokens, dtype=torch.bool) + if editing_enabled: + if prompt_mask is None: + prompt_mask = torch.zeros(block_length, device=sample.device, dtype=torch.bool) + editable = (~active_block) & (~prompt_mask.unsqueeze(0)) + editing_conf = torch.where( + editable, + sampled_probs.to(dtype=torch.float32), + torch.full_like(sampled_probs, -torch.inf, dtype=torch.float32), + ) + high_conf_edit = editing_conf > float(editing_threshold) + token_changed = sampled_tokens != sample + editing_transfer_index = high_conf_edit & token_changed & editable + + # Apply transfers + final_transfer = transfer_index | editing_transfer_index + prev_sample = sample.clone() + if final_transfer.any(): + prev_sample[final_transfer] = sampled_tokens[final_transfer] + + if not return_dict: + return prev_sample, transfer_index, editing_transfer_index, sampled_tokens, sampled_probs, model_output + return BlockRefinementSchedulerOutput( + prev_sample=prev_sample, + transfer_index=transfer_index, + editing_transfer_index=editing_transfer_index, + sampled_tokens=sampled_tokens, + sampled_probs=sampled_probs, + pred_logits=model_output, + ) + + @staticmethod + def check_eos_finished( + cur_x: torch.LongTensor, + sampled_tokens: torch.LongTensor, + final_transfer: torch.BoolTensor, + finished: torch.BoolTensor, + eos_token_id: int, + mask_token_id: int, + prompt_length: int, + ) -> torch.BoolTensor: + """ + Update per-batch finished flags when EOS tokens are committed. + + Args: + cur_x (`torch.LongTensor` of shape `(batch_size, seq_len)`): + Current full sequence including all blocks up to the current window. + sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`): + Tokens sampled by the scheduler in this step. + final_transfer (`torch.BoolTensor` of shape `(batch_size, block_length)`): + Combined mask of committed and edited positions. + finished (`torch.BoolTensor` of shape `(batch_size,)`): + Current per-batch finished flags. + eos_token_id (`int`): + EOS token ID. + mask_token_id (`int`): + Mask token ID. + prompt_length (`int`): + Number of prompt tokens at the start of the sequence. + + Returns: + `torch.BoolTensor`: Updated finished flags. + """ + batch_size = cur_x.shape[0] + for b in range(batch_size): + if finished[b]: + continue + eos_in_commits = (sampled_tokens[b][final_transfer[b]] == eos_token_id).any().item() + if not eos_in_commits: + continue + eos_pos = (cur_x[b] == eos_token_id).nonzero(as_tuple=True) + if len(eos_pos[0]) == 0: + continue + eos_pos = int(eos_pos[0][0].item()) + # The first generated token sits at index `prompt_length`; allow EOS there. + if eos_pos < prompt_length: + continue + if (cur_x[b, prompt_length:eos_pos] != mask_token_id).all().item(): + finished[b] = True + return finished + + def check_block_should_continue( + self, + step_idx: int, + masks_remaining: bool, + editing_enabled: bool, + editing_transfer_index: torch.BoolTensor, + post_steps: int, + max_post_steps: int, + finished: torch.BoolTensor, + ) -> bool: + """ + Determine whether the inner refinement loop should continue for the current block. + + Args: + step_idx (`int`): + Current refinement step index within this block. + masks_remaining (`bool`): + Whether any mask tokens remain in the block. + editing_enabled (`bool`): + Whether editing mode is active. + editing_transfer_index (`torch.BoolTensor`): + Which tokens were edited in this step. + post_steps (`int`): + Number of post-mask editing steps taken so far. + max_post_steps (`int`): + Maximum allowed post-mask editing steps. + finished (`torch.BoolTensor`): + Per-batch finished flags (from EOS detection). + + Returns: + `bool`: `True` if refinement should continue, `False` to break. + """ + if finished.all(): + return False + if not masks_remaining and not editing_enabled: + return False + if not masks_remaining and not editing_transfer_index.any(): + return False + if masks_remaining and step_idx >= self.num_inference_steps: + return False + if not masks_remaining and post_steps > max_post_steps: + return False + return True + + def add_noise( + self, + original_samples: torch.LongTensor, + attention_mask: torch.LongTensor, + *, + prompt_length: int, + block_length: int, + mask_token_id: int, + generator: torch.Generator | None = None, + ) -> tuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor, torch.BoolTensor]: + """ + Apply the forward (noising) process for semi-autoregressive block masking. + + For each block after the prompt, a random fraction of valid (non-padding) tokens are replaced with + `mask_token_id`. Two complementary views are returned: `noisy` and `noisy_rev`, where the masked positions in + one are the unmasked positions in the other. + + Args: + original_samples (`torch.LongTensor` of shape `(batch_size, seq_len)`): + Clean token IDs. + attention_mask (`torch.LongTensor` of shape `(batch_size, seq_len)`): + Padding mask (1 for valid, 0 for padding). + prompt_length (`int`): + Number of leading prompt tokens to keep unmasked. + block_length (`int`): + Block size for masking. + mask_token_id (`int`): + Token ID to use for masked positions. + generator (`torch.Generator`, *optional*): + RNG for reproducibility. + + Returns: + `tuple[torch.LongTensor, torch.LongTensor, torch.BoolTensor, torch.BoolTensor]`: + `(noisy, noisy_rev, masked, masked_rev)` — the two complementary noisy sequences and their + corresponding boolean masks. + """ + batch_size, seq_len = original_samples.shape + device = original_samples.device + + noisy = original_samples.clone() + noisy_rev = original_samples.clone() + masked = torch.zeros_like(original_samples, dtype=torch.bool) + masked_rev = torch.zeros_like(original_samples, dtype=torch.bool) + + valid = attention_mask.to(dtype=torch.bool) + for block_start in range(prompt_length, seq_len, block_length): + block_end = min(seq_len, block_start + block_length) + seg_len = block_end - block_start + if seg_len <= 0: + continue + + p_mask = torch.rand((batch_size, 1), device=device, generator=generator) + seg = torch.rand((batch_size, seg_len), device=device, generator=generator) < p_mask + seg = seg & valid[:, block_start:block_end] + seg_rev = (~seg) & valid[:, block_start:block_end] + + masked[:, block_start:block_end] = seg + masked_rev[:, block_start:block_end] = seg_rev + + noisy = torch.where(masked, torch.full_like(noisy, mask_token_id), noisy) + noisy_rev = torch.where(masked_rev, torch.full_like(noisy_rev, mask_token_id), noisy_rev) + return noisy, noisy_rev, masked, masked_rev + + +__all__ = ["BlockRefinementScheduler", "BlockRefinementSchedulerOutput"] diff --git a/diffusers/schedulers/scheduling_consistency_decoder.py b/diffusers/schedulers/scheduling_consistency_decoder.py new file mode 100644 index 0000000000000000000000000000000000000000..7f5e0cad22b9b9b2b3920c8c25eb77d9acc7fd36 --- /dev/null +++ b/diffusers/schedulers/scheduling_consistency_decoder.py @@ -0,0 +1,212 @@ +import math +from dataclasses import dataclass +from typing import Literal + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +@dataclass +class ConsistencyDecoderSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class ConsistencyDecoderScheduler(SchedulerMixin, ConfigMixin): + """ + A scheduler for the consistency decoder used in Stable Diffusion pipelines. + + This scheduler implements a two-step denoising process using consistency models for decoding latent representations + into images. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, *optional*, defaults to `1024`): + The number of diffusion steps to train the model. + sigma_data (`float`, *optional*, defaults to `0.5`): + The standard deviation of the data distribution. Used for computing the skip and output scaling factors. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1024, + sigma_data: float = 0.5, + ) -> None: + betas = betas_for_alpha_bar(num_train_timesteps) + + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + + self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod) + self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod) + + sigmas = torch.sqrt(1.0 / alphas_cumprod - 1) + + sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod) + + self.c_skip = sqrt_recip_alphas_cumprod * sigma_data**2 / (sigmas**2 + sigma_data**2) + self.c_out = sigmas * sigma_data / (sigmas**2 + sigma_data**2) ** 0.5 + self.c_in = sqrt_recip_alphas_cumprod / (sigmas**2 + sigma_data**2) ** 0.5 + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device = None, + ): + if num_inference_steps != 2: + raise ValueError("Currently more than 2 inference steps are not supported.") + + self.timesteps = torch.tensor([1008, 512], dtype=torch.long, device=device) + self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.to(device) + self.sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod.to(device) + self.c_skip = self.c_skip.to(device) + self.c_out = self.c_out.to(device) + self.c_in = self.c_in.to(device) + + @property + def init_noise_sigma(self) -> torch.Tensor: + """ + Return the standard deviation of the initial noise distribution. + + Returns: + `torch.Tensor`: + The initial noise sigma value from the precomputed `sqrt_one_minus_alphas_cumprod` at the first + timestep. + """ + return self.sqrt_one_minus_alphas_cumprod[self.timesteps[0]] + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample * self.c_in[timestep] + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> ConsistencyDecoderSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator for reproducibility. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_consistency_decoder.ConsistencyDecoderSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_consistency_decoder.ConsistencyDecoderSchedulerOutput`] or `tuple`: + If `return_dict` is `True`, + [`~schedulers.scheduling_consistency_decoder.ConsistencyDecoderSchedulerOutput`] is returned, otherwise + a tuple is returned where the first element is the sample tensor. + """ + x_0 = self.c_out[timestep] * model_output + self.c_skip[timestep] * sample + + timestep_idx = torch.where(self.timesteps == timestep)[0] + + if timestep_idx == len(self.timesteps) - 1: + prev_sample = x_0 + else: + noise = randn_tensor(x_0.shape, generator=generator, dtype=x_0.dtype, device=x_0.device) + prev_sample = ( + self.sqrt_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * x_0 + + self.sqrt_one_minus_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * noise + ) + + if not return_dict: + return (prev_sample,) + + return ConsistencyDecoderSchedulerOutput(prev_sample=prev_sample) diff --git a/diffusers/schedulers/scheduling_consistency_models.py b/diffusers/schedulers/scheduling_consistency_models.py new file mode 100644 index 0000000000000000000000000000000000000000..cc79972e8cee4c994fdc505f13f7eb896da265eb --- /dev/null +++ b/diffusers/schedulers/scheduling_consistency_models.py @@ -0,0 +1,517 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class CMStochasticIterativeSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class CMStochasticIterativeScheduler(SchedulerMixin, ConfigMixin): + """ + Multistep and onestep sampling for consistency models. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 40): + The number of diffusion steps to train the model. + sigma_min (`float`, defaults to 0.002): + Minimum noise magnitude in the sigma schedule. Defaults to 0.002 from the original implementation. + sigma_max (`float`, defaults to 80.0): + Maximum noise magnitude in the sigma schedule. Defaults to 80.0 from the original implementation. + sigma_data (`float`, defaults to 0.5): + The standard deviation of the data distribution from the EDM + [paper](https://huggingface.co/papers/2206.00364). Defaults to 0.5 from the original implementation. + s_noise (`float`, defaults to 1.0): + The amount of additional noise to counteract loss of detail during sampling. A reasonable range is [1.000, + 1.011]. Defaults to 1.0 from the original implementation. + rho (`float`, defaults to 7.0): + The parameter for calculating the Karras sigma schedule from the EDM + [paper](https://huggingface.co/papers/2206.00364). Defaults to 7.0 from the original implementation. + clip_denoised (`bool`, defaults to `True`): + Whether to clip the denoised outputs to `(-1, 1)`. + timesteps (`list` or `np.ndarray` or `torch.Tensor`, *optional*): + An explicit timestep schedule that can be optionally specified. The timesteps are expected to be in + increasing order. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 40, + sigma_min: float = 0.002, + sigma_max: float = 80.0, + sigma_data: float = 0.5, + s_noise: float = 1.0, + rho: float = 7.0, + clip_denoised: bool = True, + ) -> None: + # standard deviation of the initial noise distribution + self.init_noise_sigma = sigma_max + + ramp = np.linspace(0, 1, num_train_timesteps) + sigmas = self._convert_to_karras(ramp) + timesteps = self.sigma_to_t(sigmas) + + # setable values + self.num_inference_steps = None + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps) + self.custom_timesteps = False + self.is_scale_input_called = False + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not yet initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index, or `None` if not yet set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Scales the consistency model input by `(sigma**2 + sigma_data**2) ** 0.5`. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + # Get sigma corresponding to timestep + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + sample = sample / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + + self.is_scale_input_called = True + return sample + + def sigma_to_t(self, sigmas: float | np.ndarray): + """ + Gets scaled timesteps from the Karras sigmas for input to the consistency model. + + Args: + sigmas (`float` or `np.ndarray`): + A single Karras sigma or an array of Karras sigmas. + + Returns: + `np.ndarray`: + A scaled input timestep array. + """ + if not isinstance(sigmas, np.ndarray): + sigmas = np.array(sigmas, dtype=np.float64) + + timesteps = 1000 * 0.25 * np.log(sigmas + 1e-44) + + return timesteps + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device = None, + timesteps: list[int] | None = None, + ): + """ + Sets the timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, + `num_inference_steps` must be `None`. + """ + if num_inference_steps is None and timesteps is None: + raise ValueError("Exactly one of `num_inference_steps` or `timesteps` must be supplied.") + + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `timesteps`.") + + # Follow DDPMScheduler custom timesteps logic + if timesteps is not None: + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`timesteps` must be in descending order.") + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.custom_timesteps = True + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + self.custom_timesteps = False + + # Map timesteps to Karras sigmas directly for multistep sampling + # See https://github.com/openai/consistency_models/blob/main/cm/karras_diffusion.py#L675 + num_train_timesteps = self.config.num_train_timesteps + ramp = timesteps[::-1].copy() + ramp = ramp / (num_train_timesteps - 1) + sigmas = self._convert_to_karras(ramp) + timesteps = self.sigma_to_t(sigmas) + + sigmas = np.concatenate([sigmas, [self.config.sigma_min]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas).to(device=device) + + if str(device).startswith("mps"): + # mps does not support float64 + self.timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32) + else: + self.timesteps = torch.from_numpy(timesteps).to(device=device) + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Modified _convert_to_karras implementation that takes in ramp as argument + def _convert_to_karras(self, ramp: np.ndarray) -> np.ndarray: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + ramp (`np.ndarray`): + A ramp array of values between 0 and 1 used to interpolate between sigma_min and sigma_max. + + Returns: + `np.ndarray`: + The Karras sigma schedule array. + """ + sigma_min: float = self.config.sigma_min + sigma_max: float = self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + def get_scalings(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Computes the scaling factors for the consistency model output. + + Args: + sigma (`torch.Tensor`): + The current sigma value in the noise schedule. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing `c_skip` (scaling for the input sample) and `c_out` (scaling for the model output). + """ + sigma_data = self.config.sigma_data + + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + return c_skip, c_out + + def get_scalings_for_boundary_condition(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Gets the scalings used in the consistency model parameterization (from Appendix C of the + [paper](https://huggingface.co/papers/2303.01469)) to enforce boundary condition. + + > [!TIP] > `epsilon` in the equations for `c_skip` and `c_out` is set to `sigma_min`. + + Args: + sigma (`torch.Tensor`): + The current sigma in the Karras sigma schedule. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A two-element tuple where `c_skip` (which weights the current sample) is the first element and `c_out` + (which weights the consistency model output) is the second element. + """ + sigma_min = self.config.sigma_min + sigma_data = self.config.sigma_data + + c_skip = sigma_data**2 / ((sigma - sigma_min) ** 2 + sigma_data**2) + c_out = (sigma - sigma_min) * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + return c_skip, c_out + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> CMStochasticIterativeSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + f" `{self.__class__}.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + sigma_min = self.config.sigma_min + sigma_max = self.config.sigma_max + + if self.step_index is None: + self._init_step_index(timestep) + + # sigma_next corresponds to next_t in original implementation + sigma = self.sigmas[self.step_index] + if self.step_index + 1 < self.config.num_train_timesteps: + sigma_next = self.sigmas[self.step_index + 1] + else: + # Set sigma_next to sigma_min + sigma_next = self.sigmas[-1] + + # Get scalings for boundary conditions + c_skip, c_out = self.get_scalings_for_boundary_condition(sigma) + + # 1. Denoise model output using boundary conditions + denoised = c_out * model_output + c_skip * sample + if self.config.clip_denoised: + denoised = denoised.clamp(-1, 1) + + # 2. Sample z ~ N(0, s_noise^2 * I) + # Noise is not used for onestep sampling. + if len(self.timesteps) > 1: + noise = randn_tensor( + model_output.shape, + dtype=model_output.dtype, + device=model_output.device, + generator=generator, + ) + else: + noise = torch.zeros_like(model_output) + z = noise * self.config.s_noise + + sigma_hat = sigma_next.clamp(min=sigma_min, max=sigma_max) + + # 3. Return noisy sample + # tau = sigma_hat, eps = sigma_min + prev_sample = denoised + z * (sigma_hat**2 - sigma_min**2) ** 0.5 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return CMStochasticIterativeSchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + """ + Returns the number of training timesteps. + + Returns: + `int`: + The number of training timesteps configured for the scheduler. + """ + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py b/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..18ee272ef6193c9d5c9bfb09f3af7a7cc281449c --- /dev/null +++ b/diffusers/schedulers/scheduling_cosine_dpmsolver_multistep.py @@ -0,0 +1,755 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_dpmsolver_sde import BrownianTreeNoiseSampler +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +class CosineDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + Implements a variant of `DPMSolverMultistepScheduler` with cosine schedule, proposed by Nichol and Dhariwal (2021). + This scheduler was used in Stable Audio Open [1]. + + [1] Evans, Parker, et al. "Stable Audio Open" https://huggingface.co/papers/2407.14358 + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + sigma_min (`float`, defaults to `0.3`): + Minimum noise magnitude in the sigma schedule. This was set to 0.3 in Stable Audio Open [1]. + sigma_max (`float`, defaults to `500`): + Maximum noise magnitude in the sigma schedule. This was set to 500 in Stable Audio Open [1]. + sigma_data (`float`, defaults to `1.0`): + The standard deviation of the data distribution. This is set to 1.0 in Stable Audio Open [1]. + sigma_schedule (`str`, defaults to `"exponential"`): + Sigma schedule to compute the `sigmas`. Must be one of `"exponential"` or `"karras"`. The exponential + schedule was incorporated in [stabilityai/cosxl](https://huggingface.co/stabilityai/cosxl). The Karras + schedule is introduced in the [EDM](https://huggingface.co/papers/2206.00364) paper. + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + solver_order (`int`, defaults to `2`): + The DPMSolver order which can be `1` or `2`. It is recommended to use `solver_order=2`. + prediction_type (`str`, defaults to `"v_prediction"`): + Prediction type of the scheduler function. Must be one of `"epsilon"` (predicts the noise of the diffusion + process), `"sample"` (directly predicts the noisy sample), or `"v_prediction"` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + rho (`float`, defaults to `7.0`): + The parameter for calculating the Karras sigma schedule from the EDM + [paper](https://huggingface.co/papers/2206.00364). + solver_type (`str`, defaults to `"midpoint"`): + Solver type for the second-order solver. Must be one of `"midpoint"` or `"heun"`. The solver type slightly + affects the sample quality, especially for a small number of steps. It is recommended to use `"midpoint"`. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. Must be one of `"zero"` or + `"sigma_min"`. If `"sigma_min"`, the final sigma is the same as the last sigma in the training schedule. If + `"zero"`, the final sigma is set to 0. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.3, + sigma_max: float = 500, + sigma_data: float = 1.0, + sigma_schedule: Literal["exponential", "karras"] = "exponential", + num_train_timesteps: int = 1000, + solver_order: int = 2, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "v_prediction", + rho: float = 7.0, + solver_type: Literal["midpoint", "heun"] = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", + ) -> None: + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + ramp = torch.linspace(0, 1, num_train_timesteps) + if sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + self.timesteps = self.precondition_noise(sigmas) + + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + # setable values + self.num_inference_steps = None + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> float: + """ + The standard deviation of the initial noise distribution. + + Returns: + `float`: + The initial noise sigma value computed as `sqrt(sigma_max^2 + 1)`. + """ + return (self.config.sigma_max**2 + 1) ** 0.5 + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not yet initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index, or `None` if not yet set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs + def precondition_inputs(self, sample: torch.Tensor, sigma: float | torch.Tensor) -> torch.Tensor: + """ + Precondition the input sample by scaling it according to the EDM formulation. + + Args: + sample (`torch.Tensor`): + The input sample tensor to precondition. + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `torch.Tensor`: + The scaled input sample. + """ + c_in = self._get_conditioning_c_in(sigma) + scaled_sample = sample * c_in + return scaled_sample + + def precondition_noise(self, sigma: float | torch.Tensor) -> torch.Tensor: + """ + Precondition the noise level by computing a normalized timestep representation. + + Args: + sigma (`float` or `torch.Tensor`): + The sigma (noise level) value to precondition. + + Returns: + `torch.Tensor`: + The preconditioned noise value computed as `atan(sigma) / pi * 2`. + """ + if not isinstance(sigma, torch.Tensor): + sigma = torch.tensor([sigma]) + + return sigma.atan() / math.pi * 2 + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs + def precondition_outputs( + self, + sample: torch.Tensor, + model_output: torch.Tensor, + sigma: float | torch.Tensor, + ) -> torch.Tensor: + """ + Precondition the model outputs according to the EDM formulation. + + Args: + sample (`torch.Tensor`): + The input sample tensor. + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `torch.Tensor`: + The denoised sample computed by combining the skip connection and output scaling. + """ + sigma_data = self.config.sigma_data + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + + if self.config.prediction_type == "epsilon": + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + elif self.config.prediction_type == "v_prediction": + c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + else: + raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.") + + denoised = c_skip * sample + c_out * model_output + + return denoised + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Scale the denoising model input to match the Euler algorithm. Ensures interchangeability with schedulers that + need to scale the denoising model input depending on the current timestep. + + Args: + sample (`torch.Tensor`): + The input sample tensor. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = self.precondition_inputs(sample, sigma) + + self.is_scale_input_called = True + return sample + + def set_timesteps(self, num_inference_steps: int | None = None, device: str | torch.device | None = None) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + + ramp = torch.linspace(0, 1, self.num_inference_steps) + if self.config.sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif self.config.sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + sigmas = sigmas.to(dtype=torch.float32, device=device) + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = self.config.sigma_min + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.tensor([sigma_last], dtype=torch.float32, device=device)]) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # if a noise sampler is used, reinitialise it + self.noise_sampler = None + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_karras_sigmas + def _compute_karras_sigmas( + self, + ramp: torch.Tensor, + sigma_min: float | None = None, + sigma_max: float | None = None, + ) -> torch.Tensor: + """ + Construct the noise schedule of [Karras et al. (2022)](https://huggingface.co/papers/2206.00364). + + Args: + ramp (`torch.Tensor`): + A tensor of values in [0, 1] representing the interpolation positions. + sigma_min (`float`, *optional*): + Minimum sigma value. If `None`, uses `self.config.sigma_min`. + sigma_max (`float`, *optional*): + Maximum sigma value. If `None`, uses `self.config.sigma_max`. + + Returns: + `torch.Tensor`: + The computed Karras sigma schedule. + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_exponential_sigmas + def _compute_exponential_sigmas( + self, + ramp: torch.Tensor, + sigma_min: float | None = None, + sigma_max: float | None = None, + ) -> torch.Tensor: + """ + Compute the exponential sigma schedule. Implementation closely follows k-diffusion: + https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 + + Args: + ramp (`torch.Tensor`): + A tensor of values representing the interpolation positions. + sigma_min (`float`, *optional*): + Minimum sigma value. If `None`, uses `self.config.sigma_min`. + sigma_max (`float`, *optional*): + Maximum sigma value. If `None`, uses `self.config.sigma_max`. + + Returns: + `torch.Tensor`: + The computed exponential sigma schedule. + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), len(ramp)).exp().flip(0) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Convert sigma to alpha and sigma_t values for the diffusion process. + + Args: + sigma (`torch.Tensor`): + The sigma (noise level) value. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing `alpha_t` (always 1 since inputs are pre-scaled) and `sigma_t` (same as input + sigma). + """ + alpha_t = torch.tensor(1) # Inputs are pre-scaled before going into unet, so alpha_t = 1 + sigma_t = sigma + + return alpha_t, sigma_t + + def convert_model_output( + self, + model_output: torch.Tensor, + sample: torch.Tensor = None, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both + noise > prediction and data prediction models. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + sigma = self.sigmas[self.step_index] + x0_pred = self.precondition_outputs(sample, model_output, sigma) + + return x0_pred + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + sample: torch.Tensor = None, + noise: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + def multistep_dpm_solver_second_order_update( + self, + model_output_list: list[torch.Tensor], + sample: torch.Tensor = None, + noise: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + + # sde-dpmsolver++ + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep: int | torch.Tensor) -> None: + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final + or (self.config.lower_order_final and len(self.timesteps) < 15) + or self.config.final_sigmas_type == "zero" + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.noise_sampler is None: + seed = None + if generator is not None: + seed = ( + [g.initial_seed() for g in generator] if isinstance(generator, list) else generator.initial_seed() + ) + self.noise_sampler = BrownianTreeNoiseSampler( + model_output, + sigma_min=self.config.sigma_min, + sigma_max=self.config.sigma_max, + seed=seed, + ) + noise = self.noise_sampler(self.sigmas[self.step_index], self.sigmas[self.step_index + 1]).to( + model_output.device + ) + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in + def _get_conditioning_c_in(self, sigma: float | torch.Tensor) -> float | torch.Tensor: + """ + Compute the input conditioning factor for the EDM formulation. + + Args: + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `float` or `torch.Tensor`: + The input conditioning factor `c_in`. + """ + c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + return c_in + + def __len__(self) -> int: + """ + Returns the number of training timesteps. + + Returns: + `int`: + The number of training timesteps configured for the scheduler. + """ + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ddim.py b/diffusers/schedulers/scheduling_ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..d29bb5b2593cf99535d608e91af0a8d33a5c3e44 --- /dev/null +++ b/diffusers/schedulers/scheduling_ddim.py @@ -0,0 +1,595 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDIMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`Literal["linear", "scaled_linear", "squaredcos_cap_v2"]`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Must be one + of `"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`Literal["epsilon", "sample", "v_prediction"]`, defaults to `"epsilon"`): + Prediction type of the scheduler function. Must be one of `"epsilon"` (predicts the noise of the diffusion + process), `"sample"` (directly predicts the noisy sample), or `"v_prediction"` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`Literal["leading", "trailing", "linspace"]`, defaults to `"leading"`): + The way the timesteps should be scaled. Must be one of `"leading"`, `"trailing"`, or `"linspace"`. Refer to + Table 2 of the [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: Literal["leading", "trailing", "linspace"] = "leading", + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def scale_model_input(self, sample: torch.Tensor, timestep: int = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_variance(self, timestep: int, prev_timestep: int) -> torch.Tensor: + """ + Computes the variance of the noise added at a given diffusion step. + + For a given `timestep` and its previous step, this method calculates the variance as defined in DDIM/DDPM + literature: + var_t = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + where alpha_prod and beta_prod are cumulative products of alphas and betas, respectively. + + Args: + timestep (`int`): + The current timestep in the diffusion process. + prev_timestep (`int`): + The previous timestep in the diffusion process. If negative, uses `final_alpha_cumprod`. + + Returns: + `torch.Tensor`: + The variance for the current timestep. + """ + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str | torch.device`, *optional*): + The device to use for the timesteps. + + Raises: + ValueError: If `num_inference_steps` is larger than `self.config.num_train_timesteps`. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator: torch.Generator | None = None, + variance_noise: torch.Tensor | None = None, + return_dict: bool = True, + ) -> DDIMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`, *optional*, defaults to 0.0): + The weight of noise for added noise in diffusion step. A value of 0 corresponds to DDIM (deterministic) + and 1 corresponds to DDPM (fully stochastic). + use_clipped_model_output (`bool`, *optional*, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible sampling. + variance_noise (`torch.Tensor`, *optional*): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the pred_epsilon is always re-derived from the clipped x_0 in Glide + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if eta > 0: + if variance_noise is not None and generator is not None: + raise ValueError( + "Cannot pass both generator and variance_noise. Please make sure that either `generator` or" + " `variance_noise` stays `None`." + ) + + if variance_noise is None: + variance_noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + variance = std_dev_t * variance_noise + + prev_sample = prev_sample + variance + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ddim_cogvideox.py b/diffusers/schedulers/scheduling_ddim_cogvideox.py new file mode 100644 index 0000000000000000000000000000000000000000..11cbb8a6b9f7e372005c2ae460ac4676362fcc28 --- /dev/null +++ b/diffusers/schedulers/scheduling_ddim_cogvideox.py @@ -0,0 +1,504 @@ +# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(alphas_cumprod: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on (Algorithm 1)[https://huggingface.co/papers/2305.08891] + + Args: + alphas_cumprod (`torch.Tensor`): + The alphas cumulative products that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + + return alphas_bar + + +class CogVideoXDDIMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.00085): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.0120): + The final `beta` value. + beta_schedule (`str`, defaults to `"scaled_linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + snr_shift_scale (`float`, defaults to 3.0): + Shift scale for SNR. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.0120, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "scaled_linear", + trained_betas: np.ndarray | list[float] | None = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "leading", + rescale_betas_zero_snr: bool = False, + snr_shift_scale: float = 3.0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float64, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # Modify: SNR shift following SD3 + self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod) + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def _get_variance(self, timestep: int, prev_timestep: int) -> torch.Tensor: + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def scale_model_input(self, sample: torch.Tensor, timestep: int = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator: torch.Generator | None = None, + variance_noise: torch.Tensor | None = None, + return_dict: bool = True, + ) -> DDIMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + The weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + # To make style tests pass, commented out `pred_epsilon` as it is an unused variable + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + # pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + # pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + # pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + a_t = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 + b_t = alpha_prod_t_prev**0.5 - alpha_prod_t**0.5 * a_t + + prev_sample = a_t * sample + b_t * pred_original_sample + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ddim_inverse.py b/diffusers/schedulers/scheduling_ddim_inverse.py new file mode 100644 index 0000000000000000000000000000000000000000..cd1435090e1ccc28869cd854afff08bc697f28dd --- /dev/null +++ b/diffusers/schedulers/scheduling_ddim_inverse.py @@ -0,0 +1,385 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.schedulers.scheduling_utils import SchedulerMixin +from diffusers.utils import BaseOutput, deprecate + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDIMInverseScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMInverseScheduler` is the reverse scheduler of [`DDIMScheduler`]. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to 0, otherwise + it uses the alpha value at step `num_train_timesteps - 1`. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + order = 1 + ignore_for_config = ["kwargs"] + _deprecated_kwargs = ["set_alpha_to_zero"] + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + clip_sample_range: float = 1.0, + timestep_spacing: Literal["leading", "trailing"] = "leading", + rescale_betas_zero_snr: bool = False, + **kwargs, + ): + if kwargs.get("set_alpha_to_zero", None) is not None: + deprecation_message = ( + "The `set_alpha_to_zero` argument is deprecated. Please use `set_alpha_to_one` instead." + ) + deprecate("set_alpha_to_zero", "1.0.0", deprecation_message, standard_warn=False) + set_alpha_to_one = kwargs["set_alpha_to_zero"] + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in inverted ddim, we are looking into the next alphas_cumprod + # For the initial step, there is no current alphas_cumprod, and the index is out of bounds + # `set_alpha_to_one` decides whether we set this parameter simply to one + # in this case, self.step() just output the predicted noise + # or whether we use the initial alpha used in training the diffusion model. + self.initial_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps).copy().astype(np.int64)) + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "leading" and "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round().copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)[::-1]).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> DDIMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or + `tuple`. + + Returns: + [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + + """ + # 1. get previous step value (=t+1) + prev_timestep = timestep + timestep = min( + timestep - self.config.num_train_timesteps // self.num_inference_steps, + self.config.num_train_timesteps - 1, + ) + + # 2. compute alphas, betas + # change original implementation to exactly match noise levels for analogous forward process + alpha_prod_t = self.alphas_cumprod[timestep] if timestep >= 0 else self.initial_alpha_cumprod + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * pred_epsilon + + # 6. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if not return_dict: + return (prev_sample, pred_original_sample) + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ddim_parallel.py b/diffusers/schedulers/scheduling_ddim_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..a5bbac60275c00b50aef45204ad86507e9ae3bad --- /dev/null +++ b/diffusers/schedulers/scheduling_ddim_parallel.py @@ -0,0 +1,705 @@ +# Copyright 2025 ParaDiGMS authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput +class DDIMParallelSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDIMParallelScheduler(SchedulerMixin, ConfigMixin): + """ + Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising + diffusion probabilistic models (DDPMs) with non-Markovian guidance. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2010.02502 + + Args: + num_train_timesteps (`int`): number of diffusion steps used to train the model. + beta_start (`float`): the starting `beta` value of inference. + beta_end (`float`): the final `beta` value. + beta_schedule (`str`): + the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, optional): + option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + clip_sample (`bool`, default `True`): + option to clip predicted sample for numerical stability. + clip_sample_range (`float`, default `1.0`): + the maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, default `True`): + each diffusion step uses the value of alphas product at that step and at the previous one. For the final + step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the value of alpha at step 0. + steps_offset (`int`, default `0`): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, default `epsilon`, optional): + prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://huggingface.co/papers/2210.02303) + thresholding (`bool`, default `False`): + whether to use the "dynamic thresholding" method (introduced by Imagen, + https://huggingface.co/papers/2205.11487). Note that the thresholding method is unsuitable for latent-space + diffusion models (such as stable-diffusion). + dynamic_thresholding_ratio (`float`, default `0.995`): + the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen + (https://huggingface.co/papers/2205.11487). Valid only when `thresholding=True`. + sample_max_value (`float`, default `1.0`): + the threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, default `"leading"`): + The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample + Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + rescale_betas_zero_snr (`bool`, default `False`): + whether to rescale the betas to have zero terminal SNR (proposed by + https://huggingface.co/papers/2305.08891). This can enable the model to generate very bright and dark + samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + _is_ode_scheduler = True + + @register_to_config + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.__init__ + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: Literal["leading", "trailing", "linspace"] = "leading", + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_variance(self, timestep: int, prev_timestep: int | None = None) -> torch.Tensor: + if prev_timestep is None: + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def _batch_get_variance(self, t: torch.Tensor, prev_t: torch.Tensor) -> torch.Tensor: + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)] + alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0) + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.set_timesteps + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str | torch.device`, *optional*): + The device to use for the timesteps. + + Raises: + ValueError: If `num_inference_steps` is larger than `self.config.num_train_timesteps`. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator: torch.Generator | None = None, + variance_noise: torch.Tensor | None = None, + return_dict: bool = True, + ) -> DDIMParallelSchedulerOutput | tuple: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + eta (`float`): weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, compute "corrected" `model_output` from the clipped predicted original sample. This + correction is necessary because the predicted original sample is clipped to [-1, 1] when + `self.config.clip_sample` is `True`. If no clipping occurred, the "corrected" `model_output` matches + the input and `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + Random number generator. + variance_noise (`torch.Tensor`): instead of generating noise for the variance using `generator`, we + can directly provide the noise for the variance itself. This is useful for methods such as + CycleDiffusion. (https://huggingface.co/papers/2210.05559) + return_dict (`bool`): option for returning tuple rather than DDIMParallelSchedulerOutput class + + Returns: + [`~schedulers.scheduling_utils.DDIMParallelSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.DDIMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. + When returning a tuple, the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._get_variance(timestep, prev_timestep) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the pred_epsilon is always re-derived from the clipped x_0 in Glide + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + if eta > 0: + if variance_noise is not None and generator is not None: + raise ValueError( + "Cannot pass both generator and variance_noise. Please make sure that either `generator` or" + " `variance_noise` stays `None`." + ) + + if variance_noise is None: + variance_noise = randn_tensor( + model_output.shape, + generator=generator, + device=model_output.device, + dtype=model_output.dtype, + ) + variance = std_dev_t * variance_noise + + prev_sample = prev_sample + variance + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DDIMParallelSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def batch_step_no_noise( + self, + model_output: torch.Tensor, + timesteps: torch.Tensor, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + ) -> torch.Tensor: + """ + Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once. + Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise + is pre-sampled by the pipeline. + + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timesteps (`torch.Tensor`): + current discrete timesteps in the diffusion chain. This is now a list of integers. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + eta (`float`): weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped + predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when + `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would + coincide with the one provided as input and `use_clipped_model_output` will have not effect. + + Returns: + `torch.Tensor`: sample tensor at previous timestep. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + assert eta == 0.0 + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + t = timesteps + prev_t = t - self.config.num_train_timesteps // self.num_inference_steps + + t = t.view(-1, *([1] * (model_output.ndim - 1))) + prev_t = prev_t.view(-1, *([1] * (model_output.ndim - 1))) + + # 1. compute alphas, betas + self.alphas_cumprod = self.alphas_cumprod.to(model_output.device) + self.final_alpha_cumprod = self.final_alpha_cumprod.to(model_output.device) + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)] + alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0) + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + # 4. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 5. compute variance: "sigma_t(η)" -> see formula (16) + # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) + variance = self._batch_get_variance(t, prev_t).to(model_output.device).view(*alpha_prod_t_prev.shape) + std_dev_t = eta * variance ** (0.5) + + if use_clipped_model_output: + # the pred_epsilon is always re-derived from the clipped x_0 in Glide + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + + # 6. compute "direction pointing to x_t" of formula (12) from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon + + # 7. compute x_t without "random noise" of formula (12) from https://huggingface.co/papers/2010.02502 + prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ddpm.py b/diffusers/schedulers/scheduling_ddpm.py new file mode 100644 index 0000000000000000000000000000000000000000..972c46c6e930119886ee66db455588ef5de603bb --- /dev/null +++ b/diffusers/schedulers/scheduling_ddpm.py @@ -0,0 +1,668 @@ +# Copyright 2025 UC Berkeley Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +class DDPMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDPMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDPMScheduler` explores the connections between denoising score matching and Langevin dynamics sampling. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.0001`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.02`): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, `"squaredcos_cap_v2"`, or `"sigmoid"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. + trained_betas (`np.ndarray`, *optional*): + An array of betas to pass directly to the constructor without using `beta_start` and `beta_end`. + variance_type (`"fixed_small"`, `"fixed_small_log"`, `"fixed_large"`, `"fixed_large_log"`, `"learned"`, or `"learned_range"`, defaults to `"fixed_small"`): + Clip the variance when adding noise to the denoised sample. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to `1.0`): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + prediction_type (`"epsilon"`, `"sample"`, or `"v_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to `0.995`): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to `1.0`): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2", "sigmoid"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + variance_type: Literal[ + "fixed_small", + "fixed_small_log", + "fixed_large", + "fixed_large_log", + "learned", + "learned_range", + ] = "fixed_small", + clip_sample: bool = True, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "leading", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + elif beta_schedule == "laplace": + self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="laplace") + elif beta_schedule == "sigmoid": + # GeoDiff sigmoid schedule + betas = torch.linspace(-6, 6, num_train_timesteps) + self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.custom_timesteps = False + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.variance_type = variance_type + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device = None, + timesteps: list[int] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, + `num_inference_steps` must be `None`. + + """ + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + if timesteps is not None: + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.custom_timesteps = True + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + self.custom_timesteps = False + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def _get_variance( + self, + t: int, + predicted_variance: torch.Tensor | None = None, + variance_type: Literal[ + "fixed_small", + "fixed_small_log", + "fixed_large", + "fixed_large_log", + "learned", + "learned_range", + ] + | None = None, + ) -> torch.Tensor: + """ + Compute the variance for a given timestep according to the specified variance type. + + Args: + t (`int`): + The current timestep. + predicted_variance (`torch.Tensor`, *optional*): + The predicted variance from the model. Used only when `variance_type` is `"learned"` or + `"learned_range"`. + variance_type (`"fixed_small"`, `"fixed_small_log"`, `"fixed_large"`, `"fixed_large_log"`, `"learned"`, or `"learned_range"`, *optional*): + The type of variance to compute. If `None`, uses the variance type specified in the scheduler + configuration. + + Returns: + `torch.Tensor`: + The computed variance. + """ + prev_t = self.previous_timestep(t) + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t + + # we always take the log of variance, so clamp it to ensure it's not 0 + variance = torch.clamp(variance, min=1e-20) + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small": + variance = variance + # for rl-diffuser https://huggingface.co/papers/2205.09991 + elif variance_type == "fixed_small_log": + variance = torch.log(variance) + variance = torch.exp(0.5 * variance) + elif variance_type == "fixed_large": + variance = current_beta_t + elif variance_type == "fixed_large_log": + # Glide max_log + variance = torch.log(current_beta_t) + elif variance_type == "learned": + return predicted_variance + elif variance_type == "learned_range": + min_log = torch.log(variance) + max_log = torch.log(current_beta_t) + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> DDPMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + t = timestep + + prev_t = self.previous_timestep(t) + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in [ + "learned", + "learned_range", + ]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + current_alpha_t = alpha_prod_t / alpha_prod_t_prev + current_beta_t = 1 - current_alpha_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for the DDPMScheduler." + ) + + # 3. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t + current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + device = model_output.device + variance_noise = randn_tensor( + model_output.shape, + generator=generator, + device=device, + dtype=model_output.dtype, + ) + if self.variance_type == "fixed_small_log": + variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise + elif self.variance_type == "learned_range": + variance = self._get_variance(t, predicted_variance=predicted_variance) + variance = torch.exp(0.5 * variance) * variance_noise + else: + variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return DDPMSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + return self.config.num_train_timesteps + + def previous_timestep(self, timestep: int) -> int | torch.Tensor: + """ + Compute the previous timestep in the diffusion chain. + + Args: + timestep (`int`): + The current timestep. + + Returns: + `int` or `torch.Tensor`: + The previous timestep. + """ + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/diffusers/schedulers/scheduling_ddpm_parallel.py b/diffusers/schedulers/scheduling_ddpm_parallel.py new file mode 100644 index 0000000000000000000000000000000000000000..725958b01a44e511344e2b23dc1afe2f0565c2a8 --- /dev/null +++ b/diffusers/schedulers/scheduling_ddpm_parallel.py @@ -0,0 +1,769 @@ +# Copyright 2025 ParaDiGMS authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput +class DDPMParallelSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DDPMParallelScheduler(SchedulerMixin, ConfigMixin): + """ + Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and + Langevin dynamics sampling. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`. + [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and + [`~SchedulerMixin.from_pretrained`] functions. + + For more details, see the original paper: https://huggingface.co/papers/2006.11239 + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, `squaredcos_cap_v2` or `sigmoid`. + trained_betas (`np.ndarray`, *optional*): + Option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc. + variance_type (`str`, defaults to `"fixed_small"`): + Options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`, + `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`. + clip_sample (`bool`, defaults to `True`): + Option to clip predicted sample for numerical stability. + prediction_type (`str`, defaults to `"epsilon"`): + Prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion + process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4 + https://huggingface.co/papers/2210.02303) + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method (introduced by Imagen, + https://huggingface.co/papers/2205.11487). Note that the thresholding method is unsuitable for latent-space + diffusion models (such as stable-diffusion). + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen + (https://huggingface.co/papers/2205.11487). Valid only when `thresholding=True`. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample + Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + _is_ode_scheduler = False + + @register_to_config + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.__init__ + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2", "sigmoid"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + variance_type: Literal[ + "fixed_small", + "fixed_small_log", + "fixed_large", + "fixed_large_log", + "learned", + "learned_range", + ] = "fixed_small", + clip_sample: bool = True, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "leading", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + elif beta_schedule == "laplace": + self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="laplace") + elif beta_schedule == "sigmoid": + # GeoDiff sigmoid schedule + betas = torch.linspace(-6, 6, num_train_timesteps) + self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.custom_timesteps = False + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.variance_type = variance_type + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.set_timesteps + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device = None, + timesteps: list[int] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed, + `num_inference_steps` must be `None`. + + """ + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + if timesteps is not None: + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.custom_timesteps = True + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + self.custom_timesteps = False + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._get_variance + def _get_variance( + self, + t: int, + predicted_variance: torch.Tensor | None = None, + variance_type: Literal[ + "fixed_small", + "fixed_small_log", + "fixed_large", + "fixed_large_log", + "learned", + "learned_range", + ] + | None = None, + ) -> torch.Tensor: + """ + Compute the variance for a given timestep according to the specified variance type. + + Args: + t (`int`): + The current timestep. + predicted_variance (`torch.Tensor`, *optional*): + The predicted variance from the model. Used only when `variance_type` is `"learned"` or + `"learned_range"`. + variance_type (`"fixed_small"`, `"fixed_small_log"`, `"fixed_large"`, `"fixed_large_log"`, `"learned"`, or `"learned_range"`, *optional*): + The type of variance to compute. If `None`, uses the variance type specified in the scheduler + configuration. + + Returns: + `torch.Tensor`: + The computed variance. + """ + prev_t = self.previous_timestep(t) + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t + + # we always take the log of variance, so clamp it to ensure it's not 0 + variance = torch.clamp(variance, min=1e-20) + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small": + variance = variance + # for rl-diffuser https://huggingface.co/papers/2205.09991 + elif variance_type == "fixed_small_log": + variance = torch.log(variance) + variance = torch.exp(0.5 * variance) + elif variance_type == "fixed_large": + variance = current_beta_t + elif variance_type == "fixed_large_log": + # Glide max_log + variance = torch.log(current_beta_t) + elif variance_type == "learned": + return predicted_variance + elif variance_type == "learned_range": + min_log = torch.log(variance) + max_log = torch.log(current_beta_t) + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> DDPMParallelSchedulerOutput | tuple: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timestep (`int`): current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + generator (`torch.Generator`, *optional*): + Random number generator. + return_dict (`bool`): option for returning tuple rather than DDPMParallelSchedulerOutput class + + Returns: + [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] or `tuple`: + [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. + When returning a tuple, the first element is the sample tensor. + + """ + t = timestep + + prev_t = self.previous_timestep(t) + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in [ + "learned", + "learned_range", + ]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + current_alpha_t = alpha_prod_t / alpha_prod_t_prev + current_beta_t = 1 - current_alpha_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for the DDPMScheduler." + ) + + # 3. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t + current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + device = model_output.device + variance_noise = randn_tensor( + model_output.shape, + generator=generator, + device=device, + dtype=model_output.dtype, + ) + if self.variance_type == "fixed_small_log": + variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise + elif self.variance_type == "learned_range": + variance = self._get_variance(t, predicted_variance=predicted_variance) + variance = torch.exp(0.5 * variance) * variance_noise + else: + variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return DDPMParallelSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + def batch_step_no_noise( + self, + model_output: torch.Tensor, + timesteps: torch.Tensor, + sample: torch.Tensor, + ) -> torch.Tensor: + """ + Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once. + Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise + is pre-sampled by the pipeline. + + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): direct output from learned diffusion model. + timesteps (`torch.Tensor`): + Current discrete timesteps in the diffusion chain. This is a tensor of integers. + sample (`torch.Tensor`): + current instance of sample being created by diffusion process. + + Returns: + `torch.Tensor`: sample tensor at previous timestep. + """ + t = timesteps + num_inference_steps = self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps + prev_t = t - self.config.num_train_timesteps // num_inference_steps + + t = t.view(-1, *([1] * (model_output.ndim - 1))) + prev_t = prev_t.view(-1, *([1] * (model_output.ndim - 1))) + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in [ + "learned", + "learned_range", + ]: + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + pass + + # 1. compute alphas, betas + self.alphas_cumprod = self.alphas_cumprod.to(model_output.device) + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)] + alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0) + + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + current_alpha_t = alpha_prod_t / alpha_prod_t_prev + current_beta_t = 1 - current_alpha_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for the DDPMParallelScheduler." + ) + + # 3. Clip or threshold "predicted x_0" + if self.config.thresholding: + pred_original_sample = self._threshold_sample(pred_original_sample) + elif self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t + current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + return pred_prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self): + return self.config.num_train_timesteps + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep + def previous_timestep(self, timestep: int) -> int | torch.Tensor: + """ + Compute the previous timestep in the diffusion chain. + + Args: + timestep (`int`): + The current timestep. + + Returns: + `int` or `torch.Tensor`: + The previous timestep. + """ + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/diffusers/schedulers/scheduling_ddpm_wuerstchen.py b/diffusers/schedulers/scheduling_ddpm_wuerstchen.py new file mode 100644 index 0000000000000000000000000000000000000000..6747018cf46b8f0457efcb4e03d56131808584ae --- /dev/null +++ b/diffusers/schedulers/scheduling_ddpm_wuerstchen.py @@ -0,0 +1,269 @@ +# Copyright (c) 2022 Pablo Pernías MIT License +# Copyright 2025 UC Berkeley Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from dataclasses import dataclass +from typing import Literal + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +@dataclass +class DDPMWuerstchenSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of the previous timestep. `prev_sample` should be used as the next model input + in the denoising loop. + """ + + prev_sample: torch.Tensor + + +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use. Use values lower than 1 to prevent singularities. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine` or `exp`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DDPMWuerstchenScheduler(SchedulerMixin, ConfigMixin): + """ + Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and + Langevin dynamics sampling. + + [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__` + function, such as `scaler`. They can be accessed via `scheduler.config.scaler`. [`SchedulerMixin`] provides general + loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] + functions. + + For more details, see the original paper: https://huggingface.co/papers/2006.11239 + + Args: + scaler (`float`, defaults to `1.0`): + The scaling factor applied to the timesteps before computing the cumulative product of alphas. + s (`float`, defaults to `0.008`): + The offset added to the cosine noise schedule. + """ + + @register_to_config + def __init__( + self, + scaler: float = 1.0, + s: float = 0.008, + ) -> None: + self.scaler = scaler + self.s = torch.tensor([s]) + self._init_alpha_cumprod = torch.cos(self.s / (1 + self.s) * torch.pi * 0.5) ** 2 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + def _alpha_cumprod(self, t: torch.Tensor, device: torch.device) -> torch.Tensor: + if self.scaler > 1: + t = 1 - (1 - t) ** self.scaler + elif self.scaler < 1: + t = t**self.scaler + alpha_cumprod = torch.cos( + (t + self.s.to(device)) / (1 + self.s.to(device)) * torch.pi * 0.5 + ) ** 2 / self._init_alpha_cumprod.to(device) + return alpha_cumprod.clamp(0.0001, 0.9999) + + def scale_model_input(self, sample: torch.Tensor, timestep: torch.Tensor | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`torch.Tensor`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int | None = None, + timesteps: list[float] | torch.Tensor | None = None, + device: str | torch.device | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pretrained model. If passed, + `timesteps` must be `None`. + timesteps (`list[float]` or `torch.Tensor`, *optional*): + Custom timesteps used for the diffusion chain. If passed, `num_inference_steps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps are moved. + """ + if timesteps is None: + timesteps = torch.linspace(1.0, 0.0, num_inference_steps + 1, device=device) + if not isinstance(timesteps, torch.Tensor): + timesteps = torch.Tensor(timesteps).to(device) + self.timesteps = timesteps + + def step( + self, + model_output: torch.Tensor, + timestep: torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | list[torch.Generator] | None = None, + return_dict: bool = True, + ) -> DDPMWuerstchenSchedulerOutput | tuple[torch.Tensor]: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator` or `list[torch.Generator]`, *optional*): + A random number generator or a list of generators to make generation deterministic. + return_dict (`bool`, defaults to `True`): + Whether to return a [`~schedulers.scheduling_ddpm_wuerstchen.DDPMWuerstchenSchedulerOutput`] or a + `tuple`. + + Returns: + [`~schedulers.scheduling_ddpm_wuerstchen.DDPMWuerstchenSchedulerOutput`] or `tuple`: + If `return_dict` is `True`, a [`~schedulers.scheduling_ddpm_wuerstchen.DDPMWuerstchenSchedulerOutput`] + is returned, otherwise a `tuple` is returned where the first element is the sample tensor. + """ + dtype = model_output.dtype + device = model_output.device + t = timestep + + prev_t = self.previous_timestep(t) + + alpha_cumprod = self._alpha_cumprod(t, device).view(t.size(0), *[1 for _ in sample.shape[1:]]) + alpha_cumprod_prev = self._alpha_cumprod(prev_t, device).view(prev_t.size(0), *[1 for _ in sample.shape[1:]]) + alpha = alpha_cumprod / alpha_cumprod_prev + + mu = (1.0 / alpha).sqrt() * (sample - (1 - alpha) * model_output / (1 - alpha_cumprod).sqrt()) + + std_noise = randn_tensor(mu.shape, generator=generator, device=model_output.device, dtype=model_output.dtype) + std = ((1 - alpha) * (1.0 - alpha_cumprod_prev) / (1.0 - alpha_cumprod)).sqrt() * std_noise + pred = mu + std * (prev_t != 0).float().view(prev_t.size(0), *[1 for _ in sample.shape[1:]]) + + if not return_dict: + return (pred.to(dtype),) + + return DDPMWuerstchenSchedulerOutput(prev_sample=pred.to(dtype)) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise is added. + noise (`torch.Tensor`): + The noise to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps that determine the noise magnitude. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + device = original_samples.device + dtype = original_samples.dtype + alpha_cumprod = self._alpha_cumprod(timesteps, device=device).view( + timesteps.size(0), *[1 for _ in original_samples.shape[1:]] + ) + noisy_samples = alpha_cumprod.sqrt() * original_samples + (1 - alpha_cumprod).sqrt() * noise + return noisy_samples.to(dtype=dtype) + + def __len__(self) -> int: + return self.config.num_train_timesteps + + def previous_timestep(self, timestep: torch.Tensor) -> torch.Tensor: + """ + Compute the previous timestep in the diffusion chain. + + Args: + timestep (`torch.Tensor`): + The current timestep. + + Returns: + `torch.Tensor`: + The previous timestep. + """ + index = (self.timesteps - timestep[0]).abs().argmin().item() + prev_t = self.timesteps[index + 1][None].expand(timestep.shape[0]) + return prev_t diff --git a/diffusers/schedulers/scheduling_deis_multistep.py b/diffusers/schedulers/scheduling_deis_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..c22d2b5b439dccf31e62a50e69873a1bc62d7c04 --- /dev/null +++ b/diffusers/schedulers/scheduling_deis_multistep.py @@ -0,0 +1,1047 @@ +# Copyright 2025 FLAIR Lab and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: check https://huggingface.co/papers/2204.13902 and https://github.com/qsh-zh/deis for more info +# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DEISMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + `DEISMultistepScheduler` is a fast high order solver for diffusion ordinary differential equations (ODEs). + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.0001`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.02`): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray` or `list[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to `2`): + The DEIS order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`"epsilon"`, `"sample"`, `"v_prediction"`, or `"flow_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`), `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper), or `flow_prediction`. + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to `0.995`): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to `1.0`): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + algorithm_type (`"deis"`, defaults to `"deis"`): + The algorithm type for the solver. + solver_type (`"logrho"`, defaults to `"logrho"`): + Solver type for DEIS. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + flow_shift (`float`, *optional*, defaults to `1.0`): + The flow shift parameter for flow-based models. + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + use_dynamic_shifting (`bool`, defaults to `False`): + Whether to use dynamic shifting for the noise schedule. + time_shift_type (`"exponential"`, defaults to `"exponential"`): + The type of time shifting to apply. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + solver_order: int = 2, + prediction_type: Literal["epsilon", "sample", "v_prediction", "flow_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: Literal["deis"] = "deis", + solver_type: Literal["logrho"] = "logrho", + lower_order_final: bool = True, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + use_flow_sigmas: bool = False, + flow_shift: float = 1.0, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + use_dynamic_shifting: bool = False, + time_shift_type: Literal["exponential"] = "exponential", + ) -> None: + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if ( + sum( + [ + self.config.use_beta_sigmas, + self.config.use_exponential_sigmas, + self.config.use_karras_sigmas, + ] + ) + > 1 + ): + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DEIS + if algorithm_type not in ["deis"]: + if algorithm_type in ["dpmsolver", "dpmsolver++"]: + self.register_to_config(algorithm_type="deis") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["logrho"]: + if solver_type in ["midpoint", "heun", "bh1", "bh2"]: + self.register_to_config(solver_type="logrho") + else: + raise NotImplementedError(f"solver type {solver_type} is not implemented for {self.__class__}") + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None, mu: float | None = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + mu (`float`, *optional*): + The mu parameter for dynamic shifting. Only used when `use_dynamic_shifting=True` and + `time_shift_type="exponential"`. + """ + if mu is not None: + assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential" + self.config.flow_shift = np.exp(mu) + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Convert sigma values to alpha_t and sigma_t values. + + Args: + sigma (`torch.Tensor`): + The sigma value(s) to convert. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing (alpha_t, sigma_t) values. + """ + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DEIS algorithm needs. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + if self.config.prediction_type == "epsilon": + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DEISMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + if self.config.algorithm_type == "deis": + return (sample - alpha_t * x0_pred) / sigma_t + else: + raise NotImplementedError("only support log-rho multistep deis now") + + def deis_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DEIS (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "deis": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + else: + raise NotImplementedError("only support log-rho multistep deis now") + return x_t + + def multistep_deis_second_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order multistep DEIS. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + rho_t, rho_s0, rho_s1 = ( + sigma_t / alpha_t, + sigma_s0 / alpha_s0, + sigma_s1 / alpha_s1, + ) + + if self.config.algorithm_type == "deis": + + def ind_fn(t, b, c): + # Integrate[(log(t) - log(c)) / (log(b) - log(c)), {t}] + return t * (-np.log(c) + np.log(t) - 1) / (np.log(b) - np.log(c)) + + coef1 = ind_fn(rho_t, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s0, rho_s1) + coef2 = ind_fn(rho_t, rho_s1, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s0) + + x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1) + return x_t + else: + raise NotImplementedError("only support log-rho multistep deis now") + + def multistep_deis_third_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order multistep DEIS. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + rho_t, rho_s0, rho_s1, rho_s2 = ( + sigma_t / alpha_t, + sigma_s0 / alpha_s0, + sigma_s1 / alpha_s1, + sigma_s2 / alpha_s2, + ) + + if self.config.algorithm_type == "deis": + + def ind_fn(t, b, c, d): + # Integrate[(log(t) - log(c))(log(t) - log(d)) / (log(b) - log(c))(log(b) - log(d)), {t}] + numerator = t * ( + np.log(c) * (np.log(d) - np.log(t) + 1) + - np.log(d) * np.log(t) + + np.log(d) + + np.log(t) ** 2 + - 2 * np.log(t) + + 2 + ) + denominator = (np.log(b) - np.log(c)) * (np.log(b) - np.log(d)) + return numerator / denominator + + coef1 = ind_fn(rho_t, rho_s0, rho_s1, rho_s2) - ind_fn(rho_s0, rho_s0, rho_s1, rho_s2) + coef2 = ind_fn(rho_t, rho_s1, rho_s2, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s2, rho_s0) + coef3 = ind_fn(rho_t, rho_s2, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s2, rho_s0, rho_s1) + + x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1 + coef3 * m2) + + return x_t + else: + raise NotImplementedError("only support log-rho multistep deis now") + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep: int | torch.Tensor) -> None: + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DEIS. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + lower_order_final = ( + (self.step_index == len(self.timesteps) - 1) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.deis_first_order_update(model_output, sample=sample) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_deis_second_order_update(self.model_outputs, sample=sample) + else: + prev_sample = self.multistep_deis_third_order_update(self.model_outputs, sample=sample) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples without noise. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps at which to add noise to the samples. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_discrete_ddim.py b/diffusers/schedulers/scheduling_discrete_ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..fff98edced00137dae0406d580093d09128c6890 --- /dev/null +++ b/diffusers/schedulers/scheduling_discrete_ddim.py @@ -0,0 +1,316 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import math +from dataclasses import dataclass + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +@dataclass +class DiscreteDDIMSchedulerOutput(BaseOutput): + """ + Output class for the discrete DDIM scheduler. + + Args: + prev_sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Updated block tokens after the current denoising step. + sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`): + Token IDs sampled from the model logits, i.e. the predicted clean tokens `x0`. + sampled_probs (`torch.Tensor` of shape `(batch_size, block_length)`): + Probabilities of the sampled tokens. + pred_logits (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + The denoiser logits, passed through for self-conditioning the next step. + """ + + prev_sample: torch.LongTensor + sampled_tokens: torch.LongTensor + sampled_probs: torch.Tensor + pred_logits: torch.Tensor + + +class DiscreteDDIMScheduler(SchedulerMixin, ConfigMixin): + """ + Discrete DDIM scheduler for the uniform corruption process, following "Structured Denoising Diffusion Models in + Discrete State-Spaces" (D3PM, https://huggingface.co/papers/2107.03006). + + On the linear schedule the survival probability of a clean token at time `t` is `alpha(t) = 1 - t`. One denoising + step from time `t` to `s < t` samples every block position from the exact posterior `q(x_s | x_t, x0)`, which for + the uniform kernel decomposes into three routes: jump to the predicted clean token `x0`, stay on the current token, + or jump to a uniformly random token. Unlike masked diffusion, there is no mask token; uncommitted positions carry + random tokens. + + An optional predictor-corrector mode follows "Uniform Diffusion Models Revisited: Leave-One-Out Denoiser and + Absorbing State Reformulation" via the leave-one-out (LOO) denoiser (https://huggingface.co/papers/2605.22765). + When `corrector_steps > 0`, the pipeline runs that many Gibbs corrector sweeps after each predictor step (see + [`~DiscreteDDIMScheduler.step_correct`]), resampling the least-confident positions from the one-coordinate + conditional `Cat(alpha_s * x0_loo + (1 - alpha_s) / K)` while holding the rest fixed, which leaves the marginal + `p_s` invariant and improves generation at no training cost. + + Args: + num_inference_steps (`int`, defaults to 32): + The number of denoising steps, defining the linear time grid the posterior is evaluated on. + corrector_steps (`int`, defaults to 0): + Number of Gibbs corrector sweeps run after each predictor step. `0` recovers plain ancestral DDIM sampling. + corrector_k (`int`, defaults to 1): + Number of positions resampled per corrector sweep. + corrector_selection (`str`, defaults to `"lowest_log_margin"`): + How the resampled positions are chosen: `"lowest_log_margin"`, `"lowest_maxprob"`, `"lowest_current_prob"`, + or `"random"`. + corrector_selection_tau (`float`, defaults to 1.0): + Temperature of the Gumbel-top-k position selection (lower is greedier). + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_inference_steps: int = 32, + corrector_steps: int = 0, + corrector_k: int = 1, + corrector_selection: str = "lowest_log_margin", + corrector_selection_tau: float = 1.0, + ): + self.num_inference_steps = num_inference_steps + self.timesteps = torch.arange(num_inference_steps, dtype=torch.long) + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device | None = None) -> None: + if num_inference_steps <= 0: + raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.") + self.num_inference_steps = num_inference_steps + self.timesteps = torch.arange(num_inference_steps, device=device, dtype=torch.long) + + @staticmethod + def _sample_from_logits( + logits: torch.Tensor, + *, + temperature: float, + generator: torch.Generator | None, + ) -> tuple[torch.LongTensor, torch.Tensor]: + """Sample one token per position with optional temperature, returning tokens and their probabilities.""" + if temperature < 0: + raise ValueError(f"`temperature` must be >= 0, got {temperature}.") + + vocab_size = logits.shape[-1] + flat_logits = logits.reshape(-1, vocab_size) + probs = torch.softmax(flat_logits.float(), dim=-1) + + if temperature == 0.0: + token = flat_logits.argmax(dim=-1, keepdim=True) + else: + scaled_probs = torch.softmax(flat_logits.float() / temperature, dim=-1) + token = torch.multinomial(scaled_probs, num_samples=1, generator=generator) + + token_prob = torch.gather(probs, -1, token) + return token.view(*logits.shape[:-1]), token_prob.view(*logits.shape[:-1]) + + def _alpha(self, step_index: int) -> float: + """Survival probability `alpha = 1 - t` of a clean token at the time grid point `step_index`.""" + return step_index / self.num_inference_steps + + @staticmethod + def _to_loo_logits(logits: torch.Tensor, tokens: torch.LongTensor, alpha: float) -> torch.Tensor: + """ + Convert plain-denoiser logits to the leave-one-out posterior for the uniform kernel. + + Subtracts `log(1 + K * alpha / (1 - alpha))` from the observed token's logit (eq. 13 of + https://huggingface.co/papers/2605.22765); renormalization happens in the following softmax. + """ + if alpha <= 0.0 or alpha >= 1.0: + return logits + delta = math.log1p(logits.shape[-1] * alpha / (1.0 - alpha)) + shifted = logits.clone() + src = torch.full((*tokens.shape, 1), -delta, dtype=shifted.dtype, device=shifted.device) + shifted.scatter_add_(-1, tokens.unsqueeze(-1), src) + return shifted + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.LongTensor, + *, + temperature: float = 0.0, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> DiscreteDDIMSchedulerOutput | tuple[torch.LongTensor, torch.LongTensor, torch.Tensor]: + """ + Sample the next block from the posterior `q(x_s | x_t, x0)` of the uniform corruption process. + + With `a = alpha_t / alpha_s` (survival probability from `s` to `t`) and `b = alpha_s`, the posterior mass of + each route is + + clean: `b * (1 - a) / K + a * b * 1[x_t = x0]`, stay: `a * (1 - b) / K`, noise: `(1 - a) * (1 - b) / K`, + + so the last step (`b = 1`) deterministically commits the predicted clean tokens. + + Args: + model_output (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + Raw logits from the model for the current block. + timestep (`int` or `torch.Tensor`): + Current step index within the denoising schedule, in `[0, num_inference_steps - 1]`. + sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Current block token IDs `x_t`. + temperature (`float`): + Sampling temperature applied to the logits when drawing `x0`. + generator (`torch.Generator`, *optional*): + RNG for sampling. + return_dict (`bool`): + Whether to return a [`DiscreteDDIMSchedulerOutput`] or a plain tuple. + """ + if isinstance(timestep, torch.Tensor): + step_index = int(timestep.item()) + else: + step_index = int(timestep) + + sampled_tokens, sampled_probs = self._sample_from_logits( + model_output, temperature=temperature, generator=generator + ) + + vocab_size = model_output.shape[-1] + num_steps = self.num_inference_steps + # `step_index` counts up from 0 to `num_inference_steps - 1`: alpha(t) = 1 - t increases towards the clean end, + # with alpha_s = 1 on the final step so the predicted clean tokens are committed deterministically. + alpha_t = step_index / num_steps + alpha_s = (step_index + 1) / num_steps + survival = alpha_t / alpha_s + + same = (sample == sampled_tokens).float() + clean_mass = alpha_s * (1 - survival) / vocab_size + survival * alpha_s * same + stay_mass = survival * (1 - alpha_s) / vocab_size * torch.ones_like(same) + noise_mass = (1 - survival) * (1 - alpha_s) / vocab_size * torch.ones_like(same) + + route_probs = torch.stack([clean_mass, stay_mass, noise_mass], dim=-1) + route_probs = route_probs / route_probs.sum(dim=-1, keepdim=True) + routes = torch.multinomial(route_probs.view(-1, 3), num_samples=1, generator=generator).view_as(sample) + + random_tokens = torch.randint( + low=0, high=vocab_size, size=sample.shape, device=sample.device, generator=generator + ) + prev_sample = torch.where(routes == 0, sampled_tokens, sample) + prev_sample = torch.where(routes == 2, random_tokens, prev_sample) + + if not return_dict: + return prev_sample, sampled_tokens, sampled_probs, model_output + return DiscreteDDIMSchedulerOutput( + prev_sample=prev_sample, + sampled_tokens=sampled_tokens, + sampled_probs=sampled_probs, + pred_logits=model_output, + ) + + def _select_positions( + self, sample: torch.LongTensor, cond_log_probs: torch.Tensor, generator: torch.Generator | None + ) -> torch.LongTensor: + """Pick `corrector_k` positions per row to resample, least-confident first (Gumbel-top-k without replacement).""" + selection = self.config.corrector_selection + batch_size, seq_len = sample.shape + k_eff = min(max(1, int(self.config.corrector_k)), seq_len) + + if selection == "random": + scores = torch.rand(batch_size, seq_len, device=sample.device, generator=generator) + return torch.topk(scores, k=k_eff, dim=-1).indices + + if selection == "lowest_maxprob": + confidence = -cond_log_probs.max(dim=-1).values + elif selection == "lowest_current_prob": + confidence = -torch.gather(cond_log_probs, -1, sample.unsqueeze(-1)).squeeze(-1) + elif selection == "lowest_log_margin": + log_current = torch.gather(cond_log_probs, -1, sample.unsqueeze(-1)).squeeze(-1) + alt = cond_log_probs.clone().scatter_(-1, sample.unsqueeze(-1), float("-inf")) + confidence = -(log_current - alt.max(dim=-1).values) + else: + raise ValueError(f"Unknown `corrector_selection`: {selection!r}.") + + keys = confidence / float(self.config.corrector_selection_tau) + u = torch.rand(keys.shape, device=keys.device, generator=generator).clamp_(1e-12, 1.0 - 1e-12) + keys = keys + (-torch.log(-torch.log(u))) + return torch.topk(keys, k=k_eff, dim=-1).indices + + def step_correct( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.LongTensor, + *, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> DiscreteDDIMSchedulerOutput | tuple[torch.LongTensor, torch.LongTensor, torch.Tensor]: + """ + Run one Gibbs corrector sweep at the post-predictor time `s`, following the leave-one-out predictor-corrector + of https://huggingface.co/papers/2605.22765. + + The model logits (recomputed on the current `sample`) are converted to the LOO denoiser, the one-coordinate + conditional `p_s(x^l | x^{-l}) = Cat(alpha_s * x0_loo + (1 - alpha_s) / K)` is formed, the least-confident + `corrector_k` positions are selected, and those positions are resampled while the rest are held fixed. The + sweep preserves `p_s`, so it refines the sample without changing its marginal and needs no extra training. + + Args: + model_output (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + Raw logits from the model recomputed on the current (post-predictor) `sample`. + timestep (`int` or `torch.Tensor`): + The predictor step index just completed; the corrector runs at the following grid point `s`. + sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Current block token IDs to refine. + generator (`torch.Generator`, *optional*): + RNG for sampling. + return_dict (`bool`): + Whether to return a [`DiscreteDDIMSchedulerOutput`] or a plain tuple. + """ + if isinstance(timestep, torch.Tensor): + step_index = int(timestep.item()) + else: + step_index = int(timestep) + + # The corrector acts at the cleaner time `s` reached by the predictor. + alpha_s = self._alpha(step_index + 1) + vocab_size = model_output.shape[-1] + + # Match the reference corrector, which forms the conditional in float64 (the LOO correction reaches ~log(K)). + loo_logits = self._to_loo_logits(model_output.double(), sample, alpha_s) + loo_log_probs = torch.log_softmax(loo_logits, dim=-1) + log_uniform = math.log1p(-alpha_s) - math.log(vocab_size) + cond_log_probs = torch.logaddexp( + math.log(alpha_s) + loo_log_probs, torch.full_like(loo_log_probs, log_uniform) + ) + + positions = self._select_positions(sample, cond_log_probs, generator) + rows = torch.arange(sample.shape[0], device=sample.device).unsqueeze(-1).expand_as(positions) + chosen_probs = cond_log_probs[rows, positions].exp() + resampled = torch.multinomial( + chosen_probs.reshape(-1, vocab_size), num_samples=1, generator=generator + ).view_as(positions) + + prev_sample = sample.clone() + prev_sample[rows, positions] = resampled + sampled_probs = torch.gather(chosen_probs, -1, resampled.unsqueeze(-1)).squeeze(-1) + + if not return_dict: + return prev_sample, resampled, sampled_probs, model_output + return DiscreteDDIMSchedulerOutput( + prev_sample=prev_sample, + sampled_tokens=resampled, + sampled_probs=sampled_probs, + pred_logits=model_output, + ) + + +__all__ = ["DiscreteDDIMScheduler", "DiscreteDDIMSchedulerOutput"] diff --git a/diffusers/schedulers/scheduling_dpm_cogvideox.py b/diffusers/schedulers/scheduling_dpm_cogvideox.py new file mode 100644 index 0000000000000000000000000000000000000000..4a238a5408aed3d421e0c8cf426f05250ea1d1d7 --- /dev/null +++ b/diffusers/schedulers/scheduling_dpm_cogvideox.py @@ -0,0 +1,600 @@ +# Copyright 2025 The CogVideoX team, Tsinghua University & ZhipuAI and The HuggingFace Team. +# All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM +class DDIMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +def rescale_zero_terminal_snr(alphas_cumprod): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + the betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: rescaled betas with zero terminal SNR + """ + + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + + return alphas_bar + + +class CogVideoXDPMScheduler(SchedulerMixin, ConfigMixin): + """ + `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. Choose from + `leading`, `linspace` or `trailing`. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + snr_shift_scale (`float`, defaults to 3.0): + Shift scale for SNR. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.0120, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "scaled_linear", + trained_betas: np.ndarray | list[float] | None = None, + clip_sample: bool = True, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + clip_sample_range: float = 1.0, + sample_max_value: float = 1.0, + timestep_spacing: Literal["leading", "linspace", "trailing"] = "leading", + rescale_betas_zero_snr: bool = False, + snr_shift_scale: float = 3.0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float64, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # Modify: SNR shift following SD3 + self.alphas_cumprod = self.alphas_cumprod / (snr_shift_scale + (1 - snr_shift_scale) * self.alphas_cumprod) + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.alphas_cumprod = rescale_zero_terminal_snr(self.alphas_cumprod) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + def _get_variance(self, timestep, prev_timestep): + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None` (the default), the timesteps are not + moved. + """ + + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps) + .round()[::-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'." + ) + + self.timesteps = torch.from_numpy(timesteps).to(device) + + def get_variables( + self, + alpha_prod_t: torch.Tensor, + alpha_prod_t_prev: torch.Tensor, + alpha_prod_t_back: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Compute the variables used for DPM-Solver++ (2M) referencing the original implementation. + + Args: + alpha_prod_t (`torch.Tensor`): + The cumulative product of alphas at the current timestep. + alpha_prod_t_prev (`torch.Tensor`): + The cumulative product of alphas at the previous timestep. + alpha_prod_t_back (`torch.Tensor`, *optional*): + The cumulative product of alphas at the timestep before the previous timestep. + + Returns: + `tuple`: + A tuple containing the variables `h`, `r`, `lamb`, `lamb_next`. + """ + lamb = ((alpha_prod_t / (1 - alpha_prod_t)) ** 0.5).log() + lamb_next = ((alpha_prod_t_prev / (1 - alpha_prod_t_prev)) ** 0.5).log() + h = lamb_next - lamb + + if alpha_prod_t_back is not None: + lamb_previous = ((alpha_prod_t_back / (1 - alpha_prod_t_back)) ** 0.5).log() + h_last = lamb - lamb_previous + r = h_last / h + return h, r, lamb, lamb_next + else: + return h, None, lamb, lamb_next + + def get_mult( + self, + h: torch.Tensor, + r: torch.Tensor, + alpha_prod_t: torch.Tensor, + alpha_prod_t_prev: torch.Tensor, + alpha_prod_t_back: torch.Tensor | None = None, + ) -> tuple[torch.Tensor, torch.Tensor] | tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Compute the multipliers for the previous sample and the predicted original sample. + + Args: + h (`torch.Tensor`): + The log-SNR difference. + r (`torch.Tensor`): + The ratio of log-SNR differences. + alpha_prod_t (`torch.Tensor`): + The cumulative product of alphas at the current timestep. + alpha_prod_t_prev (`torch.Tensor`): + The cumulative product of alphas at the previous timestep. + alpha_prod_t_back (`torch.Tensor`, *optional*): + The cumulative product of alphas at the timestep before the previous timestep. + + Returns: + `tuple`: + A tuple containing the multipliers. + """ + mult1 = ((1 - alpha_prod_t_prev) / (1 - alpha_prod_t)) ** 0.5 * (-h).exp() + mult2 = (-2 * h).expm1() * alpha_prod_t_prev**0.5 + + if alpha_prod_t_back is not None: + mult3 = 1 + 1 / (2 * r) + mult4 = 1 / (2 * r) + return mult1, mult2, mult3, mult4 + else: + return mult1, mult2 + + def step( + self, + model_output: torch.Tensor, + old_pred_original_sample: torch.Tensor, + timestep: int, + timestep_back: int, + sample: torch.Tensor, + eta: float = 0.0, + use_clipped_model_output: bool = False, + generator: torch.Generator | None = None, + variance_noise: torch.Tensor | None = None, + return_dict: bool = False, + ) -> DDIMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + old_pred_original_sample (`torch.Tensor`): + The predicted original sample from the previous timestep. + timestep (`int`): + The current discrete timestep in the diffusion chain. + timestep_back (`int`): + The timestep to look back to. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`): + The weight of noise for added noise in diffusion step. + use_clipped_model_output (`bool`, defaults to `False`): + If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary + because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no + clipping has happened, "corrected" `model_output` would coincide with the one provided as input and + `use_clipped_model_output` has no effect. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + # See formulas (12) and (16) of DDIM paper https://huggingface.co/papers/2010.02502 + # Ideally, read DDIM paper in-detail understanding + + # Notation ( -> + # - pred_noise_t -> e_theta(x_t, t) + # - pred_original_sample -> f_theta(x_t, t) or x_0 + # - std_dev_t -> sigma_t + # - eta -> η + # - pred_sample_direction -> "direction pointing to x_t" + # - pred_prev_sample -> "x_t-1" + + # 1. get previous step value (=t-1) + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + alpha_prod_t_back = self.alphas_cumprod[timestep_back] if timestep_back is not None else None + + beta_prod_t = 1 - alpha_prod_t + + # 3. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (12) from https://huggingface.co/papers/2010.02502 + # To make style tests pass, commented out `pred_epsilon` as it is an unused variable + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + # pred_epsilon = model_output + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + # pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + elif self.config.prediction_type == "v_prediction": + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + # pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction`" + ) + + h, r, lamb, lamb_next = self.get_variables(alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back) + mult = list(self.get_mult(h, r, alpha_prod_t, alpha_prod_t_prev, alpha_prod_t_back)) + mult_noise = (1 - alpha_prod_t_prev) ** 0.5 * (1 - (-2 * h).exp()) ** 0.5 + + noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype) + prev_sample = mult[0] * sample - mult[1] * pred_original_sample + mult_noise * noise + + if old_pred_original_sample is None or prev_timestep < 0: + # Save a network evaluation if all noise levels are 0 or on the first step + return prev_sample, pred_original_sample + else: + denoised_d = mult[2] * pred_original_sample - mult[3] * old_pred_original_sample + noise = randn_tensor( + sample.shape, + generator=generator, + device=sample.device, + dtype=sample.dtype, + ) + x_advanced = mult[0] * sample - mult[1] * denoised_d + mult_noise * noise + + prev_sample = x_advanced + + if not return_dict: + return (prev_sample, pred_original_sample) + + return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_dpmsolver_multistep.py b/diffusers/schedulers/scheduling_dpmsolver_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..9c15df4569caf6534230f0c0d6b1db7176f63b66 --- /dev/null +++ b/diffusers/schedulers/scheduling_dpmsolver_multistep.py @@ -0,0 +1,1347 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas): + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + `DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`"epsilon"`, `"sample"`, `"v_prediction"`, or `"flow_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function. `epsilon` predicts the noise of the diffusion process, `sample` + directly predicts the noisy sample, `v_prediction` predicts the velocity (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper), and `flow_prediction` predicts the flow. + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`"dpmsolver"`, `"dpmsolver++"`, `"sde-dpmsolver"`, or `"sde-dpmsolver++"`, defaults to `"dpmsolver++"`): + Algorithm type for the solver. The `dpmsolver` type implements the algorithms in the + [DPMSolver](https://huggingface.co/papers/2206.00927) paper, and the `dpmsolver++` type implements the + algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use + `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. + solver_type (`"midpoint"` or `"heun"`, defaults to `"midpoint"`): + Solver type for the second-order solver. The solver type slightly affects the sample quality, especially + for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_lu_lambdas (`bool`, *optional*, defaults to `False`): + Whether to use the uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the noise schedule during + the sampling process. If `True`, the sigmas and time steps are determined according to a sequence of + `lambda(t)`. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + flow_shift (`float`, *optional*, defaults to 1.0): + The shift value for the timestep schedule for flow matching. + final_sigmas_type (`"zero"` or `"sigma_min"`, *optional*, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `"zero"`, the final sigma is set to 0. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`"learned"` or `"learned_range"`, *optional*): + Set to `"learned"` or `"learned_range"` for diffusion models that predict variance. If set, the model's + output contains the predicted Gaussian variance. + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + use_dynamic_shifting (`bool`, defaults to `False`): + Whether to use dynamic shifting for the timestep schedule. + time_shift_type (`"exponential"`, defaults to `"exponential"`): + The type of time shift to apply when using dynamic shifting. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + solver_order: int = 2, + prediction_type: Literal["epsilon", "sample", "v_prediction", "flow_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: Literal["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"] = "dpmsolver++", + solver_type: Literal["midpoint", "heun"] = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + use_lu_lambdas: bool = False, + use_flow_sigmas: bool = False, + flow_shift: float = 1.0, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", + lambda_min_clipped: float = -float("inf"), + variance_type: Literal["learned", "learned_range"] | None = None, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + use_dynamic_shifting: bool = False, + time_shift_type: Literal["exponential"] = "exponential", + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if ( + sum( + [ + self.config.use_beta_sigmas, + self.config.use_exponential_sigmas, + self.config.use_karras_sigmas, + ] + ) + > 1 + ): + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead" + deprecate( + "algorithm_types dpmsolver and sde-dpmsolver", + "1.0.0", + deprecation_message, + ) + + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DPM-Solver + if algorithm_type not in [ + "dpmsolver", + "dpmsolver++", + "sde-dpmsolver", + "sde-dpmsolver++", + ]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero": + raise ValueError( + f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead." + ) + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device = None, + mu: float | None = None, + timesteps: list[int] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary timesteps schedule. If `None`, timesteps will be generated + based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas` + must be `None`, and `timestep_spacing` attribute will be ignored. + """ + if mu is not None: + assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential" + self.config.flow_shift = np.exp(mu) + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.") + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot use `timesteps` with `config.use_karras_sigmas = True`") + if timesteps is not None and self.config.use_lu_lambdas: + raise ValueError("Cannot use `timesteps` with `config.use_lu_lambdas = True`") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + + if timesteps is not None: + timesteps = np.array(timesteps).astype(np.int64) + else: + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped) + last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item() + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, last_timestep - 1, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = last_timestep // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + ) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.beta_schedule != "squaredcos_cap_v2": + timesteps = timesteps.round() + elif self.config.use_lu_lambdas: + lambdas = np.flip(log_sigmas.copy()) + lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps) + sigmas = np.exp(lambdas) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.beta_schedule != "squaredcos_cap_v2": + timesteps = timesteps.round() + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Convert sigma values to alpha_t and sigma_t values. + + Args: + sigma (`torch.Tensor`): + The sigma value(s) to convert. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing (alpha_t, sigma_t) values. + """ + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + def _convert_to_lu(self, in_lambdas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model + Sampling in Around 10 Steps](https://huggingface.co/papers/2206.00927) by Lu et al. (2022). + + Args: + in_lambdas (`torch.Tensor`): + The input lambda values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted lambda values following the Lu noise schedule. + """ + + lambda_min: float = in_lambdas[-1].item() + lambda_max: float = in_lambdas[0].item() + + rho = 1.0 # 1.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = lambda_min ** (1 / rho) + max_inv_rho = lambda_max ** (1 / rho) + lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return lambdas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both + noise > prediction and data prediction models. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by the diffusion process. + noise (`torch.Tensor`, *optional*): + The noise tensor. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + x_t = ( + (alpha_t / alpha_s) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + def multistep_dpm_solver_second_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * (torch.exp(h) - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + def multistep_dpm_solver_third_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by diffusion process. + noise (`torch.Tensor`, *optional*): + The noise tensor. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + def _init_step_index(self, timestep: int | torch.Tensor) -> None: + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + variance_noise: torch.Tensor | None = None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`, *optional*): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`LEdits++`]. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If `return_dict` is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final + or (self.config.lower_order_final and len(self.timesteps) < 15) + or self.config.final_sigmas_type == "zero" + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None: + noise = randn_tensor( + model_output.shape, + generator=generator, + device=model_output.device, + dtype=torch.float32, + ) + elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]: + noise = variance_noise.to(device=model_output.device, dtype=torch.float32) + else: + noise = None + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + else: + prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample, noise=noise) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # Cast sample back to expected dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples without noise. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps at which to add noise to the samples. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py b/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py new file mode 100644 index 0000000000000000000000000000000000000000..3386514f64e9e88805f93dc4443990b5a9de121e --- /dev/null +++ b/diffusers/schedulers/scheduling_dpmsolver_multistep_inverse.py @@ -0,0 +1,1166 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin): + """ + `DPMSolverMultistepInverseScheduler` is the reverse scheduler of [`DPMSolverMultistepScheduler`]. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `dpmsolver++`): + Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The + `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) + paper, and the `dpmsolver++` type implements the algorithms in the + [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or + `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + flow_shift (`float`, *optional*, defaults to 1.0): + The flow shift factor. Valid only when `use_flow_sigmas=True`. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`str`, *optional*): + Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output + contains the predicted Gaussian variance. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + solver_order: int = 2, + prediction_type: Literal["epsilon", "sample", "v_prediction", "flow_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: Literal["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"] = "dpmsolver++", + solver_type: Literal["midpoint", "heun"] = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + use_flow_sigmas: bool = False, + flow_shift: float = 1.0, + lambda_min_clipped: float = -float("inf"), + variance_type: Literal["learned", "learned_range"] | None = None, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if ( + sum( + [ + self.config.use_beta_sigmas, + self.config.use_exponential_sigmas, + self.config.use_karras_sigmas, + ] + ) + > 1 + ): + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead" + deprecate( + "algorithm_types dpmsolver and sde-dpmsolver", + "1.0.0", + deprecation_message, + ) + + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DPM-Solver + if algorithm_type not in [ + "dpmsolver", + "dpmsolver++", + "sde-dpmsolver", + "sde-dpmsolver++", + ]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32).copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.use_karras_sigmas = use_karras_sigmas + self.use_exponential_sigmas = use_exponential_sigmas + self.use_beta_sigmas = use_beta_sigmas + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped).item() + self.noisiest_timestep = self.config.num_train_timesteps - 1 - clipped_idx + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.noisiest_timestep, num_inference_steps + 1).round()[:-1].copy().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = (self.noisiest_timestep + 1) // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(self.noisiest_timestep + 1, 0, -step_ratio).round()[::-1].copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', " + "'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + timesteps = timesteps.copy().astype(np.int64) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigma_max = ( + (1 - self.alphas_cumprod[self.noisiest_timestep]) / self.alphas_cumprod[self.noisiest_timestep] + ) ** 0.5 + sigmas = np.concatenate([sigmas, [sigma_max]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + + # when num_inference_steps == num_train_timesteps, we can end up with + # duplicates in timesteps. + _, unique_indices = np.unique(timesteps, return_index=True) + timesteps = timesteps[np.sort(unique_indices)] + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Convert sigma values to alpha_t and sigma_t values. + + Args: + sigma (`torch.Tensor`): + The sigma value(s) to convert. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing (alpha_t, sigma_t) values. + """ + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.convert_model_output + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both + noise > prediction and data prediction models. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the DPMSolverMultistepScheduler." + ) + + if self.config.thresholding: + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.dpm_solver_first_order_update + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by the diffusion process. + noise (`torch.Tensor`, *optional*): + The noise tensor. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + x_t = ( + (alpha_t / alpha_s) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_second_order_update + def multistep_dpm_solver_second_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.algorithm_type == "sde-dpmsolver": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * (torch.exp(h) - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s0) * sample + - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise + ) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_third_order_update + def multistep_dpm_solver_third_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`, *optional*): + A current instance of a sample created by diffusion process. + noise (`torch.Tensor`, *optional*): + The noise tensor. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (alpha_t / alpha_s0) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h) - 2.0 * h) / (2.0 * h) ** 2 - 0.5)) * D2 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def _init_step_index(self, timestep: int | torch.Tensor): + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + + index_candidates = (self.timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + self._step_index = step_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + variance_noise: torch.Tensor | None = None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + variance_noise (`torch.Tensor`): + Alternative to generating noise with `generator` by directly providing the noise for the variance + itself. Useful for methods such as [`CycleDiffusion`]. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final or (self.config.lower_order_final and len(self.timesteps) < 15) + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"] and variance_noise is None: + noise = randn_tensor( + model_output.shape, + generator=generator, + device=model_output.device, + dtype=model_output.dtype, + ) + elif self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]: + noise = variance_noise + else: + noise = None + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + else: + prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the clean `original_samples` using the scheduler's equivalent function. + + Args: + original_samples (`torch.Tensor`): + The original samples to add noise to. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps at which to add noise. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + step_indices = [] + for timestep in timesteps: + index_candidates = (schedule_timesteps == timestep).nonzero() + if len(index_candidates) == 0: + step_index = len(schedule_timesteps) - 1 + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + step_indices.append(step_index) + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_dpmsolver_sde.py b/diffusers/schedulers/scheduling_dpmsolver_sde.py new file mode 100644 index 0000000000000000000000000000000000000000..b288d9f61b1489d6cb37daa5b1af0a83794b78cb --- /dev/null +++ b/diffusers/schedulers/scheduling_dpmsolver_sde.py @@ -0,0 +1,825 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Callable, List, Literal, Optional, Tuple, Union + +import numpy as np +import torch +import torchsde + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DPMSolverSDE +class DPMSolverSDESchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +class BatchedBrownianTree: + """A wrapper around torchsde.BrownianTree that enables batches of entropy.""" + + def __init__( + self, + x: torch.Tensor, + t0: float, + t1: float, + seed: Optional[Union[int, List[int]]] = None, + **kwargs, + ): + t0, t1, self.sign = self.sort(t0, t1) + w0 = kwargs.get("w0", torch.zeros_like(x)) + if seed is None: + seed = torch.randint(0, 2**63 - 1, []).item() + self.batched = True + try: + assert len(seed) == x.shape[0] + w0 = w0[0] + except TypeError: + seed = [seed] + self.batched = False + self.trees = [ + torchsde.BrownianInterval( + t0=t0, + t1=t1, + size=w0.shape, + dtype=w0.dtype, + device=w0.device, + entropy=s, + tol=1e-6, + pool_size=24, + halfway_tree=True, + ) + for s in seed + ] + + @staticmethod + def sort(a: float, b: float) -> Tuple[float, float, float]: + """ + Sorts two float values and returns them along with a sign indicating if they were swapped. + + Args: + a (`float`): + The first value. + b (`float`): + The second value. + + Returns: + `Tuple[float, float, float]`: + A tuple containing the sorted values (min, max) and a sign (1.0 if a < b, -1.0 otherwise). + """ + return (a, b, 1.0) if a < b else (b, a, -1.0) + + def __call__(self, t0: float, t1: float) -> torch.Tensor: + t0, t1, sign = self.sort(t0, t1) + w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign) + return w if self.batched else w[0] + + +class BrownianTreeNoiseSampler: + """A noise sampler backed by a torchsde.BrownianTree. + + Args: + x (`torch.Tensor`): The tensor whose shape, device and dtype is used to generate random samples. + sigma_min (`float`): The low end of the valid interval. + sigma_max (`float`): The high end of the valid interval. + seed (`int` or `List[int]`): The random seed. If a list of seeds is + supplied instead of a single integer, then the noise sampler will use one BrownianTree per batch item, each + with its own seed. + transform (`callable`): A function that maps sigma to the sampler's + internal timestep. + """ + + def __init__( + self, + x: torch.Tensor, + sigma_min: float, + sigma_max: float, + seed: Optional[Union[int, List[int]]] = None, + transform: Callable[[float], float] = lambda x: x, + ): + self.transform = transform + t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max)) + self.tree = BatchedBrownianTree(x, t0, t1, seed) + + def __call__(self, sigma: float, sigma_next: float) -> torch.Tensor: + t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next)) + return self.tree(t0, t1) / (t1 - t0).abs().sqrt() + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin): + """ + DPMSolverSDEScheduler implements the stochastic sampler from the [Elucidating the Design Space of Diffusion-Based + Generative Models](https://huggingface.co/papers/2206.00364) paper. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.00085): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.012): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + noise_sampler_seed (`int`, *optional*, defaults to `None`): + The random seed to use for the noise sampler. If `None`, a random seed is generated. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + noise_sampler_seed: int | None = None, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if ( + sum( + [ + self.config.use_beta_sigmas, + self.config.use_exponential_sigmas, + self.config.use_karras_sigmas, + ] + ) + > 1 + ): + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + self.use_karras_sigmas = use_karras_sigmas + self.noise_sampler = None + self.noise_sampler_seed = noise_sampler_seed + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def init_noise_sigma(self) -> torch.Tensor: + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self) -> Union[int, None]: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self) -> Union[int, None]: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: float | torch.Tensor, + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sigma_input = sigma if self.state_in_first_order else self.mid_point_sigma + sample = sample / ((sigma_input**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device = None, + num_train_timesteps: int | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + num_train_timesteps (`int`, *optional*): + The number of train timesteps. If `None`, uses `self.config.num_train_timesteps`. + """ + self.num_inference_steps = num_inference_steps + + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(float) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + second_order_timesteps = self._second_order_timesteps(sigmas, log_sigmas) + + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]]) + + timesteps = torch.from_numpy(timesteps) + second_order_timesteps = torch.from_numpy(second_order_timesteps) + timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)]) + timesteps[1::2] = second_order_timesteps + + if str(device).startswith("mps"): + # mps does not support float64 + self.timesteps = timesteps.to(device, dtype=torch.float32) + else: + self.timesteps = timesteps.to(device=device) + + # empty first order variables + self.sample = None + self.mid_point_sigma = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.noise_sampler = None + + def _second_order_timesteps(self, sigmas: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + def sigma_fn(_t): + return np.exp(-_t) + + def t_fn(_sigma): + return -np.log(_sigma) + + midpoint_ratio = 0.5 + t = t_fn(sigmas) + delta_time = np.diff(t) + t_proposed = t[:-1] + delta_time * midpoint_ratio + sig_proposed = sigma_fn(t_proposed) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sig_proposed]) + return timesteps + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_lms_discrete.LMSDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + sigma_min: float = in_sigmas[-1].item() + sigma_max: float = in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, self.num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + @property + def state_in_first_order(self) -> bool: + return self.sample is None + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + return_dict: bool = True, + s_noise: float = 1.0, + ) -> DPMSolverSDESchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] or + tuple. + s_noise (`float`, *optional*, defaults to 1.0): + Scaling factor for noise added to the sample. + + Returns: + [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_dpmsolver_sde.DPMSolverSDESchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + # Create a noise sampler if it hasn't been created yet + if self.noise_sampler is None: + min_sigma, max_sigma = self.sigmas[self.sigmas > 0].min(), self.sigmas.max() + self.noise_sampler = BrownianTreeNoiseSampler( + sample, min_sigma.item(), max_sigma.item(), self.noise_sampler_seed + ) + + # Define functions to compute sigma and t from each other + def sigma_fn(_t: torch.Tensor) -> torch.Tensor: + return _t.neg().exp() + + def t_fn(_sigma: torch.Tensor) -> torch.Tensor: + return _sigma.log().neg() + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order + sigma = self.sigmas[self.step_index - 1] + sigma_next = self.sigmas[self.step_index] + + # Set the midpoint and step size for the current step + midpoint_ratio = 0.5 + t, t_next = t_fn(sigma), t_fn(sigma_next) + delta_time = t_next - t + t_proposed = t + delta_time * midpoint_ratio + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma if self.state_in_first_order else sigma_fn(t_proposed) + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma if self.state_in_first_order else sigma_fn(t_proposed) + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if sigma_next == 0: + derivative = (sample - pred_original_sample) / sigma + dt = sigma_next - sigma + prev_sample = sample + derivative * dt + else: + if self.state_in_first_order: + t_next = t_proposed + else: + sample = self.sample + + sigma_from = sigma_fn(t) + sigma_to = sigma_fn(t_next) + sigma_up = min( + sigma_to, + (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5, + ) + sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5 + ancestral_t = t_fn(sigma_down) + prev_sample = (sigma_fn(ancestral_t) / sigma_fn(t)) * sample - ( + t - ancestral_t + ).expm1() * pred_original_sample + prev_sample = prev_sample + self.noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * sigma_up + + if self.state_in_first_order: + # store for 2nd order step + self.sample = sample + self.mid_point_sigma = sigma_fn(t_next) + else: + # free for "first order mode" + self.sample = None + self.mid_point_sigma = None + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return DPMSolverSDESchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_dpmsolver_singlestep.py b/diffusers/schedulers/scheduling_dpmsolver_singlestep.py new file mode 100644 index 0000000000000000000000000000000000000000..002f2b844a527d710310e769b52f3c3b704ae6df --- /dev/null +++ b/diffusers/schedulers/scheduling_dpmsolver_singlestep.py @@ -0,0 +1,1312 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin): + """ + `DPMSolverSinglestepScheduler` is a fast dedicated high-order solver for diffusion ODEs. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.0001`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.02`): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray` or `list[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to `2`): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + prediction_type (`"epsilon"`, `"sample"`, `"v_prediction"`, or `"flow_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`), `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper), or `flow_prediction`. + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to `0.995`): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to `1.0`): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`"dpmsolver"`, `"dpmsolver++"`, or `"sde-dpmsolver++"`, defaults to `"dpmsolver++"`): + Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, or `sde-dpmsolver++`. The `dpmsolver` + type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927) paper, and the + `dpmsolver++` type implements the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) + paper. It is recommended to use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided + sampling like in Stable Diffusion. + solver_type (`"midpoint"` or `"heun"`, defaults to `"midpoint"`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `False`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + flow_shift (`float`, *optional*, defaults to `1.0`): + The flow shift parameter for flow-based models. + final_sigmas_type (`"zero"` or `"sigma_min"`, *optional*, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `"zero"`, the final sigma is set to 0. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`"learned"` or `"learned_range"`, *optional*): + Set to `"learned"` or `"learned_range"` for diffusion models that predict variance. If set, the model's + output contains the predicted Gaussian variance. + use_dynamic_shifting (`bool`, defaults to `False`): + Whether to use dynamic shifting for the noise schedule. + time_shift_type (`"exponential"`, defaults to `"exponential"`): + The type of time shifting to apply. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: Literal["linear", "scaled_linear", "squaredcos_cap_v2"] = "linear", + trained_betas: np.ndarray | list[float] | None = None, + solver_order: int = 2, + prediction_type: Literal["epsilon", "sample", "v_prediction", "flow_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: Literal["dpmsolver", "dpmsolver++", "sde-dpmsolver++"] = "dpmsolver++", + solver_type: Literal["midpoint", "heun"] = "midpoint", + lower_order_final: bool = False, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + use_flow_sigmas: bool = False, + flow_shift: float = 1.0, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", + lambda_min_clipped: float = -float("inf"), + variance_type: Literal["learned", "learned_range"] | None = None, + use_dynamic_shifting: bool = False, + time_shift_type: Literal["exponential"] = "exponential", + ) -> None: + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if algorithm_type == "dpmsolver": + deprecation_message = "algorithm_type `dpmsolver` is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead" + deprecate("algorithm_types=dpmsolver", "1.0.0", deprecation_message) + + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # settings for DPM-Solver + if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver++"]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero": + raise ValueError( + f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead." + ) + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.sample = None + self.order_list = self.get_order_list(num_train_timesteps) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + def get_order_list(self, num_inference_steps: int) -> list[int]: + """ + Computes the solver order at each time step. + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + + Returns: + `list[int]`: + The list of solver orders for each timestep. + """ + steps = num_inference_steps + order = self.config.solver_order + if order > 3: + raise ValueError("Order > 3 is not supported by this scheduler") + if self.config.lower_order_final: + if order == 3: + if steps % 3 == 0: + orders = [1, 2, 3] * (steps // 3 - 1) + [1, 2] + [1] + elif steps % 3 == 1: + orders = [1, 2, 3] * (steps // 3) + [1] + else: + orders = [1, 2, 3] * (steps // 3) + [1, 2] + elif order == 2: + if steps % 2 == 0: + orders = [1, 2] * (steps // 2 - 1) + [1, 1] + else: + orders = [1, 2] * (steps // 2) + [1] + elif order == 1: + orders = [1] * steps + else: + if order == 3: + orders = [1, 2, 3] * (steps // 3) + elif order == 2: + orders = [1, 2] * (steps // 2) + elif order == 1: + orders = [1] * steps + + if self.config.final_sigmas_type == "zero": + orders[-1] = 1 + + return orders + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device = None, + mu: float | None = None, + timesteps: list[int] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps schedule is used. If `timesteps` is + passed, `num_inference_steps` must be `None`. + """ + if mu is not None: + assert self.config.use_dynamic_shifting and self.config.time_shift_type == "exponential" + self.config.flow_shift = np.exp(mu) + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.") + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot use `timesteps` when `config.use_karras_sigmas=True`.") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + + num_inference_steps = num_inference_steps or len(timesteps) + self.num_inference_steps = num_inference_steps + + if timesteps is not None: + timesteps = np.array(timesteps).astype(np.int64) + else: + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped) + clipped_idx = clipped_idx.item() + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1 - clipped_idx, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f" `final_sigmas_type` must be one of `sigma_min` or `zero`, but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas).to(device=device) + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + self.model_outputs = [None] * self.config.solver_order + self.sample = None + + if not self.config.lower_order_final and num_inference_steps % self.config.solver_order != 0: + logger.warning( + "Changing scheduler {self.config} to have `lower_order_final` set to True to handle uneven amount of inference steps. Please make sure to always use an even number of `num_inference steps when using `lower_order_final=False`." + ) + self.register_to_config(lower_order_final=True) + + if not self.config.lower_order_final and self.config.final_sigmas_type == "zero": + logger.warning( + " `last_sigmas_type='zero'` is not supported for `lower_order_final=False`. Changing scheduler {self.config} to have `lower_order_final` set to True." + ) + self.register_to_config(lower_order_final=True) + + self.order_list = self.get_order_list(num_inference_steps) + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Convert sigma values to alpha_t and sigma_t values. + + Args: + sigma (`torch.Tensor`): + The sigma value(s) to convert. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing (alpha_t, sigma_t) values. + """ + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both + noise > prediction and data prediction models. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + # DPM-Solver++ needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]: + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the DPMSolverSinglestepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # DPM-Solver needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type == "dpmsolver": + if self.config.prediction_type == "epsilon": + # DPM-Solver and DPM-Solver++ only need the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the DPMSolverSinglestepScheduler." + ) + + if self.config.thresholding: + alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "dpmsolver": + x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def singlestep_dpm_solver_second_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the second-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the + time `timestep_list[-2]`. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): + The current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s1, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m1, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s1) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s1) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s1) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s1) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s1 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s1 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def singlestep_dpm_solver_third_order_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the third-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the + time `timestep_list[-3]`. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): + The current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s2, lambda_s0 - lambda_s2, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m2 + D1_0, D1_1 = (1.0 / r1) * (m1 - m2), (1.0 / r0) * (m0 - m2) + D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1) + D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s2) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1_1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s2) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "dpmsolver": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (alpha_t / alpha_s2) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1_1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (alpha_t / alpha_s2) * sample + - (sigma_t * (torch.exp(h) - 1.0)) * D0 + - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1 + - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s2 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1_1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s2 * torch.exp(-h)) * sample + + (alpha_t * (1.0 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h) + (-2.0 * h)) / (-2.0 * h) ** 2 - 0.5)) * D2 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + return x_t + + def singlestep_dpm_solver_update( + self, + model_output_list: list[torch.Tensor], + *args, + sample: torch.Tensor | None = None, + order: int = None, + noise: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the singlestep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + timestep (`int`): + The current and latter discrete timestep in the diffusion chain. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + order (`int`): + The solver order at this step. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None) + prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 2: + sample = args[2] + else: + raise ValueError("missing `sample` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if timestep_list is not None: + deprecate( + "timestep_list", + "1.0.0", + "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + if order == 1: + return self.dpm_solver_first_order_update(model_output_list[-1], sample=sample, noise=noise) + elif order == 2: + return self.singlestep_dpm_solver_second_order_update(model_output_list, sample=sample, noise=noise) + elif order == 3: + return self.singlestep_dpm_solver_third_order_update(model_output_list, sample=sample, noise=noise) + else: + raise ValueError(f"Order must be 1, 2, 3, got {order}") + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep: int | torch.Tensor) -> None: + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the singlestep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator for stochastic sampling. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.algorithm_type == "sde-dpmsolver++": + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype + ) + else: + noise = None + + order = self.order_list[self.step_index] + + # For img2img denoising might start with order>1 which is not possible + # In this case make sure that the first two steps are both order=1 + while self.model_outputs[-order] is None: + order -= 1 + + # For single-step solvers, we use the initial value at each time with order = 1. + if order == 1: + self.sample = sample + + prev_sample = self.singlestep_dpm_solver_update( + self.model_outputs, sample=self.sample, order=order, noise=noise + ) + + # upon completion increase step index by one, noise=noise + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples without noise. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps at which to add noise to the samples. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py b/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..20ad17cfdd967ab9b9bd537796397a84177718c7 --- /dev/null +++ b/diffusers/schedulers/scheduling_edm_dpmsolver_multistep.py @@ -0,0 +1,873 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm + +import math +from typing import List, Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + Implements DPMSolverMultistepScheduler in EDM formulation as presented in Karras et al. 2022 [1]. + `EDMDPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs. + + [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models." + https://huggingface.co/papers/2206.00364 + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + sigma_min (`float`, *optional*, defaults to 0.002): + Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable + range is [0, 10]. + sigma_max (`float`, *optional*, defaults to 80.0): + Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable + range is [0.2, 80.0]. + sigma_data (`float`, *optional*, defaults to 0.5): + The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1]. + sigma_schedule (`str`, *optional*, defaults to `karras`): + Sigma schedule to compute the `sigmas`. By default, we the schedule introduced in the EDM paper + (https://huggingface.co/papers/2206.00364). Other acceptable value is "exponential". The exponential + schedule was incorporated in this model: https://huggingface.co/stabilityai/cosxl. + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + rho (`float`, *optional*, defaults to 7.0): + The rho parameter in the Karras sigma schedule. This was set to 7.0 in the EDM paper [1]. + solver_order (`int`, defaults to 2): + The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided + sampling, and `solver_order=3` for unconditional sampling. + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `dpmsolver++`): + Algorithm type for the solver; can be `dpmsolver++` or `sde-dpmsolver++`. The `dpmsolver++` type implements + the algorithms in the [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to + use `dpmsolver++` or `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion. + solver_type (`str`, defaults to `midpoint`): + Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the + sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + euler_at_final (`bool`, defaults to `False`): + Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail + richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference + steps, but sometimes may result in blurring. + final_sigmas_type (`str`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.002, + sigma_max: float = 80.0, + sigma_data: float = 0.5, + sigma_schedule: Literal["karras", "exponential"] = "karras", + num_train_timesteps: int = 1000, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + rho: float = 7.0, + solver_order: int = 2, + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: Literal["dpmsolver++", "sde-dpmsolver++"] = "dpmsolver++", + solver_type: Literal["midpoint", "heun"] = "midpoint", + lower_order_final: bool = True, + euler_at_final: bool = False, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", # "zero", "sigma_min" + ): + # settings for DPM-Solver + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"]: + if algorithm_type == "deis": + self.register_to_config(algorithm_type="dpmsolver++") + else: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + if solver_type not in ["midpoint", "heun"]: + if solver_type in ["logrho", "bh1", "bh2"]: + self.register_to_config(solver_type="midpoint") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero": + raise ValueError( + f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead." + ) + + ramp = torch.linspace(0, 1, num_train_timesteps) + if sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + self.timesteps = self.precondition_noise(sigmas) + + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + # setable values + self.num_inference_steps = None + self.model_outputs = [None] * solver_order + self.lower_order_nums = 0 + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> float: + # standard deviation of the initial noise distribution + return (self.config.sigma_max**2 + 1) ** 0.5 + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs + def precondition_inputs(self, sample: torch.Tensor, sigma: float | torch.Tensor) -> torch.Tensor: + """ + Precondition the input sample by scaling it according to the EDM formulation. + + Args: + sample (`torch.Tensor`): + The input sample tensor to precondition. + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `torch.Tensor`: + The scaled input sample. + """ + c_in = self._get_conditioning_c_in(sigma) + scaled_sample = sample * c_in + return scaled_sample + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_noise + def precondition_noise(self, sigma: float | torch.Tensor) -> torch.Tensor: + """ + Precondition the noise level by applying a logarithmic transformation. + + Args: + sigma (`float` or `torch.Tensor`): + The sigma (noise level) value to precondition. + + Returns: + `torch.Tensor`: + The preconditioned noise value computed as `0.25 * log(sigma)`. + """ + if not isinstance(sigma, torch.Tensor): + sigma = torch.tensor([sigma]) + + c_noise = 0.25 * torch.log(sigma) + + return c_noise + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs + def precondition_outputs( + self, + sample: torch.Tensor, + model_output: torch.Tensor, + sigma: float | torch.Tensor, + ) -> torch.Tensor: + """ + Precondition the model outputs according to the EDM formulation. + + Args: + sample (`torch.Tensor`): + The input sample tensor. + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `torch.Tensor`: + The denoised sample computed by combining the skip connection and output scaling. + """ + sigma_data = self.config.sigma_data + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + + if self.config.prediction_type == "epsilon": + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + elif self.config.prediction_type == "v_prediction": + c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + else: + raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.") + + denoised = c_skip * sample + c_out * model_output + + return denoised + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Scale the denoising model input to match the Euler algorithm. Ensures interchangeability with schedulers that + need to scale the denoising model input depending on the current timestep. + + Args: + sample (`torch.Tensor`): + The input sample tensor. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = self.precondition_inputs(sample, sigma) + + self.is_scale_input_called = True + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + + ramp = torch.linspace(0, 1, self.num_inference_steps) + if self.config.sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(ramp) + elif self.config.sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(ramp) + + sigmas = sigmas.to(dtype=torch.float32, device=device) + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = self.config.sigma_min + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.tensor([sigma_last], dtype=torch.float32, device=device)]) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_karras_sigmas + def _compute_karras_sigmas( + self, + ramp: torch.Tensor, + sigma_min: float | None = None, + sigma_max: float | None = None, + ) -> torch.Tensor: + """ + Construct the noise schedule of [Karras et al. (2022)](https://huggingface.co/papers/2206.00364). + + Args: + ramp (`torch.Tensor`): + A tensor of values in [0, 1] representing the interpolation positions. + sigma_min (`float`, *optional*): + Minimum sigma value. If `None`, uses `self.config.sigma_min`. + sigma_max (`float`, *optional*): + Maximum sigma value. If `None`, uses `self.config.sigma_max`. + + Returns: + `torch.Tensor`: + The computed Karras sigma schedule. + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._compute_exponential_sigmas + def _compute_exponential_sigmas( + self, + ramp: torch.Tensor, + sigma_min: float | None = None, + sigma_max: float | None = None, + ) -> torch.Tensor: + """ + Compute the exponential sigma schedule. Implementation closely follows k-diffusion: + https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 + + Args: + ramp (`torch.Tensor`): + A tensor of values representing the interpolation positions. + sigma_min (`float`, *optional*): + Minimum sigma value. If `None`, uses `self.config.sigma_min`. + sigma_max (`float`, *optional*): + Maximum sigma value. If `None`, uses `self.config.sigma_max`. + + Returns: + `torch.Tensor`: + The computed exponential sigma schedule. + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), len(ramp)).exp().flip(0) + return sigmas + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _sigma_to_alpha_sigma_t(self, sigma): + alpha_t = torch.tensor(1) # Inputs are pre-scaled before going into unet, so alpha_t = 1 + sigma_t = sigma + return alpha_t, sigma_t + + def convert_model_output( + self, + model_output: torch.Tensor, + sample: torch.Tensor, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is + designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an + integral of the data prediction model. + + > [!TIP] > The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both + noise > prediction and data prediction models. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + sigma = self.sigmas[self.step_index] + x0_pred = self.precondition_outputs(sample, model_output, sigma) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + def dpm_solver_first_order_update( + self, + model_output: torch.Tensor, + sample: torch.Tensor, + noise: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + One step for the first-order DPMSolver (equivalent to DDIM). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + noise (`torch.Tensor`, *optional*): + The noise tensor to add to the original samples. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s = torch.log(alpha_s) - torch.log(sigma_s) + + h = lambda_t - lambda_s + if self.config.algorithm_type == "dpmsolver++": + x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + x_t = ( + (sigma_t / sigma_s * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + def multistep_dpm_solver_second_order_update( + self, + model_output_list: List[torch.Tensor], + sample: torch.Tensor, + noise: torch.Tensor | None = None, + ) -> torch.Tensor: + """ + One step for the second-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + noise (`torch.Tensor`, *optional*): + The noise tensor to add to the original samples. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s0, sigma_s1 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + + m0, m1 = model_output_list[-1], model_output_list[-2] + + h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 + r0 = h_0 / h + D0, D1 = m0, (1.0 / r0) * (m0 - m1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2211.01095 for detailed derivations + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1 + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + ) + elif self.config.algorithm_type == "sde-dpmsolver++": + assert noise is not None + if self.config.solver_type == "midpoint": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + elif self.config.solver_type == "heun": + x_t = ( + (sigma_t / sigma_s0 * torch.exp(-h)) * sample + + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1 + + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise + ) + + return x_t + + def multistep_dpm_solver_third_order_update( + self, + model_output_list: list[torch.Tensor], + sample: torch.Tensor, + ) -> torch.Tensor: + """ + One step for the third-order multistep DPMSolver. + + Args: + model_output_list (`list[torch.Tensor]`): + The direct outputs from learned diffusion model at current and latter timesteps. + sample (`torch.Tensor`): + A current instance of a sample created by diffusion process. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + sigma_t, sigma_s0, sigma_s1, sigma_s2 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + self.sigmas[self.step_index - 2], + ) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1) + alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1) + lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2) + + m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3] + + h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2 + r0, r1 = h_0 / h, h_1 / h + D0 = m0 + D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2) + D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1) + D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1) + if self.config.algorithm_type == "dpmsolver++": + # See https://huggingface.co/papers/2206.00927 for detailed derivations + x_t = ( + (sigma_t / sigma_s0) * sample + - (alpha_t * (torch.exp(-h) - 1.0)) * D0 + + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1 + - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2 + ) + + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep: int | torch.Tensor) -> None: + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep DPMSolver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Improve numerical stability for small number of steps + lower_order_final = (self.step_index == len(self.timesteps) - 1) and ( + self.config.euler_at_final + or (self.config.lower_order_final and len(self.timesteps) < 15) + or self.config.final_sigmas_type == "zero" + ) + lower_order_second = ( + (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15 + ) + + model_output = self.convert_model_output(model_output, sample=sample) + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.model_outputs[-1] = model_output + + if self.config.algorithm_type == "sde-dpmsolver++": + noise = randn_tensor( + model_output.shape, + generator=generator, + device=model_output.device, + dtype=model_output.dtype, + ) + else: + noise = None + + if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final: + prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise) + elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second: + prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise) + else: + prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler._get_conditioning_c_in + def _get_conditioning_c_in(self, sigma: float | torch.Tensor) -> float | torch.Tensor: + """ + Compute the input conditioning factor for the EDM formulation. + + Args: + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `float` or `torch.Tensor`: + The input conditioning factor `c_in`. + """ + c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + return c_in + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_edm_euler.py b/diffusers/schedulers/scheduling_edm_euler.py new file mode 100644 index 0000000000000000000000000000000000000000..dfb70ce68b46d4e9861fad7bf707f6a5ac62f566 --- /dev/null +++ b/diffusers/schedulers/scheduling_edm_euler.py @@ -0,0 +1,600 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete +class EDMEulerSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +class EDMEulerScheduler(SchedulerMixin, ConfigMixin): + """ + Implements the Euler scheduler in EDM formulation as presented in Karras et al. 2022 [1]. + + [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models." + https://huggingface.co/papers/2206.00364 + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + sigma_min (`float`, *optional*, defaults to `0.002`): + Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable + range is [0, 10]. + sigma_max (`float`, *optional*, defaults to `80.0`): + Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable + range is [0.2, 80.0]. + sigma_data (`float`, *optional*, defaults to `0.5`): + The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1]. + sigma_schedule (`Literal["karras", "exponential"]`, *optional*, defaults to `"karras"`): + Sigma schedule to compute the `sigmas`. By default, we use the schedule introduced in the EDM paper + (https://huggingface.co/papers/2206.00364). The `"exponential"` schedule was incorporated in this model: + https://huggingface.co/stabilityai/cosxl. + num_train_timesteps (`int`, *optional*, defaults to `1000`): + The number of diffusion steps to train the model. + prediction_type (`Literal["epsilon", "v_prediction"]`, *optional*, defaults to `"epsilon"`): + Prediction type of the scheduler function. `"epsilon"` predicts the noise of the diffusion process, and + `"v_prediction"` (see section 2.4 of [Imagen Video](https://huggingface.co/papers/2210.02303) paper). + rho (`float`, *optional*, defaults to `7.0`): + The rho parameter used for calculating the Karras sigma schedule, which is set to 7.0 in the EDM paper [1]. + final_sigmas_type (`Literal["zero", "sigma_min"]`, *optional*, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `"zero"`, the final sigma is set to 0. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + sigma_min: float = 0.002, + sigma_max: float = 80.0, + sigma_data: float = 0.5, + sigma_schedule: Literal["karras", "exponential"] = "karras", + num_train_timesteps: int = 1000, + prediction_type: Literal["epsilon", "v_prediction"] = "epsilon", + rho: float = 7.0, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", + ) -> None: + if sigma_schedule not in ["karras", "exponential"]: + raise ValueError(f"Wrong value for provided for `{sigma_schedule=}`.`") + + # setable values + self.num_inference_steps = None + + sigmas_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64 + sigmas = torch.arange(num_train_timesteps + 1, dtype=sigmas_dtype) / num_train_timesteps + if sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(sigmas) + elif sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(sigmas) + sigmas = sigmas.to(torch.float32) + + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.full((1,), fill_value=sigma_last, device=sigmas.device)]) + + self.is_scale_input_called = False + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> float: + """ + Return the standard deviation of the initial noise distribution. + + Returns: + `float`: + The initial noise sigma value computed as `(sigma_max**2 + 1) ** 0.5`. + """ + return (self.config.sigma_max**2 + 1) ** 0.5 + + @property + def step_index(self) -> int: + """ + Return the index counter for the current timestep. The index will increase by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not yet initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + Return the index for the first timestep. This should be set from the pipeline with the `set_begin_index` + method. + + Returns: + `int` or `None`: + The begin index, or `None` if not yet set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def precondition_inputs(self, sample: torch.Tensor, sigma: float | torch.Tensor) -> torch.Tensor: + """ + Precondition the input sample by scaling it according to the EDM formulation. + + Args: + sample (`torch.Tensor`): + The input sample tensor to precondition. + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `torch.Tensor`: + The scaled input sample. + """ + c_in = self._get_conditioning_c_in(sigma) + scaled_sample = sample * c_in + return scaled_sample + + def precondition_noise(self, sigma: float | torch.Tensor) -> torch.Tensor: + """ + Precondition the noise level by applying a logarithmic transformation. + + Args: + sigma (`float` or `torch.Tensor`): + The sigma (noise level) value to precondition. + + Returns: + `torch.Tensor`: + The preconditioned noise value computed as `0.25 * log(sigma)`. + """ + if not isinstance(sigma, torch.Tensor): + sigma = torch.tensor([sigma]) + + c_noise = 0.25 * torch.log(sigma) + + return c_noise + + def precondition_outputs( + self, + sample: torch.Tensor, + model_output: torch.Tensor, + sigma: float | torch.Tensor, + ) -> torch.Tensor: + """ + Precondition the model outputs according to the EDM formulation. + + Args: + sample (`torch.Tensor`): + The input sample tensor. + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `torch.Tensor`: + The denoised sample computed by combining the skip connection and output scaling. + """ + sigma_data = self.config.sigma_data + c_skip = sigma_data**2 / (sigma**2 + sigma_data**2) + + if self.config.prediction_type == "epsilon": + c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + elif self.config.prediction_type == "v_prediction": + c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5 + else: + raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.") + + denoised = c_skip * sample + c_out * model_output + + return denoised + + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Scale the denoising model input to match the Euler algorithm. Ensures interchangeability with schedulers that + need to scale the denoising model input depending on the current timestep. + + Args: + sample (`torch.Tensor`): + The input sample tensor. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = self.precondition_inputs(sample, sigma) + + self.is_scale_input_called = True + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device = None, + sigmas: torch.Tensor | list[float] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`torch.Tensor | list[float]`, *optional*): + Custom sigmas to use for the denoising process. If not defined, the default behavior when + `num_inference_steps` is passed will be used. + """ + self.num_inference_steps = num_inference_steps + + sigmas_dtype = torch.float32 if torch.backends.mps.is_available() else torch.float64 + if sigmas is None: + sigmas = torch.linspace(0, 1, self.num_inference_steps, dtype=sigmas_dtype) + elif isinstance(sigmas, float): + sigmas = torch.tensor(sigmas, dtype=sigmas_dtype) + else: + sigmas = sigmas.to(sigmas_dtype) + if self.config.sigma_schedule == "karras": + sigmas = self._compute_karras_sigmas(sigmas) + elif self.config.sigma_schedule == "exponential": + sigmas = self._compute_exponential_sigmas(sigmas) + sigmas = sigmas.to(dtype=torch.float32, device=device) + + self.timesteps = self.precondition_noise(sigmas) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + self.sigmas = torch.cat([sigmas, torch.full((1,), fill_value=sigma_last, device=sigmas.device)]) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Taken from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17 + def _compute_karras_sigmas( + self, + ramp: torch.Tensor, + sigma_min: float | None = None, + sigma_max: float | None = None, + ) -> torch.Tensor: + """ + Construct the noise schedule of [Karras et al. (2022)](https://huggingface.co/papers/2206.00364). + + Args: + ramp (`torch.Tensor`): + A tensor of values in [0, 1] representing the interpolation positions. + sigma_min (`float`, *optional*): + Minimum sigma value. If `None`, uses `self.config.sigma_min`. + sigma_max (`float`, *optional*): + Maximum sigma value. If `None`, uses `self.config.sigma_max`. + + Returns: + `torch.Tensor`: + The computed Karras sigma schedule. + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + + rho = self.config.rho + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + def _compute_exponential_sigmas( + self, + ramp: torch.Tensor, + sigma_min: float | None = None, + sigma_max: float | None = None, + ) -> torch.Tensor: + """ + Compute the exponential sigma schedule. Implementation closely follows k-diffusion: + https://github.com/crowsonkb/k-diffusion/blob/6ab5146d4a5ef63901326489f31f1d8e7dd36b48/k_diffusion/sampling.py#L26 + + Args: + ramp (`torch.Tensor`): + A tensor of values representing the interpolation positions. + sigma_min (`float`, *optional*): + Minimum sigma value. If `None`, uses `self.config.sigma_min`. + sigma_max (`float`, *optional*): + Maximum sigma value. If `None`, uses `self.config.sigma_max`. + + Returns: + `torch.Tensor`: + The computed exponential sigma schedule. + """ + sigma_min = sigma_min or self.config.sigma_min + sigma_max = sigma_max or self.config.sigma_max + sigmas = torch.linspace(math.log(sigma_min), math.log(sigma_max), len(ramp)).exp().flip(0) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: torch.Generator | None = None, + return_dict: bool = True, + pred_original_sample: torch.Tensor | None = None, + ) -> EDMEulerSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`, *optional*, defaults to `0.0`): + The amount of stochasticity to add at each step. Higher values add more noise. + s_tmin (`float`, *optional*, defaults to `0.0`): + The minimum sigma threshold below which no noise is added. + s_tmax (`float`, *optional*, defaults to `float("inf")`): + The maximum sigma threshold above which no noise is added. + s_noise (`float`, *optional*, defaults to `1.0`): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator for reproducibility. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return an [`~schedulers.scheduling_edm_euler.EDMEulerSchedulerOutput`] or tuple. + pred_original_sample (`torch.Tensor`, *optional*): + The predicted denoised sample from a previous step. If provided, skips recomputation. + + Returns: + [`~schedulers.scheduling_edm_euler.EDMEulerSchedulerOutput`] or `tuple`: + If `return_dict` is `True`, an [`~schedulers.scheduling_edm_euler.EDMEulerSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the previous sample tensor and the + second element is the predicted original sample tensor. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EDMEulerScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + + gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 + + sigma_hat = sigma * (gamma + 1) + + if gamma > 0: + noise = randn_tensor( + model_output.shape, + dtype=model_output.dtype, + device=model_output.device, + generator=generator, + ) + eps = noise * s_noise + sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5 + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if pred_original_sample is None: + pred_original_sample = self.precondition_outputs(sample, model_output, sigma_hat) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma_hat + + dt = self.sigmas[self.step_index + 1] - sigma_hat + + prev_sample = sample + derivative * dt + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return EDMEulerSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def _get_conditioning_c_in(self, sigma: float | torch.Tensor) -> float | torch.Tensor: + """ + Compute the input conditioning factor for the EDM formulation. + + Args: + sigma (`float` or `torch.Tensor`): + The current sigma (noise level) value. + + Returns: + `float` or `torch.Tensor`: + The input conditioning factor `c_in`. + """ + c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5) + return c_in + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_entropy_bound.py b/diffusers/schedulers/scheduling_entropy_bound.py new file mode 100644 index 0000000000000000000000000000000000000000..5382190ec6bf2374bab0932e1f08a0423160a543 --- /dev/null +++ b/diffusers/schedulers/scheduling_entropy_bound.py @@ -0,0 +1,185 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +from dataclasses import dataclass + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +@dataclass +class EntropyBoundSchedulerOutput(BaseOutput): + """ + Output class for the entropy bound scheduler. + + Args: + prev_sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Updated block tokens after the current denoising step. + accepted_index (`torch.BoolTensor` of shape `(batch_size, block_length)`): + Boolean mask of the positions accepted (committed) in this step. + sampled_tokens (`torch.LongTensor` of shape `(batch_size, block_length)`): + Token IDs sampled from the model logits. + sampled_probs (`torch.Tensor` of shape `(batch_size, block_length)`): + Probabilities of the sampled tokens. + pred_logits (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + The temperature-scaled logits the candidates were drawn from, for self-conditioning the next step. + """ + + prev_sample: torch.LongTensor + accepted_index: torch.BoolTensor + sampled_tokens: torch.LongTensor + sampled_probs: torch.Tensor + pred_logits: torch.Tensor + + +class EntropyBoundScheduler(SchedulerMixin, ConfigMixin): + """ + Entropy bound scheduler for the uniform corruption process. + + At each step the scheduler samples a candidate token per position and accepts the `k` lowest-entropy positions such + that `sum_i^k entropy_i - max(entropy_1, ..., entropy_k) <= entropy_bound`. The left-hand side upper-bounds the + joint mutual information between the accepted tokens, so they are approximately independent. Accepted positions + keep their sampled token; the rest are renoised with uniformly random tokens (there is no mask token). + + Proposed in "Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking" + (https://huggingface.co/papers/2505.24857). + + The sampling temperature is annealed from `t_max` on the first step down to `t_min` on the last, matching the + released checkpoint's sampler (sharper sampling as denoising advances). It is applied to the logits before both the + candidate sampling and the entropy that drives acceptance. + + Args: + entropy_bound (`float`, defaults to 0.1): + The maximum tolerated joint entropy of the accepted tokens. Larger values accept more tokens per step. + t_max (`float`, defaults to 0.8): + Sampling temperature on the first denoising step. + t_min (`float`, defaults to 0.4): + Sampling temperature on the last denoising step. + num_inference_steps (`int`, defaults to 32): + The maximum number of denoising steps. + """ + + order = 1 + + @register_to_config + def __init__( + self, entropy_bound: float = 0.1, t_max: float = 0.8, t_min: float = 0.4, num_inference_steps: int = 32 + ): + self.num_inference_steps = num_inference_steps + self.timesteps = torch.arange(num_inference_steps, dtype=torch.long) + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device | None = None) -> None: + if num_inference_steps <= 0: + raise ValueError(f"`num_inference_steps` must be > 0, got {num_inference_steps}.") + self.num_inference_steps = num_inference_steps + self.timesteps = torch.arange(num_inference_steps, device=device, dtype=torch.long) + + @staticmethod + # Copied from diffusers.schedulers.scheduling_discrete_ddim.DiscreteDDIMScheduler._sample_from_logits + def _sample_from_logits( + logits: torch.Tensor, + *, + temperature: float, + generator: torch.Generator | None, + ) -> tuple[torch.LongTensor, torch.Tensor]: + """Sample one token per position with optional temperature, returning tokens and their probabilities.""" + if temperature < 0: + raise ValueError(f"`temperature` must be >= 0, got {temperature}.") + + vocab_size = logits.shape[-1] + flat_logits = logits.reshape(-1, vocab_size) + probs = torch.softmax(flat_logits.float(), dim=-1) + + if temperature == 0.0: + token = flat_logits.argmax(dim=-1, keepdim=True) + else: + scaled_probs = torch.softmax(flat_logits.float() / temperature, dim=-1) + token = torch.multinomial(scaled_probs, num_samples=1, generator=generator) + + token_prob = torch.gather(probs, -1, token) + return token.view(*logits.shape[:-1]), token_prob.view(*logits.shape[:-1]) + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.LongTensor, + *, + entropy_bound: float | None = None, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> ( + EntropyBoundSchedulerOutput + | tuple[torch.LongTensor, torch.BoolTensor, torch.LongTensor, torch.Tensor, torch.Tensor] + ): + """ + Accept the lowest-entropy positions under the entropy bound and renoise the rest. + + Args: + model_output (`torch.Tensor` of shape `(batch_size, block_length, vocab_size)`): + Raw logits from the model for the current block. + timestep (`int` or `torch.Tensor`): + Current step index within the denoising schedule; sets the annealed sampling temperature. + sample (`torch.LongTensor` of shape `(batch_size, block_length)`): + Current block token IDs. + entropy_bound (`float`, *optional*): + Overrides the configured entropy bound for this step. + generator (`torch.Generator`, *optional*): + RNG for sampling. + return_dict (`bool`): + Whether to return an [`EntropyBoundSchedulerOutput`] or a plain tuple. + """ + if entropy_bound is None: + entropy_bound = float(self.config.entropy_bound) + + # Anneal the temperature from `t_max` to `t_min` over the schedule and scale the logits by it once, so the + # acceptance entropy is measured on the same distribution the candidates are drawn from. + fraction = (self.num_inference_steps - int(timestep)) / self.num_inference_steps + temperature = self.config.t_min + (self.config.t_max - self.config.t_min) * fraction + model_output = model_output / temperature + sampled_tokens, sampled_probs = self._sample_from_logits(model_output, temperature=1.0, generator=generator) + + token_entropy = torch.distributions.Categorical(logits=model_output).entropy() # (batch, block_length) + sorted_token_entropy, sorted_indices = torch.sort(token_entropy, dim=-1, descending=False) + cumulative_entropy = torch.cumsum(sorted_token_entropy, dim=-1) + + # `sorted_token_entropy` is the running maximum entropy (ascending order), so the left-hand side bounds the + # joint mutual information of the accepted tokens. + sorted_accepted = cumulative_entropy - sorted_token_entropy <= entropy_bound + accepted_index = torch.scatter( + input=torch.zeros_like(sorted_accepted), dim=-1, index=sorted_indices, src=sorted_accepted + ) + + random_tokens = torch.randint( + low=0, high=model_output.shape[-1], size=sample.shape, device=sample.device, generator=generator + ) + prev_sample = torch.where(accepted_index, sampled_tokens, random_tokens) + + if not return_dict: + return prev_sample, accepted_index, sampled_tokens, sampled_probs, model_output + return EntropyBoundSchedulerOutput( + prev_sample=prev_sample, + accepted_index=accepted_index, + sampled_tokens=sampled_tokens, + sampled_probs=sampled_probs, + pred_logits=model_output, + ) + + +__all__ = ["EntropyBoundScheduler", "EntropyBoundSchedulerOutput"] diff --git a/diffusers/schedulers/scheduling_euler_ancestral_discrete.py b/diffusers/schedulers/scheduling_euler_ancestral_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..8ad6abd90668670fc9834825a45ecf039263aa9d --- /dev/null +++ b/diffusers/schedulers/scheduling_euler_ancestral_discrete.py @@ -0,0 +1,528 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerAncestralDiscrete +class EulerAncestralDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Ancestral sampling with Euler method steps. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`"epsilon"`, `"sample"`, or `"v_prediction"`, defaults to `"epsilon"`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + ) -> None: + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas) + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.is_scale_input_called = False + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> torch.Tensor: + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + self.is_scale_input_called = True + return sample + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[ + ::-1 + ].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas).to(device=device) + + self.timesteps = torch.from_numpy(timesteps).to(device=device) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> EulerAncestralDiscreteSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + sigma_from = self.sigmas[self.step_index] + sigma_to = self.sigmas[self.step_index + 1] + sigma_up = (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5 + sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5 + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + + dt = sigma_down - sigma + + prev_sample = sample + derivative * dt + + device = model_output.device + noise = randn_tensor(model_output.shape, dtype=model_output.dtype, device=device, generator=generator) + + prev_sample = prev_sample + noise * sigma_up + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return EulerAncestralDiscreteSchedulerOutput( + prev_sample=prev_sample, pred_original_sample=pred_original_sample + ) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_euler_discrete.py b/diffusers/schedulers/scheduling_euler_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..eac6efe18aaf775b846e7536fd972686d3ec9683 --- /dev/null +++ b/diffusers/schedulers/scheduling_euler_discrete.py @@ -0,0 +1,906 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete +class EulerDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Euler scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`Literal["linear", "scaled_linear", "squaredcos_cap_v2"]`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`Literal["epsilon", "sample", "v_prediction"]`, defaults to `"epsilon"`, *optional*): + Prediction type of the scheduler function; can be `"epsilon"` (predicts the noise of the diffusion + process), `"sample"` (directly predicts the noisy sample`) or `"v_prediction"` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + interpolation_type (`Literal["linear", "log_linear"]`, defaults to `"linear"`, *optional*): + The interpolation type to compute intermediate sigmas for the scheduler denoising steps. Should be one of + `"linear"` or `"log_linear"`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + sigma_min (`float`, *optional*): + The minimum sigma value for the noise schedule. If not provided, defaults to the last sigma in the + schedule. + sigma_max (`float`, *optional*): + The maximum sigma value for the noise schedule. If not provided, defaults to the first sigma in the + schedule. + timestep_spacing (`Literal["linspace", "leading", "trailing"]`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + timestep_type (`Literal["discrete", "continuous"]`, defaults to `"discrete"`): + The type of timesteps to use. Can be `"discrete"` or `"continuous"`. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + final_sigmas_type (`Literal["zero", "sigma_min"]`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `"zero"`, the final sigma is set to 0. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + prediction_type: str = "epsilon", + interpolation_type: str = "linear", + use_karras_sigmas: bool | None = False, + use_exponential_sigmas: bool | None = False, + use_beta_sigmas: bool | None = False, + sigma_min: float | None = None, + sigma_max: float | None = None, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + timestep_type: Literal["discrete", "continuous"] = "discrete", + steps_offset: int = 0, + rescale_betas_zero_snr: bool = False, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + sigmas = (((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5).flip(0) + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + # setable values + self.num_inference_steps = None + + # TODO: Support the full EDM scalings for all prediction types and timestep types + if timestep_type == "continuous" and prediction_type == "v_prediction": + self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas]) + else: + self.timesteps = timesteps + + self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + self.is_scale_input_called = False + self.use_karras_sigmas = use_karras_sigmas + self.use_exponential_sigmas = use_exponential_sigmas + self.use_beta_sigmas = use_beta_sigmas + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> float | torch.Tensor: + """ + The standard deviation of the initial noise distribution. + + Returns: + `float` or `torch.Tensor`: + The standard deviation of the initial noise distribution, computed based on the maximum sigma value and + the timestep spacing configuration. + """ + max_sigma = max(self.sigmas) if isinstance(self.sigmas, list) else self.sigmas.max() + if self.config.timestep_spacing in ["linspace", "trailing"]: + return max_sigma + + return (max_sigma**2 + 1) ** 0.5 + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index for the scheduler, or `None` if not set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm. + + Args: + sample (`torch.Tensor`): + The input sample to be scaled. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample, divided by `(sigma**2 + 1) ** 0.5`. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + + self.is_scale_input_called = True + return sample + + def set_timesteps( + self, + num_inference_steps: int = None, + device: str | torch.device = None, + timesteps: list[int] | None = None, + sigmas: list[float] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If `None`, + `timesteps` or `sigmas` must be provided. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary timesteps schedule. If `None`, timesteps will be generated + based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` and `sigmas` + must be `None`, and `timestep_spacing` attribute will be ignored. + sigmas (`list[float]`, *optional*): + Custom sigmas used to support arbitrary timesteps schedule. If `None`, timesteps and sigmas will be + generated based on the relevant scheduler attributes. If `sigmas` is passed, `num_inference_steps` and + `timesteps` must be `None`, and the timesteps will be generated based on the custom sigmas schedule. + """ + + if timesteps is not None and sigmas is not None: + raise ValueError("Only one of `timesteps` or `sigmas` should be set.") + if num_inference_steps is None and timesteps is None and sigmas is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `timesteps` or `sigmas.") + if num_inference_steps is not None and (timesteps is not None or sigmas is not None): + raise ValueError("Can only pass one of `num_inference_steps` or `timesteps` or `sigmas`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_karras_sigmas = True`.") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + if ( + timesteps is not None + and self.config.timestep_type == "continuous" + and self.config.prediction_type == "v_prediction" + ): + raise ValueError( + "Cannot set `timesteps` with `config.timestep_type = 'continuous'` and `config.prediction_type = 'v_prediction'`." + ) + + if num_inference_steps is None: + num_inference_steps = len(timesteps) if timesteps is not None else len(sigmas) - 1 + self.num_inference_steps = num_inference_steps + + if sigmas is not None: + log_sigmas = np.log(np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)) + sigmas = np.array(sigmas).astype(np.float32) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas[:-1]]) + + else: + if timesteps is not None: + timesteps = np.array(timesteps).astype(np.float32) + else: + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace( + 0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32 + )[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + ) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = ( + (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + ) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.interpolation_type == "linear": + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + elif self.config.interpolation_type == "log_linear": + sigmas = torch.linspace(np.log(sigmas[-1]), np.log(sigmas[0]), num_inference_steps + 1).exp().numpy() + else: + raise ValueError( + f"{self.config.interpolation_type} is not implemented. Please specify interpolation_type to either" + " 'linear' or 'log_linear'" + ) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + + # TODO: Support the full EDM scalings for all prediction types and timestep types + if self.config.timestep_type == "continuous" and self.config.prediction_type == "v_prediction": + self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas[:-1]]).to(device=device) + else: + self.timesteps = torch.from_numpy(timesteps.astype(np.float32)).to(device=device) + + self._step_index = None + self._begin_index = None + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17 + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L26 + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> EulerDiscreteSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`, *optional*, defaults to `0.0`): + Stochasticity parameter that controls the amount of noise added during sampling. Higher values increase + randomness. + s_tmin (`float`, *optional*, defaults to `0.0`): + Minimum timestep threshold for applying stochasticity. Only timesteps above this value will have noise + added. + s_tmax (`float`, *optional*, defaults to `inf`): + Maximum timestep threshold for applying stochasticity. Only timesteps below this value will have noise + added. + s_noise (`float`, *optional*, defaults to `1.0`): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible sampling. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`: + If `return_dict` is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor and the second + element is the predicted original sample. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if not self.is_scale_input_called: + logger.warning( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + + gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 + + sigma_hat = sigma * (gamma + 1) + + if gamma > 0: + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + eps = noise * s_noise + sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5 + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + # NOTE: "original_sample" should not be an expected prediction_type but is left in for + # backwards compatibility + if self.config.prediction_type == "original_sample" or self.config.prediction_type == "sample": + pred_original_sample = model_output + elif self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma_hat * model_output + elif self.config.prediction_type == "v_prediction": + # denoised = model_output * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma_hat + + dt = self.sigmas[self.step_index + 1] - sigma_hat + + prev_sample = sample + derivative * dt + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return EulerDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.Tensor) -> torch.Tensor: + """ + Compute the velocity prediction for the given sample and noise at the specified timesteps. + + This method implements the velocity prediction used in v-prediction models, which predicts a linear combination + of the sample and noise. + + Args: + sample (`torch.Tensor`): + The input sample for which to compute the velocity. + noise (`torch.Tensor`): + The noise tensor corresponding to the sample. + timesteps (`torch.Tensor`): + The timesteps at which to compute the velocity. + + Returns: + `torch.Tensor`: + The velocity prediction computed as `sqrt(alpha_prod) * noise - sqrt(1 - alpha_prod) * sample`. + """ + if ( + isinstance(timesteps, int) + or isinstance(timesteps, torch.IntTensor) + or isinstance(timesteps, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.get_velocity()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if sample.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) + timesteps = timesteps.to(sample.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(sample.device) + timesteps = timesteps.to(sample.device) + + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + alphas_cumprod = self.alphas_cumprod.to(sample) + sqrt_alpha_prod = alphas_cumprod[step_indices] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[step_indices]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_flow_map_euler_discrete.py b/diffusers/schedulers/scheduling_flow_map_euler_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..25f8c2796beb67d65f634aa96257bf28b4d18132 --- /dev/null +++ b/diffusers/schedulers/scheduling_flow_map_euler_discrete.py @@ -0,0 +1,308 @@ +# Copyright 2026 The AnyFlow Team, NVIDIA Corp., and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass +from typing import List, Optional, Tuple, Union + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMapEulerDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor`): + Computed sample :math:`z_r` at the target flow-map timestep `r_timestep`. Should be used as the next + denoising input. + """ + + prev_sample: torch.Tensor + + +class FlowMapEulerDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Euler-style sampler for flow-map-distilled diffusion models. + + Flow-map models learn arbitrary-interval transitions :math:`z_t \\to z_r` rather than the fixed :math:`z_t \\to + z_0` mapping of consistency models, so a single distilled checkpoint can be evaluated at 1, 2, 4, 8, ... NFE + without retraining. The `step` method advances the sample from `timestep` to `r_timestep` along the predicted + velocity. + + Introduced in [AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map + Distillation](https://huggingface.co/papers/2605.13724) by Yuchao Gu, Guian Fang et al. + + This scheduler inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the + generic methods implemented for all schedulers (loading, saving, etc.). + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps used to train the underlying flow-map model. + shift (`float`, defaults to 1.0): + Multiplicative timestep shift applied to the inference schedule. ``shift=1.0`` is the identity; values + greater than 1.0 push the schedule toward more denoising at later steps (e.g., ``shift=5`` matches the + Wan2.1 default). + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + ): + # `_step_index` and `_begin_index` mirror `FlowMatchEulerDiscreteScheduler`'s state machine: + # `_step_index` advances on every `step()` so callbacks and composable schedulers can read it; + # `_begin_index` is honoured on the very first `step()` after `set_timesteps` to support + # mid-schedule restarts (e.g. image-to-image style use). + self._step_index: Optional[int] = None + self._begin_index: Optional[int] = None + self.set_timesteps(num_train_timesteps, device="cpu") + + @property + def step_index(self) -> Optional[int]: + """The index counter for current timestep. Returns ``None`` before the first :meth:`step` call after + :meth:`set_timesteps`.""" + return self._step_index + + @property + def begin_index(self) -> Optional[int]: + """The index for the first timestep — set by :meth:`set_begin_index`. Defaults to ``None``.""" + return self._begin_index + + def set_begin_index(self, begin_index: int = 0): + """Set the begin index for the scheduler. Pipelines that start mid-schedule (e.g. image-to-image) + call this between :meth:`set_timesteps` and the first :meth:`step` to anchor the rollout.""" + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """No-op identity scaling. Provided for API compatibility with other Diffusers schedulers.""" + return sample + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + noise: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + """Linearly interpolate ``sample`` toward ``noise`` according to the normalized ``timestep``.""" + timestep = timestep.to(device=sample.device, dtype=sample.dtype) + + timestep = timestep / self.config.num_train_timesteps + timestep = timestep.view(*timestep.shape, *([1] * (noise.ndim - timestep.ndim))) + sample = timestep * noise + (1.0 - timestep) * sample + return sample + + def apply_shift(self, sigmas: torch.Tensor) -> torch.Tensor: + """Apply the configured shift transformation to a sigma tensor.""" + if self.config.shift == 1.0: + return sigmas + return self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas) + + def set_timesteps( + self, + num_inference_steps: Optional[int] = None, + device: Union[str, torch.device] = None, + sigmas: Optional[List[float]] = None, + timesteps: Optional[List[float]] = None, + ) -> None: + """Build the inference timestep schedule. + + Internally tracks ``self.sigmas`` of length ``num_inference_steps + 1`` (the configured shift applied to a + linspace from ``1.0`` to ``0.0`` by default); ``self.timesteps`` exposes the first ``num_inference_steps`` + sigmas scaled by ``num_train_timesteps`` — i.e. one timestep per inference step, matching + :class:`~diffusers.schedulers.FlowMatchEulerDiscreteScheduler`. The final sigma (``0``) is the implicit + r-endpoint of the last step and is appended automatically when ``sigmas`` / ``timesteps`` are user-provided. + + Args: + num_inference_steps (`int`, *optional*): + Number of inference steps. If ``None``, must pass ``sigmas`` or ``timesteps``. + device (`str` or `torch.device`, *optional*): + Target device for ``self.sigmas`` / ``self.timesteps``. + sigmas (`List[float]`, *optional*): + Custom sigma schedule of length ``num_inference_steps``. The terminal ``0`` sigma is appended + automatically. The configured ``shift`` is applied on top. + timesteps (`List[float]`, *optional*): + Custom timestep schedule of length ``num_inference_steps``, in the same units as ``self.timesteps`` + (i.e. scaled by ``num_train_timesteps``). Converted to sigmas internally. If both ``sigmas`` and + ``timesteps`` are passed, their lengths must match. + """ + if sigmas is not None and timesteps is not None and len(sigmas) != len(timesteps): + raise ValueError("`sigmas` and `timesteps` should have the same length") + + if num_inference_steps is not None: + if (sigmas is not None and len(sigmas) != num_inference_steps) or ( + timesteps is not None and len(timesteps) != num_inference_steps + ): + raise ValueError( + "`sigmas` and `timesteps` should have the same length as `num_inference_steps` when both are provided" + ) + elif sigmas is not None: + num_inference_steps = len(sigmas) + elif timesteps is not None: + num_inference_steps = len(timesteps) + else: + raise ValueError("`num_inference_steps` must be provided when both `sigmas` and `timesteps` are `None`") + + # MPS / NPU don't support float64 — build the schedule in float64 on CPU and only move + # the final tensors to the requested device (with a float32 downcast for MPS / NPU). + device_obj = torch.device(device) if device is not None and not isinstance(device, torch.device) else device + is_mps = device_obj is not None and device_obj.type == "mps" + is_npu = device_obj is not None and device_obj.type == "npu" + out_dtype = torch.float32 if (is_mps or is_npu) else torch.float64 + + # Build the working sigma sequence (length N) before appending the terminal 0. + if sigmas is not None: + working_sigmas = torch.tensor(sigmas, dtype=torch.float64) + elif timesteps is not None: + working_sigmas = torch.tensor(timesteps, dtype=torch.float64) / self.config.num_train_timesteps + else: + working_sigmas = torch.linspace(1.0, 0.0, num_inference_steps + 1, dtype=torch.float64)[:-1] + + working_sigmas = self.apply_shift(working_sigmas) + # Append the terminal 0 sigma as the r-endpoint of the last step. `new_zeros(1)` inherits both + # device and dtype from `working_sigmas` so `torch.cat` stays device-consistent on CUDA. + full_sigmas = torch.cat([working_sigmas, working_sigmas.new_zeros(1)]) + + self.num_inference_steps = num_inference_steps + self.sigmas = full_sigmas.to(device=device, dtype=out_dtype) + self.timesteps = (self.sigmas[:-1] * self.config.num_train_timesteps).to(device=device, dtype=out_dtype) + # Reset the state machine — first `step()` after this will re-initialize `_step_index`. + self._step_index = None + self._begin_index = None + + def _init_step_index(self, timestep: Union[float, torch.FloatTensor]) -> None: + """Initialize ``self._step_index`` on the first :meth:`step` call after :meth:`set_timesteps`. + + Off-schedule timesteps are allowed (any-step sampling is documented in :meth:`step`); in that case the counter + starts at 0 so it can still be used as an observable rollout marker. + """ + if self._begin_index is not None: + self._step_index = self._begin_index + return + idx = self.index_for_timestep(timestep) + self._step_index = idx if idx is not None else 0 + + def index_for_timestep(self, timestep: Union[float, torch.FloatTensor]) -> Optional[int]: + """Return the index of ``timestep`` on the current schedule, or ``None`` if off-schedule. + + Lookup is done against ``self.timesteps`` with a small fp tolerance. Used to recover the corresponding sigma + without assuming the linear ``timesteps = sigmas * num_train_timesteps`` relationship — that way a custom + schedule (e.g. non-linear shift, manually-set timesteps) still resolves correctly. + """ + if self.timesteps is None: + return None + t_value = float(timestep.flatten()[0].item()) if torch.is_tensor(timestep) else float(timestep) + diffs = (self.timesteps.float() - t_value).abs() + idx = int(diffs.argmin().item()) + if diffs[idx].item() > 1e-3: + return None + return idx + + def step( + self, + model_output: torch.FloatTensor, + timestep: Union[float, torch.FloatTensor], + sample: torch.FloatTensor, + r_timestep: Optional[Union[float, torch.FloatTensor]] = None, + return_dict: bool = True, + ) -> Union[FlowMapEulerDiscreteSchedulerOutput, Tuple[torch.Tensor]]: + """ + Advance ``sample`` from ``timestep`` to ``r_timestep`` using the model-predicted velocity. + + Unlike a standard Euler scheduler, both endpoints of the interval can be caller-provided so that any-step + sampling is possible: a single model call can step from `t` to any chosen target `r` (including `r=0` for a + one-shot generation). When ``r_timestep`` is omitted, it defaults to the next timestep on the schedule + (matching ``FlowMatchEulerDiscreteScheduler`` semantics). + + Internally the source and target sigmas are recovered by indexing ``self.sigmas`` via + :meth:`index_for_timestep` rather than by dividing the input timesteps by ``num_train_timesteps``, so any + schedule whose timestep / sigma relationship is non-linear (for example a custom shift) stays correct. For an + off-schedule ``r_timestep``, the scheduler falls back to ``r_timestep / num_train_timesteps`` so any-step + sampling outside the schedule remains supported. + + Args: + model_output (`torch.Tensor`): + Direct output from the flow-map model (predicted mean velocity). + timestep (`float` or `torch.Tensor`): + Source timestep ``t`` in the same units as ``self.timesteps``. + sample (`torch.Tensor`): + Current sample :math:`z_t`. + r_timestep (`float` or `torch.Tensor`, *optional*): + Target timestep ``r``. Defaults to the next timestep on the schedule when ``None``; pass an explicit + value for any-step sampling. ``r_timestep == timestep`` is a no-op. + return_dict (`bool`, defaults to `True`): + Whether to return a [`FlowMapEulerDiscreteSchedulerOutput`] (the default) or a plain tuple. + + Returns: + [`FlowMapEulerDiscreteSchedulerOutput`] or `tuple`: + When ``return_dict=True``, returns a [`FlowMapEulerDiscreteSchedulerOutput`] whose ``prev_sample`` is + :math:`z_r`. Otherwise returns a 1-tuple ``(prev_sample,)``. + """ + if self.sigmas is None or self.timesteps is None: + raise ValueError("`set_timesteps` has not been called.") + + # `_step_index` is maintained purely as observable state for callbacks / composable schedulers. + # Sigma resolution stays a pure function of the passed-in (`timestep`, `r_timestep`) so the call is + # idempotent — calling `step` twice with the same arguments always returns the same `prev_sample`. + if self._step_index is None: + self._init_step_index(timestep) + + # Resolve source sigma via index lookup; fall back to / num_train_timesteps only if `timestep` is off-schedule. + t_idx = self.index_for_timestep(timestep) + if t_idx is not None: + sigma_t = self.sigmas[t_idx].to(device=sample.device, dtype=self.sigmas.dtype) + else: + t_value = timestep.to(self.sigmas.dtype) if torch.is_tensor(timestep) else torch.tensor(timestep) + sigma_t = (t_value / self.config.num_train_timesteps).to(device=sample.device, dtype=self.sigmas.dtype) + + # Resolve target sigma. None defaults to sigmas[t_idx + 1] when on-schedule; otherwise the caller's + # explicit `r_timestep` is used (sigma lookup first, fall back to scaling for off-schedule any-step). + if r_timestep is None: + if t_idx is None: + raise ValueError( + "`r_timestep` is None but `timestep` is not on the current schedule, so `r` cannot be inferred. " + "Please pass an explicit `r_timestep` for any-step sampling outside the schedule." + ) + sigma_r = self.sigmas[t_idx + 1].to(device=sample.device, dtype=self.sigmas.dtype) + else: + r_idx = self.index_for_timestep(r_timestep) + if r_idx is not None: + sigma_r = self.sigmas[r_idx].to(device=sample.device, dtype=self.sigmas.dtype) + else: + r_value = r_timestep.to(self.sigmas.dtype) if torch.is_tensor(r_timestep) else torch.tensor(r_timestep) + sigma_r = (r_value / self.config.num_train_timesteps).to(device=sample.device, dtype=self.sigmas.dtype) + + sigma_t = sigma_t.view(*sigma_t.shape, *([1] * (model_output.ndim - sigma_t.ndim))) + sigma_r = sigma_r.view(*sigma_r.shape, *([1] * (model_output.ndim - sigma_r.ndim))) + prev_sample = sample - (sigma_t - sigma_r) * model_output + prev_sample = prev_sample.to(model_output.dtype) + + # Advance state machine so downstream callbacks / composable schedulers observe correct `step_index`. + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return FlowMapEulerDiscreteSchedulerOutput(prev_sample=prev_sample) diff --git a/diffusers/schedulers/scheduling_flow_match_euler_discrete.py b/diffusers/schedulers/scheduling_flow_match_euler_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..0e4f5c6a1f97c48be6f05f6b4a671cf5ce1028a8 --- /dev/null +++ b/diffusers/schedulers/scheduling_flow_match_euler_discrete.py @@ -0,0 +1,653 @@ +# Copyright 2025 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal, Optional, Union + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMatchEulerDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.FloatTensor + + +class FlowMatchEulerDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Euler scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + shift (`float`, defaults to 1.0): + The shift value for the timestep schedule. + use_dynamic_shifting (`bool`, defaults to False): + Whether to apply timestep shifting on-the-fly based on the image resolution. + base_shift (`float`, defaults to 0.5): + Value to stabilize image generation. Increasing `base_shift` reduces variation and image is more consistent + with desired output. + max_shift (`float`, defaults to 1.15): + Value change allowed to latent vectors. Increasing `max_shift` encourages more variation and image may be + more exaggerated or stylized. + base_image_seq_len (`int`, defaults to 256): + The base image sequence length. + max_image_seq_len (`int`, defaults to 4096): + The maximum image sequence length. + invert_sigmas (`bool`, defaults to False): + Whether to invert the sigmas. + shift_terminal (`float`, defaults to None): + The end value of the shifted timestep schedule. + use_karras_sigmas (`bool`, defaults to False): + Whether to use Karras sigmas for step sizes in the noise schedule during sampling. + use_exponential_sigmas (`bool`, defaults to False): + Whether to use exponential sigmas for step sizes in the noise schedule during sampling. + use_beta_sigmas (`bool`, defaults to False): + Whether to use beta sigmas for step sizes in the noise schedule during sampling. + time_shift_type (`str`, defaults to "exponential"): + The type of dynamic resolution-dependent timestep shifting to apply. Either "exponential" or "linear". + stochastic_sampling (`bool`, defaults to False): + Whether to use stochastic sampling. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + use_dynamic_shifting: bool = False, + base_shift: float | None = 0.5, + max_shift: float | None = 1.15, + base_image_seq_len: int = 256, + max_image_seq_len: int = 4096, + invert_sigmas: bool = False, + shift_terminal: float = None, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + time_shift_type: Literal["exponential", "linear"] = "exponential", + stochastic_sampling: bool = False, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if ( + sum( + [ + self.config.use_beta_sigmas, + self.config.use_exponential_sigmas, + self.config.use_karras_sigmas, + ] + ) + > 1 + ): + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if time_shift_type not in {"exponential", "linear"}: + raise ValueError("`time_shift_type` must either be 'exponential' or 'linear'.") + + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + sigmas = timesteps / num_train_timesteps + if not use_dynamic_shifting: + # when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + + self._shift = shift + + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + + @property + def shift(self): + """ + The value used for shifting. + """ + return self._shift + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_shift(self, shift: float): + """ + Sets the shift value for the scheduler. + + Args: + shift (`float`): + The shift value to be set. + """ + self._shift = shift + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: float | torch.FloatTensor, + noise: torch.FloatTensor | None = None, + ) -> torch.FloatTensor: + """ + Forward process in flow-matching + + Args: + sample (`torch.FloatTensor`): + The input sample. + timestep (`torch.FloatTensor`): + The current timestep in the diffusion chain. + noise (`torch.FloatTensor`): + The noise tensor. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=sample.device, dtype=sample.dtype) + + if sample.device.type == "mps" and torch.is_floating_point(timestep): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) + timestep = timestep.to(sample.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(sample.device) + timestep = timestep.to(sample.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timestep] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timestep.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timestep.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(sample.shape): + sigma = sigma.unsqueeze(-1) + + sample = sigma * noise + (1.0 - sigma) * sample + + return sample + + def _sigma_to_t(self, sigma) -> float: + return sigma * self.config.num_train_timesteps + + def time_shift(self, mu: float, sigma: float, t: torch.Tensor) -> torch.Tensor: + """ + Apply time shifting to the sigmas. + + Args: + mu (`float`): + The mu parameter for the time shift. + sigma (`float`): + The sigma parameter for the time shift. + t (`torch.Tensor`): + The input timesteps. + + Returns: + `torch.Tensor`: + The time-shifted timesteps. + """ + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor: + r""" + Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config + value. + + Reference: + https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51 + + Args: + t (`torch.Tensor`): + A tensor of timesteps to be stretched and shifted. + + Returns: + `torch.Tensor`: + A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`. + """ + one_minus_z = 1 - t + scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal) + stretched_t = 1 - (one_minus_z / scale_factor) + return stretched_t + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device = None, + sigmas: list[float] | None = None, + mu: float | None = None, + timesteps: list[float] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`list[float]`, *optional*): + Custom values for sigmas to be used for each diffusion step. If `None`, the sigmas are computed + automatically. + mu (`float`, *optional*): + Determines the amount of shifting applied to sigmas when performing resolution-dependent timestep + shifting. + timesteps (`list[float]`, *optional*): + Custom values for timesteps to be used for each diffusion step. If `None`, the timesteps are computed + automatically. + """ + if self.config.use_dynamic_shifting and mu is None: + raise ValueError("`mu` must be passed when `use_dynamic_shifting` is set to be `True`") + + if sigmas is not None and timesteps is not None: + if len(sigmas) != len(timesteps): + raise ValueError("`sigmas` and `timesteps` should have the same length") + + if num_inference_steps is not None: + if (sigmas is not None and len(sigmas) != num_inference_steps) or ( + timesteps is not None and len(timesteps) != num_inference_steps + ): + raise ValueError( + "`sigmas` and `timesteps` should have the same length as num_inference_steps, if `num_inference_steps` is provided" + ) + else: + num_inference_steps = len(sigmas) if sigmas is not None else len(timesteps) + + self.num_inference_steps = num_inference_steps + + # 1. Prepare default sigmas + is_timesteps_provided = timesteps is not None + + if is_timesteps_provided: + timesteps = np.array(timesteps).astype(np.float32) + + if sigmas is None: + if timesteps is None: + timesteps = np.linspace( + self._sigma_to_t(self.sigma_max), + self._sigma_to_t(self.sigma_min), + num_inference_steps, + ) + sigmas = timesteps / self.config.num_train_timesteps + else: + sigmas = np.array(sigmas).astype(np.float32) + num_inference_steps = len(sigmas) + + # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of + # "exponential" or "linear" type is applied + if self.config.use_dynamic_shifting: + sigmas = self.time_shift(mu, 1.0, sigmas) + else: + sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas) + + # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value + if self.config.shift_terminal: + sigmas = self.stretch_shift_to_terminal(sigmas) + + # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + + # 5. Convert sigmas and timesteps to tensors and move to specified device + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + timesteps = sigmas * self.config.num_train_timesteps + + # 6. Append the terminal sigma value. + # If a model requires inverted sigma schedule for denoising but timesteps without inversion, the + # `invert_sigmas` flag can be set to `True`. This case is only required in Mochi + if self.config.invert_sigmas: + sigmas = 1.0 - sigmas + timesteps = sigmas * self.config.num_train_timesteps + sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)]) + else: + sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + self.timesteps = timesteps + self.sigmas = sigmas + self._step_index = None + self._begin_index = None + + def index_for_timestep( + self, + timestep: Union[float, torch.FloatTensor], + schedule_timesteps: Optional[torch.FloatTensor] = None, + ) -> int: + """ + Get the index for the given timestep. + + Args: + timestep (`float` or `torch.FloatTensor`): + The timestep to find the index for. + schedule_timesteps (`torch.FloatTensor`, *optional*): + The schedule timesteps to validate against. If `None`, the scheduler's timesteps are used. + + Returns: + `int`: + The index of the timestep. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep: Union[float, torch.FloatTensor]) -> None: + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor, + sample: torch.FloatTensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: torch.Generator | None = None, + per_token_timesteps: torch.Tensor | None = None, + return_dict: bool = True, + ) -> FlowMatchEulerDiscreteSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`): + s_tmin (`float`): + s_tmax (`float`): + s_noise (`float`, defaults to 1.0): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator. + per_token_timesteps (`torch.Tensor`, *optional*): + The timesteps for each token in the sample. + return_dict (`bool`, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + """ + + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `FlowMatchEulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + if per_token_timesteps is not None: + per_token_sigmas = per_token_timesteps / self.config.num_train_timesteps + + sigmas = self.sigmas[:, None, None] + lower_mask = sigmas < per_token_sigmas[None] - 1e-6 + lower_sigmas = lower_mask * sigmas + lower_sigmas, _ = lower_sigmas.max(dim=0) + + current_sigma = per_token_sigmas[..., None] + next_sigma = lower_sigmas[..., None] + dt = current_sigma - next_sigma + else: + sigma_idx = self.step_index + sigma = self.sigmas[sigma_idx] + sigma_next = self.sigmas[sigma_idx + 1] + + current_sigma = sigma + next_sigma = sigma_next + dt = sigma_next - sigma + + if self.config.stochastic_sampling: + x0 = sample - current_sigma * model_output + noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype) + prev_sample = (1.0 - next_sigma) * x0 + next_sigma * noise + else: + prev_sample = sample + dt * model_output + + # upon completion increase step index by one + self._step_index += 1 + if per_token_timesteps is None: + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + if not return_dict: + return (prev_sample,) + + return FlowMatchEulerDiscreteSchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def _time_shift_exponential(self, mu: float, sigma: float, t: torch.Tensor) -> torch.Tensor: + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + def _time_shift_linear(self, mu: float, sigma: float, t: torch.Tensor) -> torch.Tensor: + return mu / (mu + (1 / t - 1) ** sigma) + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_flow_match_heun_discrete.py b/diffusers/schedulers/scheduling_flow_match_heun_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..49892e043abd438d2a73a5e8d85bdc495e101815 --- /dev/null +++ b/diffusers/schedulers/scheduling_flow_match_heun_discrete.py @@ -0,0 +1,355 @@ +# Copyright 2025 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMatchHeunDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.FloatTensor + + +class FlowMatchHeunDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Heun scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + shift (`float`, defaults to 1.0): + The shift value for the timestep schedule. + """ + + _compatibles = [] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + ): + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + sigmas = timesteps / num_train_timesteps + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: float | torch.FloatTensor, + noise: torch.FloatTensor, + ) -> torch.FloatTensor: + """ + Forward process in flow-matching + + Args: + sample (`torch.FloatTensor`): + The input sample. + timestep (`float` or `torch.FloatTensor`): + The current timestep in the diffusion chain. + noise (`torch.FloatTensor`): + The noise tensor. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + sample = sigma * noise + (1.0 - sigma) * sample + + return sample + + def _sigma_to_t(self, sigma: float) -> float: + return sigma * self.config.num_train_timesteps + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + timesteps = np.linspace( + self._sigma_to_t(self.sigma_max), + self._sigma_to_t(self.sigma_min), + num_inference_steps, + ) + + sigmas = timesteps / self.config.num_train_timesteps + sigmas = self.config.shift * sigmas / (1 + (self.config.shift - 1) * sigmas) + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) + + timesteps = sigmas * self.config.num_train_timesteps + timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)]) + self.timesteps = timesteps.to(device=device) + + sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]]) + + # empty dt and derivative + self.prev_derivative = None + self.dt = None + + self._step_index = None + self._begin_index = None + + def index_for_timestep( + self, + timestep: float | torch.FloatTensor, + schedule_timesteps: torch.FloatTensor | None = None, + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.FloatTensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.FloatTensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep: float | torch.FloatTensor) -> None: + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def state_in_first_order(self) -> bool: + """ + Returns whether the scheduler is in the first-order state. + """ + return self.dt is None + + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor, + sample: torch.FloatTensor, + s_churn: float = 0.0, + s_tmin: float = 0.0, + s_tmax: float = float("inf"), + s_noise: float = 1.0, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> FlowMatchHeunDiscreteSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float` or `torch.FloatTensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + s_churn (`float`): + Stochasticity parameter that controls the amount of noise added during sampling. Higher values increase + randomness. + s_tmin (`float`): + Minimum timestep threshold for applying stochasticity. Only timesteps above this value will have noise + added. + s_tmax (`float`): + Maximum timestep threshold for applying stochasticity. Only timesteps below this value will have noise + added. + s_noise (`float`, defaults to 1.0): + Scaling factor for noise added to the sample. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a + [`~schedulers.scheduling_flow_match_heun_discrete.FlowMatchHeunDiscreteSchedulerOutput`] tuple. + + Returns: + [`~schedulers.scheduling_flow_match_heun_discrete.FlowMatchHeunDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, + [`~schedulers.scheduling_flow_match_heun_discrete.FlowMatchHeunDiscreteSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the sample tensor. + """ + + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `FlowMatchHeunDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order / Heun's method + sigma = self.sigmas[self.step_index - 1] + sigma_next = self.sigmas[self.step_index] + + gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0 + + sigma_hat = sigma * (gamma + 1) + + if gamma > 0: + noise = randn_tensor( + model_output.shape, + dtype=model_output.dtype, + device=model_output.device, + generator=generator, + ) + eps = noise * s_noise + sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5 + + if self.state_in_first_order: + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + denoised = sample - model_output * sigma + # 2. convert to an ODE derivative for 1st order + derivative = (sample - denoised) / sigma_hat + # 3. Delta timestep + dt = sigma_next - sigma_hat + + # store for 2nd order step + self.prev_derivative = derivative + self.dt = dt + self.sample = sample + else: + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + denoised = sample - model_output * sigma_next + # 2. 2nd order / Heun's method + derivative = (sample - denoised) / sigma_next + derivative = 0.5 * (self.prev_derivative + derivative) + + # 3. take prev timestep & sample + dt = self.dt + sample = self.sample + + # free dt and derivative + # Note, this puts the scheduler in "first order mode" + self.prev_derivative = None + self.dt = None + self.sample = None + + prev_sample = sample + derivative * dt + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return FlowMatchHeunDiscreteSchedulerOutput(prev_sample=prev_sample) + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_flow_match_lcm.py b/diffusers/schedulers/scheduling_flow_match_lcm.py new file mode 100644 index 0000000000000000000000000000000000000000..97d4ebbc8e427bd528303bd91d35a3204e9941e2 --- /dev/null +++ b/diffusers/schedulers/scheduling_flow_match_lcm.py @@ -0,0 +1,640 @@ +# Copyright 2025 Stability AI, Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class FlowMatchLCMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class FlowMatchLCMScheduler(SchedulerMixin, ConfigMixin): + """ + LCM scheduler for Flow Matching. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + shift (`float`, defaults to 1.0): + The shift value for the timestep schedule. + use_dynamic_shifting (`bool`, defaults to False): + Whether to apply timestep shifting on-the-fly based on the image resolution. + base_shift (`float`, defaults to 0.5): + Value to stabilize image generation. Increasing `base_shift` reduces variation and image is more consistent + with desired output. + max_shift (`float`, defaults to 1.15): + Value change allowed to latent vectors. Increasing `max_shift` encourages more variation and image may be + more exaggerated or stylized. + base_image_seq_len (`int`, defaults to 256): + The base image sequence length. + max_image_seq_len (`int`, defaults to 4096): + The maximum image sequence length. + invert_sigmas (`bool`, defaults to False): + Whether to invert the sigmas. + shift_terminal (`float`, defaults to None): + The end value of the shifted timestep schedule. + use_karras_sigmas (`bool`, defaults to False): + Whether to use Karras sigmas for step sizes in the noise schedule during sampling. + use_exponential_sigmas (`bool`, defaults to False): + Whether to use exponential sigmas for step sizes in the noise schedule during sampling. + use_beta_sigmas (`bool`, defaults to False): + Whether to use beta sigmas for step sizes in the noise schedule during sampling. + time_shift_type (`str`, defaults to "exponential"): + The type of dynamic resolution-dependent timestep shifting to apply. + scale_factors (`list[float]`, *optional*, defaults to `None`): + It defines how to scale the latents at which predictions are made. + upscale_mode (`str`, *optional*, defaults to "bicubic"): + Upscaling method, applied if scale-wise generation is considered. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, + use_dynamic_shifting: bool = False, + base_shift: float = 0.5, + max_shift: float = 1.15, + base_image_seq_len: int = 256, + max_image_seq_len: int = 4096, + invert_sigmas: bool = False, + shift_terminal: float | None = None, + use_karras_sigmas: bool | None = False, + use_exponential_sigmas: bool | None = False, + use_beta_sigmas: bool | None = False, + time_shift_type: Literal["exponential", "linear"] = "exponential", + scale_factors: list[float] | None = None, + upscale_mode: Literal[ + "nearest", + "linear", + "bilinear", + "bicubic", + "trilinear", + "area", + "nearest-exact", + ] = "bicubic", + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if ( + sum( + [ + self.config.use_beta_sigmas, + self.config.use_exponential_sigmas, + self.config.use_karras_sigmas, + ] + ) + > 1 + ): + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if time_shift_type not in {"exponential", "linear"}: + raise ValueError("`time_shift_type` must either be 'exponential' or 'linear'.") + + timesteps = np.linspace(1, num_train_timesteps, num_train_timesteps, dtype=np.float32)[::-1].copy() + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32) + + sigmas = timesteps / num_train_timesteps + if not use_dynamic_shifting: + # when use_dynamic_shifting is True, we apply the timestep shifting on the fly based on the image resolution + sigmas = shift * sigmas / (1 + (shift - 1) * sigmas) + + self.timesteps = sigmas * num_train_timesteps + + self._step_index = None + self._begin_index = None + + self._shift = shift + + self._init_size = None + self._scale_factors = scale_factors + self._upscale_mode = upscale_mode + + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + + @property + def shift(self): + """ + The value used for shifting. + """ + return self._shift + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_shift(self, shift: float) -> None: + self._shift = shift + + def set_scale_factors(self, scale_factors: list[float], upscale_mode: str) -> None: + """ + Sets scale factors for a scale-wise generation regime. + + Args: + scale_factors (`list[float]`): + The scale factors for each step. + upscale_mode (`str`): + Upscaling method. + """ + self._scale_factors = scale_factors + self._upscale_mode = upscale_mode + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: float | torch.FloatTensor, + noise: torch.FloatTensor | None = None, + ) -> torch.FloatTensor: + """ + Forward process in flow-matching + + Args: + sample (`torch.FloatTensor`): + The input sample. + timestep (`torch.FloatTensor`): + The current timestep in the diffusion chain. + noise (`torch.FloatTensor`): + The noise tensor. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=sample.device, dtype=sample.dtype) + + if sample.device.type == "mps" and torch.is_floating_point(timestep): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(sample.device, dtype=torch.float32) + timestep = timestep.to(sample.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(sample.device) + timestep = timestep.to(sample.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timestep] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timestep.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timestep.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(sample.shape): + sigma = sigma.unsqueeze(-1) + + sample = sigma * noise + (1.0 - sigma) * sample + + return sample + + def _sigma_to_t(self, sigma: float | torch.FloatTensor) -> float | torch.FloatTensor: + return sigma * self.config.num_train_timesteps + + def time_shift( + self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor + ) -> float | np.ndarray | torch.Tensor: + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + def stretch_shift_to_terminal(self, t: np.ndarray | torch.Tensor) -> np.ndarray | torch.Tensor: + r""" + Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config + value. + + Reference: + https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51 + + Args: + t (`torch.Tensor` or `np.ndarray`): + A tensor or numpy array of timesteps to be stretched and shifted. + + Returns: + `torch.Tensor` or `np.ndarray`: + A tensor or numpy array of adjusted timesteps such that the final value equals + `self.config.shift_terminal`. + """ + one_minus_z = 1 - t + scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal) + stretched_t = 1 - (one_minus_z / scale_factor) + return stretched_t + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + sigmas: list[float] | None = None, + mu: float | None = None, + timesteps: list[float] | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`list[float]`, *optional*): + Custom values for sigmas to be used for each diffusion step. If `None`, the sigmas are computed + automatically. + mu (`float`, *optional*): + Determines the amount of shifting applied to sigmas when performing resolution-dependent timestep + shifting. + timesteps (`list[float]`, *optional*): + Custom values for timesteps to be used for each diffusion step. If `None`, the timesteps are computed + automatically. + """ + if self.config.use_dynamic_shifting and mu is None: + raise ValueError("`mu` must be passed when `use_dynamic_shifting` is set to be `True`") + + if sigmas is not None and timesteps is not None: + if len(sigmas) != len(timesteps): + raise ValueError("`sigmas` and `timesteps` should have the same length") + + if num_inference_steps is not None: + if (sigmas is not None and len(sigmas) != num_inference_steps) or ( + timesteps is not None and len(timesteps) != num_inference_steps + ): + raise ValueError( + "`sigmas` and `timesteps` should have the same length as num_inference_steps, if `num_inference_steps` is provided" + ) + else: + num_inference_steps = len(sigmas) if sigmas is not None else len(timesteps) + + self.num_inference_steps = num_inference_steps + + # 1. Prepare default sigmas + is_timesteps_provided = timesteps is not None + + if is_timesteps_provided: + timesteps = np.array(timesteps).astype(np.float32) # type: ignore + + if sigmas is None: + if timesteps is None: + timesteps = np.linspace( # type: ignore + self._sigma_to_t(self.sigma_max), + self._sigma_to_t(self.sigma_min), + num_inference_steps, + ) + sigmas = timesteps / self.config.num_train_timesteps # type: ignore + else: + sigmas = np.array(sigmas).astype(np.float32) # type: ignore + num_inference_steps = len(sigmas) + + # 2. Perform timestep shifting. Either no shifting is applied, or resolution-dependent shifting of + # "exponential" or "linear" type is applied + if self.config.use_dynamic_shifting: + sigmas = self.time_shift(mu, 1.0, sigmas) # type: ignore + else: + sigmas = self.shift * sigmas / (1 + (self.shift - 1) * sigmas) # type: ignore + + # 3. If required, stretch the sigmas schedule to terminate at the configured `shift_terminal` value + if self.config.shift_terminal: + sigmas = self.stretch_shift_to_terminal(sigmas) # type: ignore + + # 4. If required, convert sigmas to one of karras, exponential, or beta sigma schedules + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) # type: ignore + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) # type: ignore + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) # type: ignore + + # 5. Convert sigmas and timesteps to tensors and move to specified device + sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device) # type: ignore + if not is_timesteps_provided: + timesteps = sigmas * self.config.num_train_timesteps # type: ignore + else: + timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32, device=device) # type: ignore + + # 6. Append the terminal sigma value. + # If a model requires inverted sigma schedule for denoising but timesteps without inversion, the + # `invert_sigmas` flag can be set to `True`. This case is only required in Mochi + if self.config.invert_sigmas: + sigmas = 1.0 - sigmas + timesteps = sigmas * self.config.num_train_timesteps + sigmas = torch.cat([sigmas, torch.ones(1, device=sigmas.device)]) + else: + sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)]) + + self.timesteps = timesteps + self.sigmas = sigmas + self._step_index = None + self._begin_index = None + + def index_for_timestep( + self, + timestep: float | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return int(indices[pos].item()) + + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor, + sample: torch.FloatTensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> FlowMatchLCMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_flow_match_lcm.FlowMatchLCMSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_flow_match_lcm.FlowMatchLCMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_flow_match_lcm.FlowMatchLCMSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `FlowMatchLCMScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self._scale_factors and self._upscale_mode and len(self.timesteps) != len(self._scale_factors) + 1: + raise ValueError( + "`_scale_factors` should have the same length as `timesteps` - 1, if `_scale_factors` are set." + ) + + if self._init_size is None or self.step_index is None: + self._init_size = model_output.size()[2:] + + if self.step_index is None: + self._init_step_index(timestep) + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + x0_pred = sample - sigma * model_output + + if self._scale_factors and self._upscale_mode: + if self._step_index < len(self._scale_factors): + size = [round(self._scale_factors[self._step_index] * size) for size in self._init_size] + x0_pred = torch.nn.functional.interpolate(x0_pred, size=size, mode=self._upscale_mode) + + noise = randn_tensor( + x0_pred.shape, + generator=generator, + device=x0_pred.device, + dtype=x0_pred.dtype, + ) + prev_sample = (1 - sigma_next) * x0_pred + sigma_next * noise + + # upon completion increase step index by one + self._step_index += 1 + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + if not return_dict: + return (prev_sample,) + + return FlowMatchLCMSchedulerOutput(prev_sample=prev_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def _time_shift_exponential( + self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor + ) -> float | np.ndarray | torch.Tensor: + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + def _time_shift_linear( + self, mu: float, sigma: float, t: float | np.ndarray | torch.Tensor + ) -> float | np.ndarray | torch.Tensor: + return mu / (mu + (1 / t - 1) ** sigma) + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_helios.py b/diffusers/schedulers/scheduling_helios.py new file mode 100644 index 0000000000000000000000000000000000000000..ed35245c9db322af5b75a1025f8502fdfba3503c --- /dev/null +++ b/diffusers/schedulers/scheduling_helios.py @@ -0,0 +1,867 @@ +# Copyright 2025 The Helios Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..schedulers.scheduling_utils import SchedulerMixin +from ..utils import BaseOutput, deprecate + + +@dataclass +class HeliosSchedulerOutput(BaseOutput): + prev_sample: torch.FloatTensor + model_outputs: torch.FloatTensor | None = None + last_sample: torch.FloatTensor | None = None + this_order: int | None = None + + +class HeliosScheduler(SchedulerMixin, ConfigMixin): + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, # Following Stable diffusion 3, + stages: int = 3, + stage_range: list = [0, 1 / 3, 2 / 3, 1], + gamma: float = 1 / 3, + # For UniPC + thresholding: bool = False, + prediction_type: str = "flow_prediction", + solver_order: int = 2, + predict_x0: bool = True, + solver_type: str = "bh2", + lower_order_final: bool = True, + disable_corrector: list[int] = [], + solver_p: SchedulerMixin = None, + use_flow_sigmas: bool = True, + scheduler_type: str = "unipc", # ["euler", "unipc"] + use_dynamic_shifting: bool = False, + time_shift_type: Literal["exponential", "linear"] = "exponential", + ): + self.timestep_ratios = {} # The timestep ratio for each stage + self.timesteps_per_stage = {} # The detailed timesteps per stage (fix max and min per stage) + self.sigmas_per_stage = {} # always uniform [1000, 0] + self.start_sigmas = {} # for start point / upsample renoise + self.end_sigmas = {} # for end point + self.ori_start_sigmas = {} + + # self.init_sigmas() + self.init_sigmas_for_each_stage() + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + self.gamma = gamma + + if solver_type not in ["bh1", "bh2"]: + if solver_type in ["midpoint", "heun", "logrho"]: + self.register_to_config(solver_type="bh2") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + self.predict_x0 = predict_x0 + self.model_outputs = [None] * solver_order + self.timestep_list = [None] * solver_order + self.lower_order_nums = 0 + self.disable_corrector = disable_corrector + self.solver_p = solver_p + self.last_sample = None + self._step_index = None + self._begin_index = None + + def init_sigmas(self): + """ + initialize the global timesteps and sigmas + """ + num_train_timesteps = self.config.num_train_timesteps + shift = self.config.shift + + alphas = np.linspace(1, 1 / num_train_timesteps, num_train_timesteps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(shift * sigmas / (1 + (shift - 1) * sigmas))[:-1].copy() + sigmas = torch.from_numpy(sigmas) + timesteps = (sigmas * num_train_timesteps).clone() + + self._step_index = None + self._begin_index = None + self.timesteps = timesteps + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + + def init_sigmas_for_each_stage(self): + """ + Init the timesteps for each stage + """ + self.init_sigmas() + + stage_distance = [] + stages = self.config.stages + training_steps = self.config.num_train_timesteps + stage_range = self.config.stage_range + + # Init the start and end point of each stage + for i_s in range(stages): + # To decide the start and ends point + start_indice = int(stage_range[i_s] * training_steps) + start_indice = max(start_indice, 0) + end_indice = int(stage_range[i_s + 1] * training_steps) + end_indice = min(end_indice, training_steps) + start_sigma = self.sigmas[start_indice].item() + end_sigma = self.sigmas[end_indice].item() if end_indice < training_steps else 0.0 + self.ori_start_sigmas[i_s] = start_sigma + + if i_s != 0: + ori_sigma = 1 - start_sigma + gamma = self.config.gamma + corrected_sigma = (1 / (math.sqrt(1 + (1 / gamma)) * (1 - ori_sigma) + ori_sigma)) * ori_sigma + # corrected_sigma = 1 / (2 - ori_sigma) * ori_sigma + start_sigma = 1 - corrected_sigma + + stage_distance.append(start_sigma - end_sigma) + self.start_sigmas[i_s] = start_sigma + self.end_sigmas[i_s] = end_sigma + + # Determine the ratio of each stage according to flow length + tot_distance = sum(stage_distance) + for i_s in range(stages): + if i_s == 0: + start_ratio = 0.0 + else: + start_ratio = sum(stage_distance[:i_s]) / tot_distance + if i_s == stages - 1: + end_ratio = 0.9999999999999999 + else: + end_ratio = sum(stage_distance[: i_s + 1]) / tot_distance + + self.timestep_ratios[i_s] = (start_ratio, end_ratio) + + # Determine the timesteps and sigmas for each stage + for i_s in range(stages): + timestep_ratio = self.timestep_ratios[i_s] + # timestep_max = self.timesteps[int(timestep_ratio[0] * training_steps)] + timestep_max = min(self.timesteps[int(timestep_ratio[0] * training_steps)], 999) + timestep_min = self.timesteps[min(int(timestep_ratio[1] * training_steps), training_steps - 1)] + timesteps = np.linspace(timestep_max, timestep_min, training_steps + 1) + self.timesteps_per_stage[i_s] = ( + timesteps[:-1] if isinstance(timesteps, torch.Tensor) else torch.from_numpy(timesteps[:-1]) + ) + stage_sigmas = np.linspace(0.999, 0, training_steps + 1) + self.sigmas_per_stage[i_s] = torch.from_numpy(stage_sigmas[:-1]) + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def _sigma_to_t(self, sigma): + return sigma * self.config.num_train_timesteps + + def set_timesteps( + self, + num_inference_steps: int, + stage_index: int | None = None, + device: str | torch.device = None, + sigmas: bool | None = None, + mu: bool | None = None, + is_amplify_first_chunk: bool = False, + ): + """ + Setting the timesteps and sigmas for each stage + """ + if self.config.scheduler_type == "dmd": + if is_amplify_first_chunk: + num_inference_steps = num_inference_steps * 2 + 1 + else: + num_inference_steps = num_inference_steps + 1 + + self.num_inference_steps = num_inference_steps + self.init_sigmas() + + if self.config.stages == 1: + if sigmas is None: + sigmas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)[:-1].astype( + np.float32 + ) + if self.config.shift != 1.0: + assert not self.config.use_dynamic_shifting + sigmas = self.time_shift(self.config.shift, 1.0, sigmas) + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = torch.from_numpy(sigmas) + else: + stage_timesteps = self.timesteps_per_stage[stage_index] + timesteps = np.linspace( + stage_timesteps[0].item(), + stage_timesteps[-1].item(), + num_inference_steps, + ) + + stage_sigmas = self.sigmas_per_stage[stage_index] + ratios = np.linspace(stage_sigmas[0].item(), stage_sigmas[-1].item(), num_inference_steps) + sigmas = torch.from_numpy(ratios) + + self.timesteps = torch.from_numpy(timesteps).to(device=device) + self.sigmas = torch.cat([sigmas, torch.zeros(1)]).to(device=device) + + self._step_index = None + self.reset_scheduler_history() + + if self.config.scheduler_type == "dmd": + self.timesteps = self.timesteps[:-1] + self.sigmas = torch.cat([self.sigmas[:-2], self.sigmas[-1:]]) + + if self.config.use_dynamic_shifting: + assert self.config.shift == 1.0 + self.sigmas = self.time_shift(mu, 1.0, self.sigmas) + if self.config.stages == 1: + self.timesteps = self.sigmas[:-1] * self.config.num_train_timesteps + else: + self.timesteps = self.timesteps_per_stage[stage_index].min() + self.sigmas[:-1] * ( + self.timesteps_per_stage[stage_index].max() - self.timesteps_per_stage[stage_index].min() + ) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler.time_shift + def time_shift(self, mu: float, sigma: float, t: torch.Tensor): + """ + Apply time shifting to the sigmas. + + Args: + mu (`float`): + The mu parameter for the time shift. + sigma (`float`): + The sigma parameter for the time shift. + t (`torch.Tensor`): + The input timesteps. + + Returns: + `torch.Tensor`: + The time-shifted timesteps. + """ + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler._time_shift_exponential + def _time_shift_exponential(self, mu, sigma, t): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler._time_shift_linear + def _time_shift_linear(self, mu, sigma, t): + return mu / (mu + (1 / t - 1) ** sigma) + + # ---------------------------------- Euler ---------------------------------- + def index_for_timestep(self, timestep, schedule_timesteps=None): + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def _init_step_index(self, timestep): + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step_euler( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor = None, + sample: torch.FloatTensor = None, + generator: torch.Generator | None = None, + sigma: torch.FloatTensor | None = None, + sigma_next: torch.FloatTensor | None = None, + return_dict: bool = True, + ) -> HeliosSchedulerOutput | tuple: + assert (sigma is None) == (sigma_next is None), "sigma and sigma_next must both be None or both be not None" + + if sigma is None and sigma_next is None: + if ( + isinstance(timestep, int) + or isinstance(timestep, torch.IntTensor) + or isinstance(timestep, torch.LongTensor) + ): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` as a timestep." + ), + ) + + if self.step_index is None: + self._step_index = 0 + + # Upcast to avoid precision issues when computing prev_sample + sample = sample.to(torch.float32) + + if sigma is None and sigma_next is None: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + + prev_sample = sample + (sigma_next - sigma) * model_output + + # Cast sample back to model compatible dtype + prev_sample = prev_sample.to(model_output.dtype) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return HeliosSchedulerOutput(prev_sample=prev_sample) + + # ---------------------------------- UniPC ---------------------------------- + def _sigma_to_alpha_sigma_t(self, sigma): + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = torch.clamp(sigma, min=1e-8) + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + sigma: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + r""" + Convert the model output to the corresponding type the UniPC algorithm needs. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + flag = False + if sigma is None: + flag = True + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + + if self.predict_x0: + if self.config.prediction_type == "epsilon": + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + if flag: + sigma_t = self.sigmas[self.step_index] + else: + sigma_t = sigma + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the UniPCMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + else: + if self.config.prediction_type == "epsilon": + return model_output + elif self.config.prediction_type == "sample": + epsilon = (sample - alpha_t * model_output) / sigma_t + return epsilon + elif self.config.prediction_type == "v_prediction": + epsilon = alpha_t * model_output + sigma_t * sample + return epsilon + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the UniPCMultistepScheduler." + ) + + def multistep_uni_p_bh_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + order: int = None, + sigma: torch.Tensor = None, + sigma_next: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the UniP (B(h) version). Alternatively, `self.solver_p` is used if is specified. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model at the current timestep. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + order (`int`): + The order of UniP at this timestep (corresponds to the *p* in UniPC-p). + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if order is None: + if len(args) > 2: + order = args[2] + else: + raise ValueError("missing `order` as a required keyword argument") + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + model_output_list = self.model_outputs + + s0 = self.timestep_list[-1] + m0 = model_output_list[-1] + x = sample + + if self.solver_p: + x_t = self.solver_p.step(model_output, s0, x).prev_sample + return x_t + + if sigma_next is None and sigma is None: + sigma_t, sigma_s0 = self.sigmas[self.step_index + 1], self.sigmas[self.step_index] + else: + sigma_t, sigma_s0 = sigma_next, sigma + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + h = lambda_t - lambda_s0 + device = sample.device + + rks = [] + D1s = [] + for i in range(1, order): + si = self.step_index - i + mi = model_output_list[-(i + 1)] + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + rk = (lambda_si - lambda_s0) / h + rks.append(rk) + D1s.append((mi - m0) / rk) + + rks.append(1.0) + rks = torch.tensor(rks, device=device) + + R = [] + b = [] + + hh = -h if self.predict_x0 else h + h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 + h_phi_k = h_phi_1 / hh - 1 + + factorial_i = 1 + + if self.config.solver_type == "bh1": + B_h = hh + elif self.config.solver_type == "bh2": + B_h = torch.expm1(hh) + else: + raise NotImplementedError() + + for i in range(1, order + 1): + R.append(torch.pow(rks, i - 1)) + b.append(h_phi_k * factorial_i / B_h) + factorial_i *= i + 1 + h_phi_k = h_phi_k / hh - 1 / factorial_i + + R = torch.stack(R) + b = torch.tensor(b, device=device) + + if len(D1s) > 0: + D1s = torch.stack(D1s, dim=1) # (B, K) + # for order 2, we use a simplified version + if order == 2: + rhos_p = torch.tensor([0.5], dtype=x.dtype, device=device) + else: + rhos_p = torch.linalg.solve(R[:-1, :-1], b[:-1]).to(device).to(x.dtype) + else: + D1s = None + + if self.predict_x0: + x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0 + if D1s is not None: + pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s) + else: + pred_res = 0 + x_t = x_t_ - alpha_t * B_h * pred_res + else: + x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0 + if D1s is not None: + pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s) + else: + pred_res = 0 + x_t = x_t_ - sigma_t * B_h * pred_res + + x_t = x_t.to(x.dtype) + return x_t + + def multistep_uni_c_bh_update( + self, + this_model_output: torch.Tensor, + *args, + last_sample: torch.Tensor = None, + this_sample: torch.Tensor = None, + order: int = None, + sigma_before: torch.Tensor = None, + sigma: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the UniC (B(h) version). + + Args: + this_model_output (`torch.Tensor`): + The model outputs at `x_t`. + this_timestep (`int`): + The current timestep `t`. + last_sample (`torch.Tensor`): + The generated sample before the last predictor `x_{t-1}`. + this_sample (`torch.Tensor`): + The generated sample after the last predictor `x_{t}`. + order (`int`): + The `p` of UniC-p at this step. The effective order of accuracy should be `order + 1`. + + Returns: + `torch.Tensor`: + The corrected sample tensor at the current timestep. + """ + this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None) + if last_sample is None: + if len(args) > 1: + last_sample = args[1] + else: + raise ValueError("missing `last_sample` as a required keyword argument") + if this_sample is None: + if len(args) > 2: + this_sample = args[2] + else: + raise ValueError("missing `this_sample` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if this_timestep is not None: + deprecate( + "this_timestep", + "1.0.0", + "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + model_output_list = self.model_outputs + + m0 = model_output_list[-1] + x = last_sample + x_t = this_sample + model_t = this_model_output + + if sigma_before is None and sigma is None: + sigma_t, sigma_s0 = self.sigmas[self.step_index], self.sigmas[self.step_index - 1] + else: + sigma_t, sigma_s0 = sigma, sigma_before + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + h = lambda_t - lambda_s0 + device = this_sample.device + + rks = [] + D1s = [] + for i in range(1, order): + si = self.step_index - (i + 1) + mi = model_output_list[-(i + 1)] + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + rk = (lambda_si - lambda_s0) / h + rks.append(rk) + D1s.append((mi - m0) / rk) + + rks.append(1.0) + rks = torch.tensor(rks, device=device) + + R = [] + b = [] + + hh = -h if self.predict_x0 else h + h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 + h_phi_k = h_phi_1 / hh - 1 + + factorial_i = 1 + + if self.config.solver_type == "bh1": + B_h = hh + elif self.config.solver_type == "bh2": + B_h = torch.expm1(hh) + else: + raise NotImplementedError() + + for i in range(1, order + 1): + R.append(torch.pow(rks, i - 1)) + b.append(h_phi_k * factorial_i / B_h) + factorial_i *= i + 1 + h_phi_k = h_phi_k / hh - 1 / factorial_i + + R = torch.stack(R) + b = torch.tensor(b, device=device) + + if len(D1s) > 0: + D1s = torch.stack(D1s, dim=1) + else: + D1s = None + + # for order 1, we use a simplified version + if order == 1: + rhos_c = torch.tensor([0.5], dtype=x.dtype, device=device) + else: + rhos_c = torch.linalg.solve(R, b).to(device).to(x.dtype) + + if self.predict_x0: + x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0 + if D1s is not None: + corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s) + else: + corr_res = 0 + D1_t = model_t - m0 + x_t = x_t_ - alpha_t * B_h * (corr_res + rhos_c[-1] * D1_t) + else: + x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0 + if D1s is not None: + corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s) + else: + corr_res = 0 + D1_t = model_t - m0 + x_t = x_t_ - sigma_t * B_h * (corr_res + rhos_c[-1] * D1_t) + x_t = x_t.to(x.dtype) + return x_t + + def step_unipc( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor = None, + sample: torch.Tensor = None, + return_dict: bool = True, + model_outputs: list = None, + timestep_list: list = None, + sigma_before: torch.Tensor = None, + sigma: torch.Tensor = None, + sigma_next: torch.Tensor = None, + cus_step_index: int = None, + cus_lower_order_num: int = None, + cus_this_order: int = None, + cus_last_sample: torch.Tensor = None, + ) -> HeliosSchedulerOutput | tuple: + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if cus_step_index is None: + if self.step_index is None: + self._step_index = 0 + else: + self._step_index = cus_step_index + + if cus_lower_order_num is not None: + self.lower_order_nums = cus_lower_order_num + + if cus_this_order is not None: + self.this_order = cus_this_order + + if cus_last_sample is not None: + self.last_sample = cus_last_sample + + use_corrector = ( + self.step_index > 0 and self.step_index - 1 not in self.disable_corrector and self.last_sample is not None + ) + + # Convert model output using the proper conversion method + model_output_convert = self.convert_model_output(model_output, sample=sample, sigma=sigma) + + if model_outputs is not None and timestep_list is not None: + self.model_outputs = model_outputs[:-1] + self.timestep_list = timestep_list[:-1] + + if use_corrector: + sample = self.multistep_uni_c_bh_update( + this_model_output=model_output_convert, + last_sample=self.last_sample, + this_sample=sample, + order=self.this_order, + sigma_before=sigma_before, + sigma=sigma, + ) + + if model_outputs is not None and timestep_list is not None: + model_outputs[-1] = model_output_convert + self.model_outputs = model_outputs[1:] + self.timestep_list = timestep_list[1:] + else: + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.timestep_list[i] = self.timestep_list[i + 1] + self.model_outputs[-1] = model_output_convert + self.timestep_list[-1] = timestep + + if self.config.lower_order_final: + this_order = min(self.config.solver_order, len(self.timesteps) - self.step_index) + else: + this_order = self.config.solver_order + self.this_order = min(this_order, self.lower_order_nums + 1) # warmup for multistep + assert self.this_order > 0 + + self.last_sample = sample + prev_sample = self.multistep_uni_p_bh_update( + model_output=model_output, # pass the original non-converted model output, in case solver-p is used + sample=sample, + order=self.this_order, + sigma=sigma, + sigma_next=sigma_next, + ) + + if cus_lower_order_num is None: + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + if cus_step_index is None: + self._step_index += 1 + + if not return_dict: + return (prev_sample, model_outputs, self.last_sample, self.this_order) + + return HeliosSchedulerOutput( + prev_sample=prev_sample, + model_outputs=model_outputs, + last_sample=self.last_sample, + this_order=self.this_order, + ) + + # ---------------------------------- Merge ---------------------------------- + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor = None, + sample: torch.FloatTensor = None, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> HeliosSchedulerOutput | tuple: + if self.config.scheduler_type == "euler": + return self.step_euler( + model_output=model_output, + timestep=timestep, + sample=sample, + generator=generator, + return_dict=return_dict, + ) + elif self.config.scheduler_type == "unipc": + return self.step_unipc( + model_output=model_output, + timestep=timestep, + sample=sample, + return_dict=return_dict, + ) + else: + raise NotImplementedError + + def reset_scheduler_history(self): + self.model_outputs = [None] * self.config.solver_order + self.timestep_list = [None] * self.config.solver_order + self.lower_order_nums = 0 + self.disable_corrector = self.config.disable_corrector + self.solver_p = self.config.solver_p + self.last_sample = None + self._step_index = None + self._begin_index = None + + def __len__(self): + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_helios_dmd.py b/diffusers/schedulers/scheduling_helios_dmd.py new file mode 100644 index 0000000000000000000000000000000000000000..1f4afa0e3128bee25bcc241e687d1a98f8d8fc09 --- /dev/null +++ b/diffusers/schedulers/scheduling_helios_dmd.py @@ -0,0 +1,331 @@ +# Copyright 2025 The Helios Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..schedulers.scheduling_utils import SchedulerMixin +from ..utils import BaseOutput + + +@dataclass +class HeliosDMDSchedulerOutput(BaseOutput): + prev_sample: torch.FloatTensor + model_outputs: torch.FloatTensor | None = None + last_sample: torch.FloatTensor | None = None + this_order: int | None = None + + +class HeliosDMDScheduler(SchedulerMixin, ConfigMixin): + _compatibles = [] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + shift: float = 1.0, # Following Stable diffusion 3, + stages: int = 3, + stage_range: list = [0, 1 / 3, 2 / 3, 1], + gamma: float = 1 / 3, + prediction_type: str = "flow_prediction", + use_flow_sigmas: bool = True, + use_dynamic_shifting: bool = False, + time_shift_type: Literal["exponential", "linear"] = "linear", + ): + self.timestep_ratios = {} # The timestep ratio for each stage + self.timesteps_per_stage = {} # The detailed timesteps per stage (fix max and min per stage) + self.sigmas_per_stage = {} # always uniform [1000, 0] + self.start_sigmas = {} # for start point / upsample renoise + self.end_sigmas = {} # for end point + self.ori_start_sigmas = {} + + # self.init_sigmas() + self.init_sigmas_for_each_stage() + self.sigma_min = self.sigmas[-1].item() + self.sigma_max = self.sigmas[0].item() + self.gamma = gamma + + self.last_sample = None + self._step_index = None + self._begin_index = None + + def init_sigmas(self): + """ + initialize the global timesteps and sigmas + """ + num_train_timesteps = self.config.num_train_timesteps + shift = self.config.shift + + alphas = np.linspace(1, 1 / num_train_timesteps, num_train_timesteps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(shift * sigmas / (1 + (shift - 1) * sigmas))[:-1].copy() + sigmas = torch.from_numpy(sigmas) + timesteps = (sigmas * num_train_timesteps).clone() + + self._step_index = None + self._begin_index = None + self.timesteps = timesteps + self.sigmas = sigmas.to("cpu") # to avoid too much CPU/GPU communication + + def init_sigmas_for_each_stage(self): + """ + Init the timesteps for each stage + """ + self.init_sigmas() + + stage_distance = [] + stages = self.config.stages + training_steps = self.config.num_train_timesteps + stage_range = self.config.stage_range + + # Init the start and end point of each stage + for i_s in range(stages): + # To decide the start and ends point + start_indice = int(stage_range[i_s] * training_steps) + start_indice = max(start_indice, 0) + end_indice = int(stage_range[i_s + 1] * training_steps) + end_indice = min(end_indice, training_steps) + start_sigma = self.sigmas[start_indice].item() + end_sigma = self.sigmas[end_indice].item() if end_indice < training_steps else 0.0 + self.ori_start_sigmas[i_s] = start_sigma + + if i_s != 0: + ori_sigma = 1 - start_sigma + gamma = self.config.gamma + corrected_sigma = (1 / (math.sqrt(1 + (1 / gamma)) * (1 - ori_sigma) + ori_sigma)) * ori_sigma + # corrected_sigma = 1 / (2 - ori_sigma) * ori_sigma + start_sigma = 1 - corrected_sigma + + stage_distance.append(start_sigma - end_sigma) + self.start_sigmas[i_s] = start_sigma + self.end_sigmas[i_s] = end_sigma + + # Determine the ratio of each stage according to flow length + tot_distance = sum(stage_distance) + for i_s in range(stages): + if i_s == 0: + start_ratio = 0.0 + else: + start_ratio = sum(stage_distance[:i_s]) / tot_distance + if i_s == stages - 1: + end_ratio = 0.9999999999999999 + else: + end_ratio = sum(stage_distance[: i_s + 1]) / tot_distance + + self.timestep_ratios[i_s] = (start_ratio, end_ratio) + + # Determine the timesteps and sigmas for each stage + for i_s in range(stages): + timestep_ratio = self.timestep_ratios[i_s] + # timestep_max = self.timesteps[int(timestep_ratio[0] * training_steps)] + timestep_max = min(self.timesteps[int(timestep_ratio[0] * training_steps)], 999) + timestep_min = self.timesteps[min(int(timestep_ratio[1] * training_steps), training_steps - 1)] + timesteps = np.linspace(timestep_max, timestep_min, training_steps + 1) + self.timesteps_per_stage[i_s] = ( + timesteps[:-1] if isinstance(timesteps, torch.Tensor) else torch.from_numpy(timesteps[:-1]) + ) + stage_sigmas = np.linspace(0.999, 0, training_steps + 1) + self.sigmas_per_stage[i_s] = torch.from_numpy(stage_sigmas[:-1]) + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def _sigma_to_t(self, sigma): + return sigma * self.config.num_train_timesteps + + def set_timesteps( + self, + num_inference_steps: int, + stage_index: int | None = None, + device: str | torch.device = None, + sigmas: bool | None = None, + mu: bool | None = None, + is_amplify_first_chunk: bool = False, + ): + """ + Setting the timesteps and sigmas for each stage + """ + if is_amplify_first_chunk: + num_inference_steps = num_inference_steps * 2 + 1 + else: + num_inference_steps = num_inference_steps + 1 + + self.num_inference_steps = num_inference_steps + self.init_sigmas() + + if self.config.stages == 1: + if sigmas is None: + sigmas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)[:-1].astype( + np.float32 + ) + if self.config.shift != 1.0: + assert not self.config.use_dynamic_shifting + sigmas = self.time_shift(self.config.shift, 1.0, sigmas) + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = torch.from_numpy(sigmas) + else: + stage_timesteps = self.timesteps_per_stage[stage_index] + timesteps = np.linspace( + stage_timesteps[0].item(), + stage_timesteps[-1].item(), + num_inference_steps, + ) + + stage_sigmas = self.sigmas_per_stage[stage_index] + ratios = np.linspace(stage_sigmas[0].item(), stage_sigmas[-1].item(), num_inference_steps) + sigmas = torch.from_numpy(ratios) + + self.timesteps = torch.from_numpy(timesteps).to(device=device) + self.sigmas = torch.cat([sigmas, torch.zeros(1)]).to(device=device) + + self._step_index = None + self.reset_scheduler_history() + + self.timesteps = self.timesteps[:-1] + self.sigmas = torch.cat([self.sigmas[:-2], self.sigmas[-1:]]) + + if self.config.use_dynamic_shifting: + assert self.config.shift == 1.0 + self.sigmas = self.time_shift(mu, 1.0, self.sigmas) + if self.config.stages == 1: + self.timesteps = self.sigmas[:-1] * self.config.num_train_timesteps + else: + self.timesteps = self.timesteps_per_stage[stage_index].min() + self.sigmas[:-1] * ( + self.timesteps_per_stage[stage_index].max() - self.timesteps_per_stage[stage_index].min() + ) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler.time_shift + def time_shift(self, mu: float, sigma: float, t: torch.Tensor): + """ + Apply time shifting to the sigmas. + + Args: + mu (`float`): + The mu parameter for the time shift. + sigma (`float`): + The sigma parameter for the time shift. + t (`torch.Tensor`): + The input timesteps. + + Returns: + `torch.Tensor`: + The time-shifted timesteps. + """ + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler._time_shift_exponential + def _time_shift_exponential(self, mu, sigma, t): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler._time_shift_linear + def _time_shift_linear(self, mu, sigma, t): + return mu / (mu + (1 / t - 1) ** sigma) + + # ---------------------------------- For DMD ---------------------------------- + def add_noise(self, original_samples, noise, timestep, sigmas, timesteps): + sigmas = sigmas.to(noise.device) + timesteps = timesteps.to(noise.device) + timestep_id = torch.argmin((timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1) + sigma = sigmas[timestep_id].reshape(-1, 1, 1, 1, 1) + sample = (1 - sigma) * original_samples + sigma * noise + return sample.type_as(noise) + + def convert_flow_pred_to_x0(self, flow_pred, xt, timestep, sigmas, timesteps): + # use higher precision for calculations + original_dtype = flow_pred.dtype + device = flow_pred.device + flow_pred, xt, sigmas, timesteps = (x.double().to(device) for x in (flow_pred, xt, sigmas, timesteps)) + + timestep_id = torch.argmin((timesteps.unsqueeze(0) - timestep.unsqueeze(1)).abs(), dim=1) + sigma_t = sigmas[timestep_id].reshape(-1, 1, 1, 1, 1) + x0_pred = xt - sigma_t * flow_pred + return x0_pred.to(original_dtype) + + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor = None, + sample: torch.FloatTensor = None, + generator: torch.Generator | None = None, + return_dict: bool = True, + cur_sampling_step: int = 0, + dmd_noisy_tensor: torch.FloatTensor | None = None, + dmd_sigmas: torch.FloatTensor | None = None, + dmd_timesteps: torch.FloatTensor | None = None, + all_timesteps: torch.FloatTensor | None = None, + ) -> HeliosDMDSchedulerOutput | tuple: + pred_image_or_video = self.convert_flow_pred_to_x0( + flow_pred=model_output, + xt=sample, + timestep=torch.full((model_output.shape[0],), timestep, dtype=torch.long, device=model_output.device), + sigmas=dmd_sigmas, + timesteps=dmd_timesteps, + ) + if cur_sampling_step < len(all_timesteps) - 1: + prev_sample = self.add_noise( + pred_image_or_video, + dmd_noisy_tensor, + torch.full( + (model_output.shape[0],), + all_timesteps[cur_sampling_step + 1], + dtype=torch.long, + device=model_output.device, + ), + sigmas=dmd_sigmas, + timesteps=dmd_timesteps, + ) + else: + prev_sample = pred_image_or_video + + if not return_dict: + return (prev_sample,) + + return HeliosDMDSchedulerOutput(prev_sample=prev_sample) + + def reset_scheduler_history(self): + self._step_index = None + self._begin_index = None + + def __len__(self): + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_heun_discrete.py b/diffusers/schedulers/scheduling_heun_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..df2593493aeb33868910ae8d7c28b2506cd08cab --- /dev/null +++ b/diffusers/schedulers/scheduling_heun_discrete.py @@ -0,0 +1,711 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->HeunDiscrete +class HeunDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + Scheduler with Heun steps for discrete beta schedules. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, `"squaredcos_cap_v2"`, or `"exp"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, `squaredcos_cap_v2`, or `exp`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + prediction_type (`"epsilon"`, `"sample"`, or `"v_prediction"`, defaults to `"epsilon"`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to 1.0): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + prediction_type: str = "epsilon", + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + clip_sample: bool = False, + clip_sample_range: float = 1.0, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + ) -> None: + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="cosine") + elif beta_schedule == "exp": + self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="exp") + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + self.use_karras_sigmas = use_karras_sigmas + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + @property + def init_noise_sigma(self): + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: float | torch.Tensor, + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device = None, + num_train_timesteps: int | None = None, + timesteps: list[int] | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*, defaults to `None`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str`, `torch.device`, *optional*, defaults to `None`): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + num_train_timesteps (`int`, *optional*, defaults to `None`): + The number of diffusion steps used when training the model. If `None`, the default + `num_train_timesteps` attribute is used. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, timesteps will be + generated based on the `timestep_spacing` attribute. If `timesteps` is passed, `num_inference_steps` + must be `None`, and `timestep_spacing` attribute will be ignored. + """ + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.") + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + if timesteps is not None and self.config.use_karras_sigmas: + raise ValueError("Cannot use `timesteps` with `config.use_karras_sigmas = True`") + if timesteps is not None and self.config.use_exponential_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_exponential_sigmas = True`.") + if timesteps is not None and self.config.use_beta_sigmas: + raise ValueError("Cannot set `timesteps` with `config.use_beta_sigmas = True`.") + + num_inference_steps = num_inference_steps or len(timesteps) + self.num_inference_steps = num_inference_steps + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + if timesteps is not None: + timesteps = np.array(timesteps, dtype=np.float32) + else: + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]]) + + timesteps = torch.from_numpy(timesteps) + timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)]) + + self.timesteps = timesteps.to(device=device, dtype=torch.float32) + + # empty dt and derivative + self.prev_derivative = None + self.dt = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + @property + def state_in_first_order(self): + return self.dt is None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor | np.ndarray, + timestep: float | torch.Tensor, + sample: torch.Tensor | np.ndarray, + return_dict: bool = True, + ) -> HeunDiscreteSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_heun_discrete.HeunDiscreteSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_heun_discrete.HeunDiscreteSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_heun_discrete.HeunDiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order / Heun's method + sigma = self.sigmas[self.step_index - 1] + sigma_next = self.sigmas[self.step_index] + + # currently only gamma=0 is supported. This usually works best anyways. + # We can support gamma in the future but then need to scale the timestep before + # passing it to the model which requires a change in API + gamma = 0 + sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma_hat if self.state_in_first_order else sigma_next + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma_hat if self.state_in_first_order else sigma_next + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if self.config.clip_sample: + pred_original_sample = pred_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + if self.state_in_first_order: + # 2. Convert to an ODE derivative for 1st order + derivative = (sample - pred_original_sample) / sigma_hat + # 3. delta timestep + dt = sigma_next - sigma_hat + + # store for 2nd order step + self.prev_derivative = derivative + self.dt = dt + self.sample = sample + else: + # 2. 2nd order / Heun's method + derivative = (sample - pred_original_sample) / sigma_next + derivative = (self.prev_derivative + derivative) / 2 + + # 3. take prev timestep & sample + dt = self.dt + sample = self.sample + + # free dt and derivative + # Note, this puts the scheduler in "first order mode" + self.prev_derivative = None + self.dt = None + self.sample = None + + prev_sample = sample + derivative * dt + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return HeunDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ipndm.py b/diffusers/schedulers/scheduling_ipndm.py new file mode 100644 index 0000000000000000000000000000000000000000..f2e744e0f2afd29b49b6d06c961a9cb69b9cb74c --- /dev/null +++ b/diffusers/schedulers/scheduling_ipndm.py @@ -0,0 +1,279 @@ +# Copyright 2025 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +class IPNDMScheduler(SchedulerMixin, ConfigMixin): + """ + A fourth-order Improved Pseudo Linear Multistep scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + trained_betas (`np.ndarray` or `List[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + trained_betas: np.ndarray | list[float] | None = None, + ): + # set `betas`, `alphas`, `timesteps` + self.set_timesteps(num_train_timesteps) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://huggingface.co/papers/2202.09778 + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.ets = [] + self._step_index = None + self._begin_index = None + + @property + def step_index(self) -> int | None: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + + Returns: + `int` or `None`: + The index counter for current timestep. + """ + return self._step_index + + @property + def begin_index(self) -> int | None: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The index for the first timestep. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device | None = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + steps = torch.linspace(1, 0, num_inference_steps + 1)[:-1] + steps = torch.cat([steps, torch.tensor([0.0])]) + + if self.config.trained_betas is not None: + self.betas = torch.tensor(self.config.trained_betas, dtype=torch.float32) + else: + self.betas = torch.sin(steps * math.pi / 2) ** 2 + + self.alphas = (1.0 - self.betas**2) ** 0.5 + + timesteps = (torch.atan2(self.betas, self.alphas) / math.pi * 2)[:-1] + self.timesteps = timesteps.to(device) + + self.ets = [] + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + if self.step_index is None: + self._init_step_index(timestep) + + timestep_index = self.step_index + prev_timestep_index = self.step_index + 1 + + ets = sample * self.betas[timestep_index] + model_output * self.alphas[timestep_index] + self.ets.append(ets) + + if len(self.ets) == 1: + ets = self.ets[-1] + elif len(self.ets) == 2: + ets = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + ets = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + ets = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep_index, prev_timestep_index, ets) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample( + self, + sample: torch.Tensor, + timestep_index: int, + prev_timestep_index: int, + ets: torch.Tensor, + ) -> torch.Tensor: + """ + Predicts the previous sample based on the current sample, timestep indices, and running model outputs. + + Args: + sample (`torch.Tensor`): + The current sample. + timestep_index (`int`): + Index of the current timestep in the schedule. + prev_timestep_index (`int`): + Index of the previous timestep in the schedule. + ets (`torch.Tensor`): + The running sequence of model outputs. + + Returns: + `torch.Tensor`: + The predicted previous sample. + """ + alpha = self.alphas[timestep_index] + sigma = self.betas[timestep_index] + + next_alpha = self.alphas[prev_timestep_index] + next_sigma = self.betas[prev_timestep_index] + + pred = (sample - sigma * ets) / max(alpha, 1e-8) + prev_sample = next_alpha * pred + ets * next_sigma + + return prev_sample + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py b/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..0a7bffb01a869bafd548aee9e273097adfd18064 --- /dev/null +++ b/diffusers/schedulers/scheduling_k_dpm_2_ancestral_discrete.py @@ -0,0 +1,750 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->KDPM2AncestralDiscrete +class KDPM2AncestralDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + KDPM2DiscreteScheduler with ancestral sampling is inspired by the DPMSolver2 and Algorithm 2 from the [Elucidating + the Design Space of Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) paper. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.00085): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.012): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray` or `list[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + use_karras_sigmas (`bool`, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + prediction_type (`str`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ) -> None: + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> torch.Tensor: + """ + The standard deviation of the initial noise distribution. + + Returns: + `torch.Tensor`: + The standard deviation of the initial noise distribution. + """ + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self) -> int | None: + """ + The index counter for the current timestep. It increases by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if it has not been initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int | None: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index, or `None` if it has not been set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: float | torch.Tensor, + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + else: + sigma = self.sigmas_interpol[self.step_index - 1] + + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device | None = None, + num_train_timesteps: int | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + num_train_timesteps (`int`, *optional*): + The number of diffusion steps used to train the model. If `None`, `self.config.num_train_timesteps` is + used. + """ + self.num_inference_steps = num_inference_steps + + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + self.log_sigmas = torch.from_numpy(log_sigmas).to(device) + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + + # compute up and down sigmas + sigmas_next = sigmas.roll(-1) + sigmas_next[-1] = 0.0 + sigmas_up = (sigmas_next**2 * (sigmas**2 - sigmas_next**2) / sigmas**2) ** 0.5 + sigmas_down = (sigmas_next**2 - sigmas_up**2) ** 0.5 + sigmas_down[-1] = 0.0 + + # compute interpolated sigmas + sigmas_interpol = sigmas.log().lerp(sigmas_down.log(), 0.5).exp() + sigmas_interpol[-2:] = 0.0 + + # set sigmas + self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]]) + self.sigmas_interpol = torch.cat( + [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]] + ) + self.sigmas_up = torch.cat([sigmas_up[:1], sigmas_up[1:].repeat_interleave(2), sigmas_up[-1:]]) + self.sigmas_down = torch.cat([sigmas_down[:1], sigmas_down[1:].repeat_interleave(2), sigmas_down[-1:]]) + + if str(device).startswith("mps"): + timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32) + else: + timesteps = torch.from_numpy(timesteps).to(device) + + sigmas_interpol = sigmas_interpol.cpu() + log_sigmas = self.log_sigmas.cpu() + timesteps_interpol = np.array( + [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol] + ) + + timesteps_interpol = torch.from_numpy(timesteps_interpol).to(device, dtype=timesteps.dtype) + interleaved_timesteps = torch.stack((timesteps_interpol[:-2, None], timesteps[1:, None]), dim=-1).flatten() + + self.timesteps = torch.cat([timesteps[:1], interleaved_timesteps]) + + self.sample = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + @property + def state_in_first_order(self) -> bool: + """ + Whether the scheduler is currently performing a first-order update. + + Returns: + `bool`: + `True` if the scheduler is performing a first-order update, otherwise `False`. + """ + return self.sample is None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor | np.ndarray, + timestep: float | torch.Tensor, + sample: torch.Tensor | np.ndarray, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> KDPM2AncestralDiscreteSchedulerOutput | tuple[torch.Tensor, torch.Tensor]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor` or `np.ndarray`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor` or `np.ndarray`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a + [`~schedulers.scheduling_k_dpm_2_ancestral_discrete.KDPM2AncestralDiscreteSchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_k_dpm_2_ancestral_discrete.KDPM2AncestralDiscreteSchedulerOutput`] or + `tuple[torch.Tensor, torch.Tensor]`: + If `return_dict` is `True`, + [`~schedulers.scheduling_k_dpm_2_ancestral_discrete.KDPM2AncestralDiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_interpol = self.sigmas_interpol[self.step_index] + sigma_up = self.sigmas_up[self.step_index] + sigma_down = self.sigmas_down[self.step_index - 1] + else: + # 2nd order / KPDM2's method + sigma = self.sigmas[self.step_index - 1] + sigma_interpol = self.sigmas_interpol[self.step_index - 1] + sigma_up = self.sigmas_up[self.step_index - 1] + sigma_down = self.sigmas_down[self.step_index - 1] + + # currently only gamma=0 is supported. This usually works best anyways. + # We can support gamma in the future but then need to scale the timestep before + # passing it to the model which requires a change in API + gamma = 0 + sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if self.state_in_first_order: + # 2. Convert to an ODE derivative for 1st order + derivative = (sample - pred_original_sample) / sigma_hat + # 3. delta timestep + dt = sigma_interpol - sigma_hat + + # store for 2nd order step + self.sample = sample + self.dt = dt + prev_sample = sample + derivative * dt + else: + # DPM-Solver-2 + # 2. Convert to an ODE derivative for 2nd order + derivative = (sample - pred_original_sample) / sigma_interpol + # 3. delta timestep + dt = sigma_down - sigma_hat + + sample = self.sample + self.sample = None + + prev_sample = sample + derivative * dt + noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator + ) + prev_sample = prev_sample + noise * sigma_up + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return KDPM2AncestralDiscreteSchedulerOutput( + prev_sample=prev_sample, pred_original_sample=pred_original_sample + ) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + """ + Return the number of training timesteps configured for the scheduler. + + Returns: + `int`: + The number of training timesteps. + """ + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_k_dpm_2_discrete.py b/diffusers/schedulers/scheduling_k_dpm_2_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..f836df424ad0b858f09f1093ad4c92bb7ce1d0ec --- /dev/null +++ b/diffusers/schedulers/scheduling_k_dpm_2_discrete.py @@ -0,0 +1,715 @@ +# Copyright 2025 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->KDPM2Discrete +class KDPM2DiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class KDPM2DiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + KDPM2DiscreteScheduler is inspired by the DPMSolver2 and Algorithm 2 from the [Elucidating the Design Space of + Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) paper. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.00085`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.012`): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear` or `scaled_linear`. + trained_betas (`np.ndarray` or `list[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + use_karras_sigmas (`bool`, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + prediction_type (`str`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 2 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, # sensible defaults + beta_end: float = 0.012, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ) -> None: + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # set all values + self.set_timesteps(num_train_timesteps, None, num_train_timesteps) + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> torch.Tensor: + """ + The standard deviation of the initial noise distribution. + + Returns: + `torch.Tensor`: + The standard deviation of the initial noise distribution. + """ + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self) -> int | None: + """ + The index counter for the current timestep. It increases by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if it has not been initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int | None: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index, or `None` if it has not been set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input( + self, + sample: torch.Tensor, + timestep: float | torch.Tensor, + ) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + else: + sigma = self.sigmas_interpol[self.step_index] + + sample = sample / ((sigma**2 + 1) ** 0.5) + return sample + + def set_timesteps( + self, + num_inference_steps: int, + device: str | torch.device | None = None, + num_train_timesteps: int | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + num_train_timesteps (`int`, *optional*): + The number of diffusion steps used to train the model. If `None`, `self.config.num_train_timesteps` is + used. + """ + self.num_inference_steps = num_inference_steps + + num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + self.log_sigmas = torch.from_numpy(log_sigmas).to(device=device) + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + sigmas = torch.from_numpy(sigmas).to(device=device) + + # interpolate sigmas + sigmas_interpol = sigmas.log().lerp(sigmas.roll(1).log(), 0.5).exp() + + self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]]) + self.sigmas_interpol = torch.cat( + [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]] + ) + + timesteps = torch.from_numpy(timesteps).to(device) + + # interpolate timesteps + sigmas_interpol = sigmas_interpol.cpu() + log_sigmas = self.log_sigmas.cpu() + timesteps_interpol = np.array( + [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol] + ) + timesteps_interpol = torch.from_numpy(timesteps_interpol).to(device, dtype=timesteps.dtype) + interleaved_timesteps = torch.stack((timesteps_interpol[1:-1, None], timesteps[1:, None]), dim=-1).flatten() + + self.timesteps = torch.cat([timesteps[:1], interleaved_timesteps]) + + self.sample = None + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def state_in_first_order(self) -> bool: + """ + Whether the scheduler is currently performing a first-order update. + + Returns: + `bool`: + `True` if the scheduler is performing a first-order update, otherwise `False`. + """ + return self.sample is None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def step( + self, + model_output: torch.Tensor | np.ndarray, + timestep: float | torch.Tensor, + sample: torch.Tensor | np.ndarray, + return_dict: bool = True, + ) -> KDPM2DiscreteSchedulerOutput | tuple[torch.Tensor, torch.Tensor]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor` or `np.ndarray`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor` or `np.ndarray`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_k_dpm_2_discrete.KDPM2DiscreteSchedulerOutput`] or + tuple. + + Returns: + [`~schedulers.scheduling_k_dpm_2_discrete.KDPM2DiscreteSchedulerOutput`] or `tuple[torch.Tensor, + torch.Tensor]`: + If return_dict is `True`, [`~schedulers.scheduling_k_dpm_2_discrete.KDPM2DiscreteSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if self.step_index is None: + self._init_step_index(timestep) + + if self.state_in_first_order: + sigma = self.sigmas[self.step_index] + sigma_interpol = self.sigmas_interpol[self.step_index + 1] + sigma_next = self.sigmas[self.step_index + 1] + else: + # 2nd order / KDPM2's method + sigma = self.sigmas[self.step_index - 1] + sigma_interpol = self.sigmas_interpol[self.step_index] + sigma_next = self.sigmas[self.step_index] + + # currently only gamma=0 is supported. This usually works best anyways. + # We can support gamma in the future but then need to scale the timestep before + # passing it to the model which requires a change in API + gamma = 0 + sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = sample - sigma_input * model_output + elif self.config.prediction_type == "v_prediction": + sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol + pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + ( + sample / (sigma_input**2 + 1) + ) + elif self.config.prediction_type == "sample": + raise NotImplementedError("prediction_type not implemented yet: sample") + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + if self.state_in_first_order: + # 2. Convert to an ODE derivative for 1st order + derivative = (sample - pred_original_sample) / sigma_hat + # 3. delta timestep + dt = sigma_interpol - sigma_hat + + # store for 2nd order step + self.sample = sample + else: + # DPM-Solver-2 + # 2. Convert to an ODE derivative for 2nd order + derivative = (sample - pred_original_sample) / sigma_interpol + + # 3. delta timestep + dt = sigma_next - sigma_hat + + sample = self.sample + self.sample = None + + # upon completion increase step index by one + self._step_index += 1 + + prev_sample = sample + derivative * dt + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return KDPM2DiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_lcm.py b/diffusers/schedulers/scheduling_lcm.py new file mode 100644 index 0000000000000000000000000000000000000000..9c47fefac935161e372a1d8a42fc28c070353e8d --- /dev/null +++ b/diffusers/schedulers/scheduling_lcm.py @@ -0,0 +1,775 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class LCMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + denoised (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images, *optional*): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. `denoised` can + be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + denoised: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class LCMScheduler(SchedulerMixin, ConfigMixin): + """ + `LCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config + attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be + accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving + functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.00085`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.012`): + The final `beta` value. + beta_schedule (`str`, defaults to `"scaled_linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray` or `list[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + original_inference_steps (`int`, defaults to `50`): + The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we + will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule. + clip_sample (`bool`, defaults to `False`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to `1.0`): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to `0.995`): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to `1.0`): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + timestep_scaling (`float`, defaults to `10.0`): + The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions + `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation + error at the default of `10.0` is already pretty small). + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.012, + beta_schedule: str = "scaled_linear", + trained_betas: np.ndarray | list[float] | None = None, + original_inference_steps: int = 50, + clip_sample: bool = False, + clip_sample_range: float = 1.0, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + timestep_scaling: float = 10.0, + rescale_betas_zero_snr: bool = False, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + self.custom_timesteps = False + + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def step_index(self) -> int | None: + """ + The index counter for current timestep. It will increase by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int | None: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index for the scheduler, or `None` if not set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + original_inference_steps: int | None = None, + timesteps: list[int] | None = None, + strength: float = 1.0, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + original_inference_steps (`int`, *optional*): + The original number of inference steps, which will be used to generate a linearly-spaced timestep + schedule (which is different from the standard `diffusers` implementation). We will then take + `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as + our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep + schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`. + strength (`float`, defaults to `1.0`): + Strength factor used to generate a partial timestep schedule (e.g. for image-to-image). A value of + `1.0` uses the full schedule. + + Returns: + `None` + """ + # 0. Check inputs + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.") + + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + # 1. Calculate the LCM original training/distillation timestep schedule. + original_steps = ( + original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps + ) + + if original_steps > self.config.num_train_timesteps: + raise ValueError( + f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + # LCM Timesteps Setting + # The skipping step parameter k from the paper. + k = self.config.num_train_timesteps // original_steps + # LCM Training/Distillation Steps Schedule + # Currently, only a linearly-spaced schedule is supported (same as in the LCM distillation scripts). + lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1 + + # 2. Calculate the LCM inference timestep schedule. + if timesteps is not None: + # 2.1 Handle custom timestep schedules. + train_timesteps = set(lcm_origin_timesteps) + non_train_timesteps = [] + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[i] not in train_timesteps: + non_train_timesteps.append(timesteps[i]) + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1 + if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1: + logger.warning( + f"The first timestep on the custom timestep schedule is {timesteps[0]}, not" + f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get" + f" unexpected results when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule + if non_train_timesteps: + logger.warning( + f"The custom timestep schedule contains the following timesteps which are not on the original" + f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results" + f" when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule is longer than original_steps + if len(timesteps) > original_steps: + logger.warning( + f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the" + f" the length of the timestep schedule used for training: {original_steps}. You may get some" + f" unexpected results when using this timestep schedule." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.num_inference_steps = len(timesteps) + self.custom_timesteps = True + + # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps) + init_timestep = min(int(self.num_inference_steps * strength), self.num_inference_steps) + t_start = max(self.num_inference_steps - init_timestep, 0) + timesteps = timesteps[t_start * self.order :] + # TODO: also reset self.num_inference_steps? + else: + # 2.2 Create the "standard" LCM inference timestep schedule. + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + skipping_step = len(lcm_origin_timesteps) // num_inference_steps + + if skipping_step < 1: + raise ValueError( + f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}." + ) + + self.num_inference_steps = num_inference_steps + + if num_inference_steps > original_steps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:" + f" {original_steps} because the final timestep schedule will be a subset of the" + f" `original_inference_steps`-sized initial timestep schedule." + ) + + # LCM Inference Steps Schedule + lcm_origin_timesteps = lcm_origin_timesteps[::-1].copy() + # Select (approximately) evenly spaced indices from lcm_origin_timesteps. + inference_indices = np.linspace(0, len(lcm_origin_timesteps), num=num_inference_steps, endpoint=False) + inference_indices = np.floor(inference_indices).astype(np.int64) + timesteps = lcm_origin_timesteps[inference_indices] + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long) + + self._step_index = None + self._begin_index = None + + def get_scalings_for_boundary_condition_discrete(self, timestep: int) -> tuple[float, float]: + """ + Computes the boundary condition scalings (`c_skip` and `c_out`) for the given discrete timestep, as used in the + Latent Consistency Model. + + Args: + timestep (`int`): + The discrete timestep for which to compute the scalings. + + Returns: + `tuple[float, float]`: + A tuple containing `c_skip` (scaling for the input sample) and `c_out` (scaling for the predicted + denoised sample). + """ + self.sigma_data = 0.5 # Default: 0.5 + scaled_timestep = timestep * self.config.timestep_scaling + + c_skip = self.sigma_data**2 / (scaled_timestep**2 + self.sigma_data**2) + c_out = scaled_timestep / (scaled_timestep**2 + self.sigma_data**2) ** 0.5 + return c_skip, c_out + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> LCMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible sampling. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # 1. get previous step value + prev_step_index = self.step_index + 1 + if prev_step_index < len(self.timesteps): + prev_timestep = self.timesteps[prev_step_index] + else: + prev_timestep = timestep + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + # 3. Get scalings for boundary conditions + c_skip, c_out = self.get_scalings_for_boundary_condition_discrete(timestep) + + # 4. Compute the predicted original sample x_0 based on the model parameterization + if self.config.prediction_type == "epsilon": # noise-prediction + predicted_original_sample = (sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt() + elif self.config.prediction_type == "sample": # x-prediction + predicted_original_sample = model_output + elif self.config.prediction_type == "v_prediction": # v-prediction + predicted_original_sample = alpha_prod_t.sqrt() * sample - beta_prod_t.sqrt() * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for `LCMScheduler`." + ) + + # 5. Clip or threshold "predicted x_0" + if self.config.thresholding: + predicted_original_sample = self._threshold_sample(predicted_original_sample) + elif self.config.clip_sample: + predicted_original_sample = predicted_original_sample.clamp( + -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 6. Denoise model output using boundary conditions + denoised = c_out * predicted_original_sample + c_skip * sample + + # 7. Sample and inject noise z ~ N(0, I) for MultiStep Inference + # Noise is not used on the final timestep of the timestep schedule. + # This also means that noise is not used for one-step sampling. + if self.step_index != self.num_inference_steps - 1: + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=denoised.dtype + ) + prev_sample = alpha_prod_t_prev.sqrt() * denoised + beta_prod_t_prev.sqrt() * noise + else: + prev_sample = denoised + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample, denoised) + + return LCMSchedulerOutput(prev_sample=prev_sample, denoised=denoised) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + """ + Returns the number of train timesteps. + + Returns: + `int`: + The number of diffusion steps used to train the model. + """ + return self.config.num_train_timesteps + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep + def previous_timestep(self, timestep): + """ + Compute the previous timestep in the diffusion chain. + + Args: + timestep (`int`): + The current timestep. + + Returns: + `int` or `torch.Tensor`: + The previous timestep. + """ + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/diffusers/schedulers/scheduling_lms_discrete.py b/diffusers/schedulers/scheduling_lms_discrete.py new file mode 100644 index 0000000000000000000000000000000000000000..e917adda9516e840f4997f03afdef47c37082fbf --- /dev/null +++ b/diffusers/schedulers/scheduling_lms_discrete.py @@ -0,0 +1,675 @@ +# Copyright 2025 Katherine Crowson and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import math +import warnings +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin + + +if is_scipy_available(): + import scipy.stats + from scipy import integrate + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->LMSDiscrete +class LMSDiscreteSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class LMSDiscreteScheduler(SchedulerMixin, ConfigMixin): + """ + A linear multistep scheduler for discrete beta schedules. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.0001`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.02`): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + prediction_type (`"epsilon"`, `"sample"`, or `"v_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + prediction_type: Literal["epsilon", "sample", "v_prediction"] = "epsilon", + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + ): + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32) + self.sigmas = torch.from_numpy(sigmas) + + # setable values + self.num_inference_steps = None + self.use_karras_sigmas = use_karras_sigmas + self.set_timesteps(num_train_timesteps, None) + self.derivatives = [] + self.is_scale_input_called = False + + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def init_noise_sigma(self) -> float | torch.Tensor: + """ + The standard deviation of the initial noise distribution. + + Returns: + `float` or `torch.Tensor`: + The standard deviation of the initial noise distribution, computed based on the maximum sigma value and + the timestep spacing configuration. + """ + # standard deviation of the initial noise distribution + if self.config.timestep_spacing in ["linspace", "trailing"]: + return self.sigmas.max() + + return (self.sigmas.max() ** 2 + 1) ** 0.5 + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index for the scheduler, or `None` if not set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: float | torch.Tensor) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`float` or `torch.Tensor`): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + sample = sample / ((sigma**2 + 1) ** 0.5) + self.is_scale_input_called = True + return sample + + def get_lms_coefficient(self, order: int, t: int, current_order: int) -> float: + """ + Compute the linear multistep coefficient. + + Args: + order (`int`): + The order of the linear multistep method. + t (`int`): + The current timestep index. + current_order (`int`): + The current order for which to compute the coefficient. + + Returns: + `float`: + The computed linear multistep coefficient. + """ + + def lms_derivative(tau): + prod = 1.0 + for k in range(order): + if current_order == k: + continue + prod *= (tau - self.sigmas[t - k]) / (self.sigmas[t - current_order] - self.sigmas[t - k]) + return prod + + integrated_coeff = integrate.quad(lms_derivative, self.sigmas[t], self.sigmas[t + 1], epsrel=1e-4)[0] + + return integrated_coeff + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[ + ::-1 + ].copy() + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + + if self.config.use_karras_sigmas: + sigmas = self._convert_to_karras(in_sigmas=sigmas) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_exponential_sigmas: + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + elif self.config.use_beta_sigmas: + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + + sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas).to(device=device) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.float32) + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + self.derivatives = [] + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + def _convert_to_karras(self, in_sigmas: torch.Tensor) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + sigma_min: float = in_sigmas[-1].item() + sigma_max: float = in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, self.num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def step( + self, + model_output: torch.Tensor, + timestep: float | torch.Tensor, + sample: torch.Tensor, + order: int = 4, + return_dict: bool = True, + ) -> LMSDiscreteSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`float` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + order (`int`, defaults to 4): + The order of the linear multistep method. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if not self.is_scale_input_called: + warnings.warn( + "The `scale_model_input` function should be called before `step` to ensure correct denoising. " + "See `StableDiffusionPipeline` for a usage example." + ) + + if self.step_index is None: + self._init_step_index(timestep) + + sigma = self.sigmas[self.step_index] + + # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise + if self.config.prediction_type == "epsilon": + pred_original_sample = sample - sigma * model_output + elif self.config.prediction_type == "v_prediction": + # * c_out + input * c_skip + pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1)) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`" + ) + + # 2. Convert to an ODE derivative + derivative = (sample - pred_original_sample) / sigma + self.derivatives.append(derivative) + if len(self.derivatives) > order: + self.derivatives.pop(0) + + # 3. Compute linear multistep coefficients + order = min(self.step_index + 1, order) + lms_coeffs = [self.get_lms_coefficient(order, self.step_index, curr_order) for curr_order in range(order)] + + # 4. Compute previous sample based on the derivatives path + prev_sample = sample + sum( + coeff * derivative for coeff, derivative in zip(lms_coeffs, reversed(self.derivatives)) + ) + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return ( + prev_sample, + pred_original_sample, + ) + + return LMSDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + noisy_samples = original_samples + noise * sigma + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_ltx_euler_ancestral_rf.py b/diffusers/schedulers/scheduling_ltx_euler_ancestral_rf.py new file mode 100644 index 0000000000000000000000000000000000000000..6b12c5ef96b2397c0d482f7c32e83ff9a3ce7787 --- /dev/null +++ b/diffusers/schedulers/scheduling_ltx_euler_ancestral_rf.py @@ -0,0 +1,429 @@ +# Copyright 2025 Lightricks and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +LTXEulerAncestralRFScheduler + +This scheduler implements a K-diffusion style Euler-Ancestral sampler specialized for flow / CONST parameterization, +closely mirroring ComfyUI's `sample_euler_ancestral_RF` implementation used for LTX-Video. + +Reference implementation (ComfyUI): + comfy.k_diffusion.sampling.sample_euler_ancestral_RF +""" + +from dataclasses import dataclass + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class LTXEulerAncestralRFSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor`): + Updated sample for the next step in the denoising process. + """ + + prev_sample: torch.FloatTensor + + +class LTXEulerAncestralRFScheduler(SchedulerMixin, ConfigMixin): + """ + Euler-Ancestral scheduler for LTX-Video (RF / CONST parametrization). + + This scheduler is intended for models where the network is trained with a CONST-like parameterization (as in LTXV / + FLUX). It approximates ComfyUI's `sample_euler_ancestral_RF` sampler and is useful when reproducing ComfyUI + workflows inside diffusers. + + The scheduler can either: + - reuse the [`FlowMatchEulerDiscreteScheduler`] sigma / timestep logic when only `num_inference_steps` is provided + (default diffusers-style usage), or + - follow an explicit ComfyUI-style sigma schedule when `sigmas` (or `timesteps`) are passed to [`set_timesteps`]. + + Args: + num_train_timesteps (`int`, defaults to 1000): + Included for config compatibility; not used to build the schedule. + eta (`float`, defaults to 1.0): + Stochasticity parameter. `eta=0.0` yields deterministic DDIM-like sampling; `eta=1.0` matches ComfyUI's + default RF behavior. + s_noise (`float`, defaults to 1.0): + Global scaling factor for the stochastic noise term. + """ + + # Allow config migration from the flow-match scheduler and back. + _compatibles = ["FlowMatchEulerDiscreteScheduler"] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + eta: float = 1.0, + s_noise: float = 1.0, + ) -> None: + # Note: num_train_timesteps is kept only for config compatibility. + self.num_inference_steps: int | None = None + self.sigmas: torch.Tensor | None = None + self.timesteps: torch.Tensor | None = None + self._step_index: int | None = None + self._begin_index: int | None = None + + @property + def step_index(self) -> int | None: + """ + The index counter for the current timestep. It increases by one after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self) -> int | None: + """ + The index for the first timestep. It can be set from a pipeline with `set_begin_index` to support + image-to-image like workflows that start denoising part-way through the schedule. + """ + return self._begin_index + + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Included for API compatibility; not strictly needed here but kept to allow pipelines that call + `set_begin_index`. + + Args: + begin_index (`int`, defaults to `0`): + The index of the first timestep in the denoising schedule. + """ + self._begin_index = begin_index + + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Map a (continuous) `timestep` value to an index into `self.timesteps`. + + This follows the convention used in other discrete schedulers: if the same timestep value appears multiple + times in the schedule (which can happen when starting in the middle of the schedule), the *second* occurrence + is used for the first `step` call so that no sigma is accidentally skipped. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search. If `None`, `self.timesteps` is used. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + if self.timesteps is None: + raise ValueError("Timesteps have not been set. Call `set_timesteps` first.") + schedule_timesteps = self.timesteps + + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(schedule_timesteps.device) + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + if len(indices) == 0: + raise ValueError( + "Passed `timestep` is not in `self.timesteps`. Make sure to use values from `scheduler.timesteps`." + ) + + return indices[pos].item() + + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the internal step index based on a given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep used to initialize the step index. + """ + if self.timesteps is None: + raise ValueError("Timesteps have not been set. Call `set_timesteps` first.") + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + sigmas: list[float] | torch.Tensor | None = None, + timesteps: list[float] | torch.Tensor | None = None, + mu: float | None = None, + **kwargs: object, + ) -> None: + """ + Set the sigma / timestep schedule for sampling. + + When `sigmas` or `timesteps` are provided explicitly, they are used as the RF sigma schedule (ComfyUI-style) + and are expected to include the terminal 0.0. When both are `None`, the scheduler reuses the + [`FlowMatchEulerDiscreteScheduler`] logic to generate sigmas from `num_inference_steps` and the stored config + (including any resolution-dependent shifting, Karras/beta schedules, etc.). + + Args: + num_inference_steps (`int`, *optional*): + Number of denoising steps. If provided together with explicit `sigmas`/`timesteps`, they are expected + to be consistent and are otherwise ignored with a warning. + device (`str` or `torch.device`, *optional*): + Device to move the internal tensors to. + sigmas (`list[float]` or `torch.Tensor`, *optional*): + Explicit sigma schedule, e.g. `[1.0, 0.99, ..., 0.0]`. + timesteps (`list[float]` or `torch.Tensor`, *optional*): + Optional alias for `sigmas`. If `sigmas` is None and `timesteps` is provided, timesteps are treated as + sigmas. + mu (`float`, *optional*): + Optional shift parameter used when delegating to [`FlowMatchEulerDiscreteScheduler.set_timesteps`] and + `config.use_dynamic_shifting` is `True`. + kwargs (`object`): + Additional keyword arguments accepted for API compatibility. + """ + # 1. Auto-generate schedule (FlowMatch-style) when no explicit sigmas/timesteps are given + if sigmas is None and timesteps is None: + if num_inference_steps is None: + raise ValueError( + "LTXEulerAncestralRFScheduler.set_timesteps requires either explicit `sigmas`/`timesteps` " + "or a `num_inference_steps` value." + ) + + # We reuse FlowMatchEulerDiscreteScheduler to construct a sigma schedule that is + # consistent with the original LTX training setup (including optional time shifting, + # Karras / exponential / beta schedules, etc.). + from .scheduling_flow_match_euler_discrete import FlowMatchEulerDiscreteScheduler + + base_scheduler = FlowMatchEulerDiscreteScheduler.from_config(self.config) + base_scheduler.set_timesteps( + num_inference_steps=num_inference_steps, + device=device, + sigmas=None, + mu=mu, + timesteps=None, + ) + + self.num_inference_steps = base_scheduler.num_inference_steps + # Keep sigmas / timesteps on the requested device so step() can operate on-device without + # extra transfers. + self.sigmas = base_scheduler.sigmas.to(device=device) + self.timesteps = base_scheduler.timesteps.to(device=device) + self._step_index = None + self._begin_index = None + return + + # 2. Explicit sigma schedule (ComfyUI-style path) + if sigmas is None: + # `timesteps` is treated as sigmas in RF / flow-matching setups. + sigmas = timesteps + + if isinstance(sigmas, list): + sigmas_tensor = torch.tensor(sigmas, dtype=torch.float32) + elif isinstance(sigmas, torch.Tensor): + sigmas_tensor = sigmas.to(dtype=torch.float32) + else: + raise TypeError(f"`sigmas` must be a list or torch.Tensor, got {type(sigmas)}.") + + if sigmas_tensor.ndim != 1: + raise ValueError(f"`sigmas` must be a 1D tensor, got shape {tuple(sigmas_tensor.shape)}.") + + if sigmas_tensor[-1].abs().item() > 1e-6: + logger.warning( + "The last sigma in the schedule is not zero (%.6f). " + "For best compatibility with ComfyUI's RF sampler, the terminal sigma " + "should be 0.0.", + sigmas_tensor[-1].item(), + ) + + # Move to device once, then derive timesteps. + if device is not None: + sigmas_tensor = sigmas_tensor.to(device) + + # Internal sigma schedule stays in [0, 1] (as provided). + self.sigmas = sigmas_tensor + # Timesteps are scaled to match the training setup of LTX (FlowMatch-style), + # where the network expects timesteps on [0, num_train_timesteps]. + # This keeps the transformer conditioning in the expected range while the RF + # scheduler still operates on the raw sigma values. + num_train = float(getattr(self.config, "num_train_timesteps", 1000)) + self.timesteps = sigmas_tensor * num_train + + if num_inference_steps is not None and num_inference_steps != len(sigmas) - 1: + logger.warning( + "Provided `num_inference_steps=%d` does not match `len(sigmas)-1=%d`. " + "Overriding `num_inference_steps` with `len(sigmas)-1`.", + num_inference_steps, + len(sigmas) - 1, + ) + + self.num_inference_steps = len(sigmas) - 1 + self._step_index = None + self._begin_index = None + + def _sigma_broadcast(self, sigma: torch.Tensor, sample: torch.Tensor) -> torch.Tensor: + """ + Helper to broadcast a scalar sigma to the shape of `sample`. + + Args: + sigma (`torch.Tensor`): + The scalar sigma tensor to broadcast. + sample (`torch.Tensor`): + The sample tensor whose number of dimensions determines the broadcast shape. + + Returns: + `torch.Tensor`: + The sigma tensor reshaped for broadcasting with `sample`. + """ + while sigma.ndim < sample.ndim: + sigma = sigma.view(*sigma.shape, 1) + return sigma + + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.Tensor, + sample: torch.FloatTensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> LTXEulerAncestralRFSchedulerOutput | tuple[torch.FloatTensor]: + """ + Perform a single Euler-Ancestral RF update step. + + Args: + model_output (`torch.FloatTensor`): + Raw model output at the current step. Interpreted under the CONST parametrization as `v_t`, with + denoised state reconstructed as `x0 = x_t - sigma_t * v_t`. + timestep (`float` or `torch.Tensor`): + The current sigma value (must match one entry in `self.timesteps`). + sample (`torch.FloatTensor`): + Current latent sample `x_t`. + generator (`torch.Generator`, *optional*): + Optional generator for reproducible noise. + return_dict (`bool`, defaults to `True`): + If `True`, return a `LTXEulerAncestralRFSchedulerOutput`; otherwise return a tuple where the first + element is the updated sample. + + Returns: + [`~schedulers.scheduling_ltx_euler_ancestral_rf.LTXEulerAncestralRFSchedulerOutput`] or `tuple`: + If `return_dict` is `True`, + [`~schedulers.scheduling_ltx_euler_ancestral_rf.LTXEulerAncestralRFSchedulerOutput`] is returned. + Otherwise, a tuple is returned where the first element is the updated sample. + """ + + if isinstance(timestep, (int, torch.IntTensor, torch.LongTensor)): + raise ValueError( + ( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `LTXEulerAncestralRFScheduler.step()` is not supported. Make sure to pass" + " one of the `scheduler.timesteps` values as `timestep`." + ), + ) + + if self.sigmas is None or self.timesteps is None: + raise ValueError("Scheduler has not been initialized. Call `set_timesteps` before `step`.") + + if self._step_index is None: + self._init_step_index(timestep) + + i = self._step_index + if i >= len(self.sigmas) - 1: + # Already at the end; simply return the current sample. + prev_sample = sample + else: + # Work in float32 for numerical stability + sample_f = sample.to(torch.float32) + model_output_f = model_output.to(torch.float32) + + sigma = self.sigmas[i] + sigma_next = self.sigmas[i + 1] + + sigma_b = self._sigma_broadcast(sigma.view(1), sample_f) + sigma_next_b = self._sigma_broadcast(sigma_next.view(1), sample_f) + + # Approximate denoised x0 under CONST parametrization: + # x0 = x_t - sigma_t * v_t + denoised = sample_f - sigma_b * model_output_f + + if sigma_next.abs().item() < 1e-8: + # Final denoising step + x = denoised + else: + eta = float(self.config.eta) + s_noise = float(self.config.s_noise) + + # Downstep computation (ComfyUI RF variant) + downstep_ratio = 1.0 + (sigma_next / sigma - 1.0) * eta + sigma_down = sigma_next * downstep_ratio + + alpha_ip1 = 1.0 - sigma_next + alpha_down = 1.0 - sigma_down + + # Deterministic part (Euler step in (x, x0)-space) + sigma_down_b = self._sigma_broadcast(sigma_down.view(1), sample_f) + alpha_ip1_b = self._sigma_broadcast(alpha_ip1.view(1), sample_f) + alpha_down_b = self._sigma_broadcast(alpha_down.view(1), sample_f) + + sigma_ratio = sigma_down_b / sigma_b + x = sigma_ratio * sample_f + (1.0 - sigma_ratio) * denoised + + # Stochastic ancestral noise + if eta > 0.0 and s_noise > 0.0: + renoise_coeff = ( + (sigma_next_b**2 - sigma_down_b**2 * alpha_ip1_b**2 / (alpha_down_b**2 + 1e-12)) + .clamp(min=0.0) + .sqrt() + ) + + noise = randn_tensor( + sample_f.shape, generator=generator, device=sample_f.device, dtype=sample_f.dtype + ) + x = (alpha_ip1_b / (alpha_down_b + 1e-12)) * x + noise * renoise_coeff * s_noise + + prev_sample = x.to(sample.dtype) + + # Advance internal step index + self._step_index = min(self._step_index + 1, len(self.sigmas) - 1) + + if not return_dict: + return (prev_sample,) + + return LTXEulerAncestralRFSchedulerOutput(prev_sample=prev_sample) + + def __len__(self) -> int: + """ + Return the number of training timesteps for compatibility with scheduler utilities. + + Returns: + `int`: + The configured number of training timesteps. + """ + return int(getattr(self.config, "num_train_timesteps", 1000)) diff --git a/diffusers/schedulers/scheduling_minimax_h3.py b/diffusers/schedulers/scheduling_minimax_h3.py new file mode 100644 index 0000000000000000000000000000000000000000..db3217060c157ca05170b43e811c21b15a499aed --- /dev/null +++ b/diffusers/schedulers/scheduling_minimax_h3.py @@ -0,0 +1,288 @@ +# Copyright 2025 The MiniMax authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +"""Rectified-flow Euler scheduler for MiniMax-H3. + +Three things make this incompatible with a [`FlowMatchEulerDiscreteScheduler`] config, which is why +it is a separate class: + +1. **The velocity sign is reversed.** MiniMax-H3's transformer predicts a *data-ward* velocity, so + `x0 = x_t + sigma * v` instead of diffusers' `x0 = x_t - sigma * v`. +2. **Timesteps are `t = 1 - sigma` in `[0, 1]`**, with `t = 1` meaning *clean*. Flow-match + schedulers expose `timesteps = sigma * num_train_timesteps`, i.e. the opposite direction on a + 1000x scale. The transformer's AdaLN consumes the H3 convention directly. +3. **The sigma grid starts from `linspace(1, 0, num_inference_steps)`** — the terminal zero is part + of the requested step count, and duplicates created by the shift are collapsed with + `unique_consecutive`. `FlowMatchEulerDiscreteScheduler` instead builds + `linspace(1, 1/num_train_timesteps, ...)` and appends a terminal sigma afterwards, so + `len(sigmas)` and every interior value differ. + +Everything else is ordinary rectified flow: the exponential shift `sigma' = s*sigma / (1 + (s-1)*sigma)`, +and an Euler update written as the `x_t` / `x0` blend `x_next = r*x_t + (1 - r)*x0` with +`r = sigma_next / sigma`, evaluated in float32. Despite the reference class being named +"euler ancestral", `eta` is 0 — no noise is ever re-injected. + +MiniMax-H3 runs **two schedules per request**, one per modality (`shift=12.0` for video, +`shift=3.0` for audio). The modality is not a property of the scheduler: a pipeline holds two +instances, e.g. `scheduler` and `audio_scheduler`. +""" + +from dataclasses import dataclass + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +@dataclass +class MiniMaxH3SchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.FloatTensor`): + Computed sample `x_{t+1}` for the next step of the denoising loop. + """ + + prev_sample: torch.FloatTensor + + +class MiniMaxH3Scheduler(SchedulerMixin, ConfigMixin): + r""" + Rectified-flow Euler scheduler (`eta = 0`) with an exponential sigma shift, as used by MiniMax-H3. + + Args: + shift (`float`, defaults to `12.0`): + Exponential shift applied to the sigma grid, `sigma' = s*sigma / (1 + (s-1)*sigma)`. The + released checkpoints use `12.0` for video latents and `3.0` for audio latents. + """ + + _compatibles = [] + order = 1 + + @register_to_config + def __init__(self, shift: float = 12.0): + if shift <= 0: + raise ValueError(f"`shift` must be positive, got {shift}.") + + self.num_inference_steps: int | None = None + self.sigmas: torch.Tensor | None = None + self.timesteps: torch.Tensor | None = None + self._shift = float(shift) + self._step_index: int | None = None + self._begin_index: int | None = None + + @property + def shift(self) -> float: + """The exponential shift currently applied to the sigma grid.""" + return self._shift + + @property + def step_index(self) -> int | None: + """Index of the step the scheduler is about to take. Increases by one after each `step`.""" + return self._step_index + + @property + def begin_index(self) -> int | None: + """Index of the first step, set from a pipeline through [`~MiniMaxH3Scheduler.set_begin_index`].""" + return self._begin_index + + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_shift(self, shift: float) -> None: + """ + Overrides the configured sigma shift; call before [`~MiniMaxH3Scheduler.set_timesteps`]. + + MiniMax-H3 exposes this per request as `flow_shift` (video) / `audio_flow_shift` (audio). + + Args: + shift (`float`): + The exponential shift to use for the next schedule. + """ + if shift <= 0: + raise ValueError(f"`shift` must be positive, got {shift}.") + self._shift = float(shift) + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + sigmas: list[float] | torch.Tensor | None = None, + ) -> None: + r""" + Build the sigma / timestep schedule. + + The grid is `linspace(1, 0, num_inference_steps)` pushed through the exponential shift, with + consecutive duplicates collapsed. The terminal `0` is already part of that grid — the shift maps + `0` to exactly `0` — so the schedule holds `num_inference_steps` sigmas and drives + `num_inference_steps - 1` model evaluations, exposed as `self.timesteps = 1 - sigmas[:-1]`. + + Args: + num_inference_steps (`int`, *optional*): + Number of sigma grid points, terminal `0` included. Ignored when `sigmas` is given. + device (`str` or `torch.device`, *optional*): + Device the schedule tensors are moved to. The grid itself is always built on CPU in + float32 so the schedule does not depend on the accelerator. + sigmas (`list[float]` or `torch.Tensor`, *optional*): + A fully-formed sigma schedule, used verbatim (no shifting, no deduplication). It must be + strictly decreasing and terminate at `0.0`. + """ + if sigmas is None: + if num_inference_steps is None or num_inference_steps < 2: + raise ValueError( + "`set_timesteps` requires either an explicit `sigmas` schedule or " + f"`num_inference_steps` >= 2, got {num_inference_steps}." + ) + + # The rectified-flow sigma range is fixed at [1.0, 0.0]. + base = torch.linspace(1.0, 0.0, int(num_inference_steps), dtype=torch.float32) + sigmas = self._shift * base / (1 + (self._shift - 1) * base) + # The shift compresses the grid near sigma = 1; collapse any float32 collisions it creates. + sigmas = torch.unique_consecutive(sigmas) + else: + sigmas = torch.as_tensor(sigmas, dtype=torch.float32).flatten().cpu() + if sigmas.numel() < 2 or not bool((sigmas[1:] < sigmas[:-1]).all()) or sigmas[-1].item() != 0.0: + raise ValueError("`sigmas` must hold at least two strictly decreasing values ending at 0.0.") + + self.sigmas = sigmas.to(device=device) + # t = 1 - sigma, and t = 1 is clean. The terminal sigma has no model evaluation. + self.timesteps = (1.0 - sigmas[:-1]).to(device=device) + self.num_inference_steps = int(self.timesteps.numel()) + self._step_index = None + self._begin_index = None + + def index_for_timestep(self, timestep: float | torch.Tensor) -> int: + """ + Map a timestep value to its index in the schedule. + + Args: + timestep (`float` or `torch.Tensor`): + A value taken from `self.timesteps`. The schedule is strictly increasing in `t`, so the + match is unique. + + Returns: + `int`: The index of `timestep`. + """ + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + indices = (self.timesteps == timestep).nonzero() + if len(indices) == 0: + raise ValueError( + "Passed `timestep` is not in `self.timesteps`. Make sure to use values from `scheduler.timesteps`." + ) + return indices[0].item() + + def scale_noise( + self, + sample: torch.FloatTensor, + timestep: float | torch.FloatTensor, + noise: torch.FloatTensor, + ) -> torch.FloatTensor: + r""" + Rectified-flow forward process, in MiniMax-H3's `t` convention: `x_t = t*x_0 + (1 - t)*noise`. + + MiniMax-H3 uses this to noise its conditioning anchors, where `t` is the `noise_aug` level + rather than a schedule entry, so `timestep` is taken at face value and is *not* looked up in + `self.timesteps`. + + Args: + sample (`torch.FloatTensor`): + The clean sample `x_0`. + timestep (`float` or `torch.FloatTensor`): + The target time in `[0, 1]`; `1` returns `sample` unchanged. + noise (`torch.FloatTensor`): + The noise to mix in. + + Returns: + `torch.FloatTensor`: The noised sample. + """ + if not isinstance(timestep, torch.Tensor): + timestep = torch.tensor(timestep, dtype=sample.dtype, device=sample.device) + timestep = timestep.to(device=sample.device, dtype=sample.dtype) + while timestep.ndim < sample.ndim: + timestep = timestep.unsqueeze(-1) + return timestep * sample + (1.0 - timestep) * noise + + def step( + self, + model_output: torch.FloatTensor, + timestep: float | torch.FloatTensor, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> MiniMaxH3SchedulerOutput | tuple: + r""" + Take one Euler (`eta = 0`) step. + + The model output is a data-ward velocity, so the denoised estimate is + `x0 = x_t + (1 - t) * v` — note the `+`, the opposite of the usual flow-match convention. + The update is then the blend `x_next = r*x_t + (1 - r)*x0` with `r = sigma_next / sigma`, + evaluated in float32 for half-precision samples. + + Args: + model_output (`torch.FloatTensor`): + The transformer's velocity prediction at `timestep`. + timestep (`float` or `torch.FloatTensor`): + The current timestep, one of `self.timesteps` (so `timestep == 1 - sigma`). + sample (`torch.FloatTensor`): + The current sample `x_t`. + return_dict (`bool`, defaults to `True`): + Whether to return a [`MiniMaxH3SchedulerOutput`] instead of a plain tuple. + + Returns: + [`MiniMaxH3SchedulerOutput`] or `tuple`: the sample for the next step. + """ + if isinstance(timestep, int) or (isinstance(timestep, torch.Tensor) and not timestep.is_floating_point()): + raise ValueError( + "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to" + " `MiniMaxH3Scheduler.step()` is not supported. Make sure to pass one of the" + " `scheduler.timesteps` values." + ) + + if self._step_index is None: + self._step_index = self.index_for_timestep(timestep) if self._begin_index is None else self._begin_index + + # x0 from the data-ward velocity. The sigma used here is recovered from the *timestep* the + # transformer was conditioned on, whereas the Euler ratio below uses the sigma grid: for + # sigma < 0.5 the float32 round trip `1 - (1 - sigma)` is not exact, and the reference keeps + # the two sources apart. + if not isinstance(timestep, torch.Tensor): + timestep = torch.tensor(timestep, dtype=sample.dtype) + sigma_from_timestep = 1 - timestep.to(device=sample.device, dtype=sample.dtype) + while sigma_from_timestep.ndim < sample.ndim: + sigma_from_timestep = sigma_from_timestep.unsqueeze(-1) + denoised = sample + sigma_from_timestep * model_output + + # Euler with eta = 0, written as an x_t / x0 blend and evaluated in float32. + compute_dtype = torch.float32 if sample.dtype in (torch.float16, torch.bfloat16) else sample.dtype + sigma = self.sigmas[self._step_index].to(device=sample.device, dtype=compute_dtype) + sigma_next = self.sigmas[self._step_index + 1].to(device=sample.device, dtype=compute_dtype) + ratio = sigma_next / sigma + prev_sample = ratio * sample.to(dtype=compute_dtype) + (1.0 - ratio) * denoised.to(dtype=compute_dtype) + prev_sample = prev_sample.to(dtype=sample.dtype) + + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + return MiniMaxH3SchedulerOutput(prev_sample=prev_sample) diff --git a/diffusers/schedulers/scheduling_pndm.py b/diffusers/schedulers/scheduling_pndm.py new file mode 100644 index 0000000000000000000000000000000000000000..f87dff14375df852d0cdffe769f5c71586ef67b7 --- /dev/null +++ b/diffusers/schedulers/scheduling_pndm.py @@ -0,0 +1,504 @@ +# Copyright 2025 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class PNDMScheduler(SchedulerMixin, ConfigMixin): + """ + `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step + method. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.0001`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.02`): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + skip_prk_steps (`bool`, defaults to `False`): + Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before + PLMS steps. + set_alpha_to_one (`bool`, defaults to `False`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + prediction_type (`"epsilon"` or `"v_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process) + or `v_prediction` (see section 2.4 of [Imagen Video](https://huggingface.co/papers/2210.02303) paper). + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + skip_prk_steps: bool = False, + set_alpha_to_one: bool = False, + prediction_type: Literal["epsilon", "v_prediction"] = "epsilon", + timestep_spacing: Literal["linspace", "leading", "trailing"] = "leading", + steps_offset: int = 0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://huggingface.co/papers/2202.09778 + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.cur_model_output = 0 + self.counter = 0 + self.cur_sample = None + self.ets = [] + + # setable values + self.num_inference_steps = None + self._timesteps = np.arange(0, num_train_timesteps)[::-1].copy() + self.prk_timesteps = None + self.plms_timesteps = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + self._timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps).round().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).round() + self._timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio))[::-1].astype( + np.int64 + ) + self._timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + if self.config.skip_prk_steps: + # for some models like stable diffusion the prk steps can/should be skipped to + # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation + # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51 + self.prk_timesteps = np.array([]) + self.plms_timesteps = np.concatenate([self._timesteps[:-1], self._timesteps[-2:-1], self._timesteps[-1:]])[ + ::-1 + ].copy() + else: + prk_timesteps = np.array(self._timesteps[-self.pndm_order :]).repeat(2) + np.tile( + np.array([0, self.config.num_train_timesteps // num_inference_steps // 2]), self.pndm_order + ) + self.prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1].copy() + self.plms_timesteps = self._timesteps[:-3][ + ::-1 + ].copy() # we copy to avoid having negative strides which are not supported by torch.from_numpy + + timesteps = np.concatenate([self.prk_timesteps, self.plms_timesteps]).astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.ets = [] + self.counter = 0 + self.cur_model_output = 0 + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`] + or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.counter < len(self.prk_timesteps) and not self.config.skip_prk_steps: + return self.step_prk(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + else: + return self.step_plms(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + + def step_prk( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential + equation. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + diff_to_prev = 0 if self.counter % 2 else self.config.num_train_timesteps // self.num_inference_steps // 2 + prev_timestep = timestep - diff_to_prev + timestep = self.prk_timesteps[self.counter // 4 * 4] + + if self.counter % 4 == 0: + self.cur_model_output += 1 / 6 * model_output + self.ets.append(model_output) + self.cur_sample = sample + elif (self.counter - 1) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 2) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 3) % 4 == 0: + model_output = self.cur_model_output + 1 / 6 * model_output + self.cur_model_output = 0 + + # cur_sample should not be `None` + cur_sample = self.cur_sample if self.cur_sample is not None else sample + + prev_sample = self._get_prev_sample(cur_sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def step_plms( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if not self.config.skip_prk_steps and len(self.ets) < 3: + raise ValueError( + f"{self.__class__} can only be run AFTER scheduler has been run " + "in 'prk' mode for at least 12 iterations " + "See: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_pndm.py " + "for more information." + ) + + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + if self.counter != 1: + self.ets = self.ets[-3:] + self.ets.append(model_output) + else: + prev_timestep = timestep + timestep = timestep + self.config.num_train_timesteps // self.num_inference_steps + + if len(self.ets) == 1 and self.counter == 0: + model_output = model_output + self.cur_sample = sample + elif len(self.ets) == 1 and self.counter == 1: + model_output = (model_output + self.ets[-1]) / 2 + sample = self.cur_sample + self.cur_sample = None + elif len(self.ets) == 2: + model_output = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + model_output = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + model_output = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample( + self, sample: torch.Tensor, timestep: int, prev_timestep: int, model_output: torch.Tensor + ) -> torch.Tensor: + """ + Compute the previous sample x_(t-δ) from the current sample x_t using formula (9) from the [PNDM + paper](https://huggingface.co/papers/2202.09778). + + Args: + sample (`torch.Tensor`): + The current sample x_t. + timestep (`int`): + The current timestep t. + prev_timestep (`int`): + The previous timestep (t-δ). + model_output (`torch.Tensor`): + The model output e_θ(x_t, t). + + Returns: + `torch.Tensor`: + The previous sample x_(t-δ). + """ + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if self.config.prediction_type == "v_prediction": + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + elif self.config.prediction_type != "epsilon": + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`" + ) + + # corresponds to (α_(t−δ) - α_t) divided by + # denominator of x_t in formula (9) and plus 1 + # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) = + # sqrt(α_(t−δ)) / sqrt(α_t)) + sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) + + # corresponds to denominator of e_θ(x_t, t) in formula (9) + model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + ( + alpha_prod_t * beta_prod_t * alpha_prod_t_prev + ) ** (0.5) + + # full formula (9) + prev_sample = ( + sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff + ) + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_repaint.py b/diffusers/schedulers/scheduling_repaint.py new file mode 100644 index 0000000000000000000000000000000000000000..97bd97909f4d70d4045cba5b9ae5538f215fc790 --- /dev/null +++ b/diffusers/schedulers/scheduling_repaint.py @@ -0,0 +1,419 @@ +# Copyright 2025 ETH Zurich Computer Vision Lab and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +@dataclass +class RePaintSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample (x_{0}) based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class RePaintScheduler(SchedulerMixin, ConfigMixin): + """ + `RePaintScheduler` is a scheduler for DDPM inpainting inside a given mask. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.0001`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.02`): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, `squaredcos_cap_v2`, or `sigmoid`. + eta (`float`, defaults to `0.0`): + The weight of noise added during a diffusion step. A value of `0.0` corresponds to DDIM and a value of + `1.0` corresponds to DDPM. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + clip_sample (`bool`, defaults to `True`): + Clip the predicted sample between -1 and 1 for numerical stability. + + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + eta: float = 0.0, + trained_betas: np.ndarray | None = None, + clip_sample: bool = True, + ) -> None: + if trained_betas is not None: + self.betas = torch.from_numpy(trained_betas) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + elif beta_schedule == "sigmoid": + # GeoDiff sigmoid schedule + betas = torch.linspace(-6, 6, num_train_timesteps) + self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + self.final_alpha_cumprod = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.eta = eta + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, + num_inference_steps: int, + jump_length: int = 10, + jump_n_sample: int = 10, + device: str | torch.device | None = None, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + jump_length (`int`, defaults to `10`): + The number of steps taken forward in time before going backward in time for a single jump (“j” in + RePaint paper). Take a look at Figure 9 and 10 in the paper. + jump_n_sample (`int`, defaults to `10`): + The number of times to make a forward time jump for a given chosen time sample. Take a look at Figure 9 + and 10 in the paper. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + + """ + num_inference_steps = min(self.config.num_train_timesteps, num_inference_steps) + self.num_inference_steps = num_inference_steps + + timesteps = [] + + jumps = {} + for j in range(0, num_inference_steps - jump_length, jump_length): + jumps[j] = jump_n_sample - 1 + + t = num_inference_steps + while t >= 1: + t = t - 1 + timesteps.append(t) + + if jumps.get(t, 0) > 0: + jumps[t] = jumps[t] - 1 + for _ in range(jump_length): + t = t + 1 + timesteps.append(t) + + timesteps = np.array(timesteps) * (self.config.num_train_timesteps // self.num_inference_steps) + self.timesteps = torch.from_numpy(timesteps).to(device) + + def _get_variance(self, t: int) -> torch.Tensor: + """ + Compute the variance for a given timestep. + + Args: + t (`int`): + The current timestep. + + Returns: + `torch.Tensor`: + The computed variance. + """ + prev_timestep = t - self.config.num_train_timesteps // self.num_inference_steps + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from + # https://huggingface.co/papers/2006.11239) and sample from it to get + # previous sample x_{t-1} ~ N(pred_prev_sample, variance) == add + # variance to pred_sample + # Is equivalent to formula (16) in https://huggingface.co/papers/2010.02502 + # without eta. + # variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * self.betas[t] + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + original_image: torch.Tensor, + mask: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> RePaintSchedulerOutput | tuple[torch.Tensor, torch.Tensor]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + original_image (`torch.Tensor`): + The original image to inpaint on. + mask (`torch.Tensor`): + The mask where a value of 0.0 indicates which part of the original image to inpaint. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] instead of a + tuple. + + Returns: + [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] or `tuple[torch.Tensor, torch.Tensor]`: + If return_dict is `True`, [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] is returned, + otherwise a tuple is returned where the first element is the previous sample and the second is the + predicted original sample. + """ + t = timestep + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + # 1. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + pred_original_sample = (sample - beta_prod_t**0.5 * model_output) / alpha_prod_t**0.5 + + # 3. Clip "predicted x_0" + if self.config.clip_sample: + pred_original_sample = torch.clamp(pred_original_sample, -1, 1) + + # We choose to follow RePaint Algorithm 1 to get x_{t-1}, however we + # substitute formula (7) in the algorithm coming from DDPM paper + # (formula (4) Algorithm 2 - Sampling) with formula (12) from DDIM paper. + # DDIM schedule gives the same results as DDPM with eta = 1.0 + # Noise is being reused in 7. and 8., but no impact on quality has + # been observed. + + # 5. Add noise + device = model_output.device + noise = randn_tensor(model_output.shape, generator=generator, device=device, dtype=model_output.dtype) + std_dev_t = self.eta * self._get_variance(timestep) ** 0.5 + + variance = 0 + if t > 0 and self.eta > 0: + variance = std_dev_t * noise + + # 6. compute "direction pointing to x_t" of formula (12) + # from https://huggingface.co/papers/2010.02502 + pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** 0.5 * model_output + + # 7. compute x_{t-1} of formula (12) from https://huggingface.co/papers/2010.02502 + prev_unknown_part = alpha_prod_t_prev**0.5 * pred_original_sample + pred_sample_direction + variance + + # 8. Algorithm 1 Line 5 https://huggingface.co/papers/2201.09865 + # The computation reported in Algorithm 1 Line 5 is incorrect. Line 5 refers to formula (8a) of the same paper, + # which tells to sample from a Gaussian distribution with mean "(alpha_prod_t_prev**0.5) * original_image" + # and variance "(1 - alpha_prod_t_prev)". This means that the standard Gaussian distribution "noise" should be + # scaled by the square root of the variance (as it is done here), however Algorithm 1 Line 5 tells to scale by the variance. + prev_known_part = (alpha_prod_t_prev**0.5) * original_image + ((1 - alpha_prod_t_prev) ** 0.5) * noise + + # 9. Algorithm 1 Line 8 https://huggingface.co/papers/2201.09865 + pred_prev_sample = mask * prev_known_part + (1.0 - mask) * prev_unknown_part + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return RePaintSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + def undo_step( + self, + sample: torch.Tensor, + timestep: int, + generator: torch.Generator | None = None, + ) -> torch.Tensor: + """ + Add noise to a sample to move it forward by one diffusion step. + + Args: + sample (`torch.Tensor`): + The sample to which noise is added. + timestep (`int`): + The current discrete timestep in the diffusion chain. + generator (`torch.Generator`, *optional*): + A random number generator. + + Returns: + `torch.Tensor`: + The sample at the next timestep in the diffusion chain. + """ + n = self.config.num_train_timesteps // self.num_inference_steps + + for i in range(n): + beta = self.betas[timestep + i] + if sample.device.type == "mps": + # randn does not work reproducibly on mps + noise = randn_tensor(sample.shape, dtype=sample.dtype, generator=generator) + noise = noise.to(sample.device) + else: + noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype) + + # 10. Algorithm 1 Line 10 https://huggingface.co/papers/2201.09865 + sample = (1 - beta) ** 0.5 * sample + beta**0.5 * noise + + return sample + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Indicate that adding noise for RePaint training is not supported by this scheduler. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise would be added. + noise (`torch.Tensor`): + The noise that would be added to the samples. + timesteps (`torch.IntTensor`): + The timesteps that would determine the noise level for each sample. + + Raises: + `NotImplementedError`: + RePaint training should use [`DDPMScheduler.add_noise`] instead. + """ + raise NotImplementedError("Use `DDPMScheduler.add_noise()` to train for sampling with RePaint.") + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_sasolver.py b/diffusers/schedulers/scheduling_sasolver.py new file mode 100644 index 0000000000000000000000000000000000000000..93ea0b99e2c0c3de8cb59751cda394ecc0b0aa0c --- /dev/null +++ b/diffusers/schedulers/scheduling_sasolver.py @@ -0,0 +1,1423 @@ +# Copyright 2025 Shuchen Xue, etc. in University of Chinese Academy of Sciences Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: check https://huggingface.co/papers/2309.05019 +# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py + +import math +from typing import Callable, List, Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class SASolverScheduler(SchedulerMixin, ConfigMixin): + """ + `SASolverScheduler` is a fast dedicated high-order solver for diffusion SDEs. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + predictor_order (`int`, defaults to 2): + The predictor order which can be `1` or `2` or `3` or '4'. It is recommended to use `predictor_order=2` for + guided sampling, and `predictor_order=3` for unconditional sampling. + corrector_order (`int`, defaults to 2): + The corrector order which can be `1` or `2` or `3` or '4'. It is recommended to use `corrector_order=2` for + guided sampling, and `corrector_order=3` for unconditional sampling. + prediction_type (`"epsilon"`, `"sample"`, `"v_prediction"`, or `"flow_prediction"`, defaults to `"epsilon"`): + Prediction type of the scheduler function. `epsilon` predicts the noise of the diffusion process, `sample` + directly predicts the noisy sample, `v_prediction` predicts the velocity (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper), and `flow_prediction` predicts the flow. + tau_func (`Callable`, *optional*): + Stochasticity during the sampling. Default in init is `lambda t: 1 if t >= 200 and t <= 800 else 0`. + SA-Solver will sample from vanilla diffusion ODE if tau_func is set to `lambda t: 0`. SA-Solver will sample + from vanilla diffusion SDE if tau_func is set to `lambda t: 1`. For more details, please check + https://huggingface.co/papers/2309.05019 + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and + `algorithm_type="dpmsolver++"`. + algorithm_type (`str`, defaults to `data_prediction`): + Algorithm type for the solver; can be `data_prediction` or `noise_prediction`. It is recommended to use + `data_prediction` with `solver_order=2` for guided sampling like in Stable Diffusion. + lower_order_final (`bool`, defaults to `True`): + Whether to use lower-order solvers in the final steps. Default = True. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + flow_shift (`float`, *optional*, defaults to `1.0`): + The shift value for the timestep schedule for flow matching. + lambda_min_clipped (`float`, defaults to `-inf`): + Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the + cosine (`squaredcos_cap_v2`) noise schedule. + variance_type (`str`, *optional*): + Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output + contains the predicted Gaussian variance. + timestep_spacing (`str`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + predictor_order: int = 2, + corrector_order: int = 2, + prediction_type: str = "epsilon", + tau_func: Callable | None = None, + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + algorithm_type: str = "data_prediction", + lower_order_final: bool = True, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + use_flow_sigmas: bool = False, + flow_shift: float = 1.0, + lambda_min_clipped: float = -float("inf"), + variance_type: str | None = None, + timestep_spacing: str = "linspace", + steps_offset: int = 0, + ): + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = ( + torch.linspace( + beta_start**0.5, + beta_end**0.5, + num_train_timesteps, + dtype=torch.float32, + ) + ** 2 + ) + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + if algorithm_type not in ["data_prediction", "noise_prediction"]: + raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}") + + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.timestep_list = [None] * max(predictor_order, corrector_order - 1) + self.model_outputs = [None] * max(predictor_order, corrector_order - 1) + + if tau_func is None: + self.tau_func = lambda t: 1 if t >= 200 and t <= 800 else 0 + else: + self.tau_func = tau_func + self.predict_x0 = algorithm_type == "data_prediction" + self.lower_order_nums = 0 + self.last_sample = None + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self): + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self): + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps(self, num_inference_steps: int = None, device: str | torch.device = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + # Clipping the minimum of all lambda(t) for numerical stability. + # This is critical for cosine (squaredcos_cap_v2) noise schedule. + clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped) + last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item() + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].copy().astype(np.int64) + ) + + elif self.config.timestep_spacing == "leading": + step_ratio = last_timestep // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + if self.config.use_karras_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_beta_sigmas: + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + elif self.config.use_flow_sigmas: + alphas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1) + sigmas = 1.0 - alphas + sigmas = np.flip(self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas))[:-1].copy() + timesteps = (sigmas * self.config.num_train_timesteps).copy() + sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32) + else: + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + self.model_outputs = [ + None, + ] * max(self.config.predictor_order, self.config.corrector_order - 1) + self.lower_order_nums = 0 + self.last_sample = None + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma, log_sigmas): + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma): + """ + Convert sigma values to alpha_t and sigma_t values. + + Args: + sigma (`torch.Tensor`): + The sigma value(s) to convert. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing (alpha_t, sigma_t) values. + """ + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor | None = None, + **kwargs, + ) -> torch.Tensor: + """ + Convert the model output to the corresponding type the data_prediction/noise_prediction algorithm needs. + Noise_prediction is designed to discretize an integral of the noise prediction model, and data_prediction is + designed to discretize an integral of the data prediction model. + + > [!TIP] > The algorithm and model type are decoupled. You can use either data_prediction or noise_prediction + for both > noise prediction and data prediction models. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + # SA-Solver_data_prediction needs to solve an integral of the data prediction model. + if self.config.algorithm_type in ["data_prediction"]: + if self.config.prediction_type == "epsilon": + # SA-Solver only needs the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + model_output = model_output[:, :3] + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the SASolverScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + + # SA-Solver_noise_prediction needs to solve an integral of the noise prediction model. + elif self.config.algorithm_type in ["noise_prediction"]: + if self.config.prediction_type == "epsilon": + # SA-Solver only needs the "mean" output. + if self.config.variance_type in ["learned", "learned_range"]: + epsilon = model_output[:, :3] + else: + epsilon = model_output + elif self.config.prediction_type == "sample": + epsilon = (sample - alpha_t * model_output) / sigma_t + elif self.config.prediction_type == "v_prediction": + epsilon = alpha_t * model_output + sigma_t * sample + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the SASolverScheduler." + ) + + if self.config.thresholding: + alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep] + x0_pred = (sample - sigma_t * epsilon) / alpha_t + x0_pred = self._threshold_sample(x0_pred) + epsilon = (sample - alpha_t * x0_pred) / sigma_t + + return epsilon + + def get_coefficients_exponential_negative( + self, order: int, interval_start: torch.Tensor, interval_end: torch.Tensor + ) -> torch.Tensor: + r""" + Compute the integral of $\exp(-x) \cdot x^{order}$ over the interval `[interval_start, interval_end]`. This is + used as a building block of the SA-Solver algorithm (negative exponential case). + + Args: + order (`int`): + The order of the polynomial factor $x^{order}$ in the integrand. Must be in `[0, 1, 2, 3]`. + interval_start (`torch.Tensor`): + The lower bound of the integration interval. + interval_end (`torch.Tensor`): + The upper bound of the integration interval. + + Returns: + `torch.Tensor`: + The value of the integral $\int_{interval_start}^{interval_end} \exp(-x) \cdot x^{order} \, dx$. + """ + assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3" + + if order == 0: + return torch.exp(-interval_end) * (torch.exp(interval_end - interval_start) - 1) + elif order == 1: + return torch.exp(-interval_end) * ( + (interval_start + 1) * torch.exp(interval_end - interval_start) - (interval_end + 1) + ) + elif order == 2: + return torch.exp(-interval_end) * ( + (interval_start**2 + 2 * interval_start + 2) * torch.exp(interval_end - interval_start) + - (interval_end**2 + 2 * interval_end + 2) + ) + elif order == 3: + return torch.exp(-interval_end) * ( + (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6) + * torch.exp(interval_end - interval_start) + - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6) + ) + + def get_coefficients_exponential_positive( + self, order: int, interval_start: torch.Tensor, interval_end: torch.Tensor, tau: torch.Tensor + ) -> torch.Tensor: + r""" + Compute the integral of $\exp(x(1+\tau^2)) \cdot x^{order}$ over the interval `[interval_start, interval_end]`. + This is used as a building block of the SA-Solver algorithm (positive exponential case). + + Args: + order (`int`): + The order of the polynomial factor $x^{order}$ in the integrand. Must be in `[0, 1, 2, 3]`. + interval_start (`torch.Tensor`): + The lower bound of the integration interval. + interval_end (`torch.Tensor`): + The upper bound of the integration interval. + tau (`torch.Tensor`): + The stochasticity coefficient $\tau$ that scales the exponent in the integrand. + + Returns: + `torch.Tensor`: + The value of the integral $\int_{interval_start}^{interval_end} \exp(x(1+\tau^2)) \cdot x^{order} \, + dx$. + """ + assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3" + + # after change of variable(cov) + interval_end_cov = (1 + tau**2) * interval_end + interval_start_cov = (1 + tau**2) * interval_start + + if order == 0: + return ( + torch.exp(interval_end_cov) * (1 - torch.exp(-(interval_end_cov - interval_start_cov))) / (1 + tau**2) + ) + elif order == 1: + return ( + torch.exp(interval_end_cov) + * ( + (interval_end_cov - 1) + - (interval_start_cov - 1) * torch.exp(-(interval_end_cov - interval_start_cov)) + ) + / ((1 + tau**2) ** 2) + ) + elif order == 2: + return ( + torch.exp(interval_end_cov) + * ( + (interval_end_cov**2 - 2 * interval_end_cov + 2) + - (interval_start_cov**2 - 2 * interval_start_cov + 2) + * torch.exp(-(interval_end_cov - interval_start_cov)) + ) + / ((1 + tau**2) ** 3) + ) + elif order == 3: + return ( + torch.exp(interval_end_cov) + * ( + (interval_end_cov**3 - 3 * interval_end_cov**2 + 6 * interval_end_cov - 6) + - (interval_start_cov**3 - 3 * interval_start_cov**2 + 6 * interval_start_cov - 6) + * torch.exp(-(interval_end_cov - interval_start_cov)) + ) + / ((1 + tau**2) ** 4) + ) + + def lagrange_polynomial_coefficient(self, order: int, lambda_list: List[torch.Tensor]) -> List[List[torch.Tensor]]: + r""" + Compute the coefficients of the Lagrange polynomial of order `order` that interpolates the points $(x_i, y_i) = + (\lambda_i, i)$ for $i \in [0, order]`. The returned coefficients describe the polynomial in its expanded form + $\sum_{k=0}^{order} c_k \cdot x^k$. + + Args: + order (`int`): + The order of the Lagrange polynomial. Must be in `[0, 1, 2, 3]`. + lambda_list (`List[torch.Tensor]`): + A list of `order + 1` $\lambda$ values (i.e. the $x$-coordinates of the interpolation nodes). Must + satisfy `len(lambda_list) == order + 1`. + + Returns: + `List[List[torch.Tensor]]`: + A list of `order + 1` coefficient lists — one for each output channel — where each inner list contains + the polynomial coefficients $c_0, c_1, \ldots, c_{order}$. + """ + + assert order in [0, 1, 2, 3] + assert order == len(lambda_list) - 1 + if order == 0: + return [[1]] + elif order == 1: + return [ + [ + 1 / (lambda_list[0] - lambda_list[1]), + -lambda_list[1] / (lambda_list[0] - lambda_list[1]), + ], + [ + 1 / (lambda_list[1] - lambda_list[0]), + -lambda_list[0] / (lambda_list[1] - lambda_list[0]), + ], + ] + elif order == 2: + denominator1 = (lambda_list[0] - lambda_list[1]) * (lambda_list[0] - lambda_list[2]) + denominator2 = (lambda_list[1] - lambda_list[0]) * (lambda_list[1] - lambda_list[2]) + denominator3 = (lambda_list[2] - lambda_list[0]) * (lambda_list[2] - lambda_list[1]) + return [ + [ + 1 / denominator1, + (-lambda_list[1] - lambda_list[2]) / denominator1, + lambda_list[1] * lambda_list[2] / denominator1, + ], + [ + 1 / denominator2, + (-lambda_list[0] - lambda_list[2]) / denominator2, + lambda_list[0] * lambda_list[2] / denominator2, + ], + [ + 1 / denominator3, + (-lambda_list[0] - lambda_list[1]) / denominator3, + lambda_list[0] * lambda_list[1] / denominator3, + ], + ] + elif order == 3: + denominator1 = ( + (lambda_list[0] - lambda_list[1]) + * (lambda_list[0] - lambda_list[2]) + * (lambda_list[0] - lambda_list[3]) + ) + denominator2 = ( + (lambda_list[1] - lambda_list[0]) + * (lambda_list[1] - lambda_list[2]) + * (lambda_list[1] - lambda_list[3]) + ) + denominator3 = ( + (lambda_list[2] - lambda_list[0]) + * (lambda_list[2] - lambda_list[1]) + * (lambda_list[2] - lambda_list[3]) + ) + denominator4 = ( + (lambda_list[3] - lambda_list[0]) + * (lambda_list[3] - lambda_list[1]) + * (lambda_list[3] - lambda_list[2]) + ) + return [ + [ + 1 / denominator1, + (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1, + ( + lambda_list[1] * lambda_list[2] + + lambda_list[1] * lambda_list[3] + + lambda_list[2] * lambda_list[3] + ) + / denominator1, + (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1, + ], + [ + 1 / denominator2, + (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2, + ( + lambda_list[0] * lambda_list[2] + + lambda_list[0] * lambda_list[3] + + lambda_list[2] * lambda_list[3] + ) + / denominator2, + (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2, + ], + [ + 1 / denominator3, + (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3, + ( + lambda_list[0] * lambda_list[1] + + lambda_list[0] * lambda_list[3] + + lambda_list[1] * lambda_list[3] + ) + / denominator3, + (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3, + ], + [ + 1 / denominator4, + (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4, + ( + lambda_list[0] * lambda_list[1] + + lambda_list[0] * lambda_list[2] + + lambda_list[1] * lambda_list[2] + ) + / denominator4, + (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4, + ], + ] + + def get_coefficients_fn( + self, + order: int, + interval_start: torch.Tensor, + interval_end: torch.Tensor, + lambda_list: List[torch.Tensor], + tau: torch.Tensor, + ) -> List[torch.Tensor]: + r""" + Compute the gradient coefficients used by the SA-Predictor and SA-Corrector. The coefficients are obtained by + combining the Lagrange polynomial interpolation with the closed-form exponential integrals. + + Args: + order (`int`): + The order of the solver. Must be in `[1, 2, 3, 4]` and equal to `len(lambda_list)`. + interval_start (`torch.Tensor`): + The lower bound $\lambda_{s_0}$ of the integration interval. + interval_end (`torch.Tensor`): + The upper bound $\lambda_t$ of the integration interval. + lambda_list (`List[torch.Tensor]`): + The $\lambda$ values of the past `order` timesteps used as the interpolation nodes of the Lagrange + polynomial. + tau (`torch.Tensor`): + The stochasticity coefficient $\tau$ at the current step. + + Returns: + `List[torch.Tensor]`: + A list of `order` gradient coefficients, one per past model output. + """ + assert order in [1, 2, 3, 4] + assert order == len(lambda_list), "the length of lambda list must be equal to the order" + coefficients = [] + lagrange_coefficient = self.lagrange_polynomial_coefficient(order - 1, lambda_list) + for i in range(order): + coefficient = 0 + for j in range(order): + if self.predict_x0: + coefficient += lagrange_coefficient[i][j] * self.get_coefficients_exponential_positive( + order - 1 - j, interval_start, interval_end, tau + ) + else: + coefficient += lagrange_coefficient[i][j] * self.get_coefficients_exponential_negative( + order - 1 - j, interval_start, interval_end + ) + coefficients.append(coefficient) + assert len(coefficients) == order, "the length of coefficients does not match the order" + return coefficients + + def stochastic_adams_bashforth_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor, + noise: torch.Tensor, + order: int, + tau: torch.Tensor, + **kwargs, + ) -> torch.Tensor: + """ + One step for the SA-Predictor. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model at the current timestep. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + noise (`torch.Tensor`): + A noise tensor of the same shape as `sample`, drawn from a standard normal distribution. Used to inject + stochasticity into the update. + order (`int`): + The order of SA-Predictor at this timestep. + tau (`torch.Tensor`): + The stochasticity coefficient tau at the current step. + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if noise is None: + if len(args) > 2: + noise = args[2] + else: + raise ValueError("missing `noise` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if tau is None: + if len(args) > 4: + tau = args[4] + else: + raise ValueError("missing `tau` as a required keyword argument") + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + model_output_list = self.model_outputs + sigma_t, sigma_s0 = ( + self.sigmas[self.step_index + 1], + self.sigmas[self.step_index], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + gradient_part = torch.zeros_like(sample) + h = lambda_t - lambda_s0 + lambda_list = [] + + for i in range(order): + si = self.step_index - i + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + lambda_list.append(lambda_si) + + gradient_coefficients = self.get_coefficients_fn(order, lambda_s0, lambda_t, lambda_list, tau) + + x = sample + + if self.predict_x0: + if ( + order == 2 + ): ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling. + # The added term is O(h^3). Empirically we find it will slightly improve the image quality. + # ODE case + # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2])) + # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2])) + temp_sigma = self.sigmas[self.step_index - 1] + temp_alpha_s, temp_sigma_s = self._sigma_to_alpha_sigma_t(temp_sigma) + temp_lambda_s = torch.log(temp_alpha_s) - torch.log(temp_sigma_s) + gradient_coefficients[0] += ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h**2 / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2)) + / (lambda_s0 - temp_lambda_s) + ) + gradient_coefficients[1] -= ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h**2 / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2)) + / (lambda_s0 - temp_lambda_s) + ) + + for i in range(order): + if self.predict_x0: + gradient_part += ( + (1 + tau**2) + * sigma_t + * torch.exp(-(tau**2) * lambda_t) + * gradient_coefficients[i] + * model_output_list[-(i + 1)] + ) + else: + gradient_part += -(1 + tau**2) * alpha_t * gradient_coefficients[i] * model_output_list[-(i + 1)] + + if self.predict_x0: + noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise + else: + noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise + + if self.predict_x0: + x_t = torch.exp(-(tau**2) * h) * (sigma_t / sigma_s0) * x + gradient_part + noise_part + else: + x_t = (alpha_t / alpha_s0) * x + gradient_part + noise_part + + x_t = x_t.to(x.dtype) + return x_t + + def stochastic_adams_moulton_update( + self, + this_model_output: torch.Tensor, + *args, + last_sample: torch.Tensor, + last_noise: torch.Tensor, + this_sample: torch.Tensor, + order: int, + tau: torch.Tensor, + **kwargs, + ) -> torch.Tensor: + """ + One step for the SA-Corrector. + + Args: + this_model_output (`torch.Tensor`): + The model outputs at `x_t`. + last_sample (`torch.Tensor`): + The generated sample before the last predictor `x_{t-1}`. + last_noise (`torch.Tensor`): + The noise tensor injected by the last predictor at the previous step. Used to ensure consistent + stochasticity between the predictor and the corrector. + this_sample (`torch.Tensor`): + The generated sample after the last predictor `x_{t}`. + order (`int`): + The order of SA-Corrector at this step. + tau (`torch.Tensor`): + The stochasticity coefficient tau at the current step. + + Returns: + `torch.Tensor`: + The corrected sample tensor at the current timestep. + """ + + this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None) + if last_sample is None: + if len(args) > 1: + last_sample = args[1] + else: + raise ValueError("missing `last_sample` as a required keyword argument") + if last_noise is None: + if len(args) > 2: + last_noise = args[2] + else: + raise ValueError("missing `last_noise` as a required keyword argument") + if this_sample is None: + if len(args) > 3: + this_sample = args[3] + else: + raise ValueError("missing `this_sample` as a required keyword argument") + if order is None: + if len(args) > 4: + order = args[4] + else: + raise ValueError("missing `order` as a required keyword argument") + if tau is None: + if len(args) > 5: + tau = args[5] + else: + raise ValueError("missing `tau` as a required keyword argument") + if this_timestep is not None: + deprecate( + "this_timestep", + "1.0.0", + "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + model_output_list = self.model_outputs + sigma_t, sigma_s0 = ( + self.sigmas[self.step_index], + self.sigmas[self.step_index - 1], + ) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + gradient_part = torch.zeros_like(this_sample) + h = lambda_t - lambda_s0 + lambda_list = [] + for i in range(order): + si = self.step_index - i + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si]) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + lambda_list.append(lambda_si) + + model_prev_list = model_output_list + [this_model_output] + + gradient_coefficients = self.get_coefficients_fn(order, lambda_s0, lambda_t, lambda_list, tau) + + x = last_sample + + if self.predict_x0: + if ( + order == 2 + ): ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling. + # The added term is O(h^3). Empirically we find it will slightly improve the image quality. + # ODE case + # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h) + # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h) + gradient_coefficients[0] += ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2 * h)) + ) + gradient_coefficients[1] -= ( + 1.0 + * torch.exp((1 + tau**2) * lambda_t) + * (h / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2 * h)) + ) + + for i in range(order): + if self.predict_x0: + gradient_part += ( + (1 + tau**2) + * sigma_t + * torch.exp(-(tau**2) * lambda_t) + * gradient_coefficients[i] + * model_prev_list[-(i + 1)] + ) + else: + gradient_part += -(1 + tau**2) * alpha_t * gradient_coefficients[i] * model_prev_list[-(i + 1)] + + if self.predict_x0: + noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * last_noise + else: + noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * last_noise + + if self.predict_x0: + x_t = torch.exp(-(tau**2) * h) * (sigma_t / sigma_s0) * x + gradient_part + noise_part + else: + x_t = (alpha_t / alpha_s0) * x + gradient_part + noise_part + + x_t = x_t.to(x.dtype) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep): + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator=None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the SA-Solver. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + use_corrector = self.step_index > 0 and self.last_sample is not None + + model_output_convert = self.convert_model_output(model_output, sample=sample) + + if use_corrector: + current_tau = self.tau_func(self.timestep_list[-1]) + sample = self.stochastic_adams_moulton_update( + this_model_output=model_output_convert, + last_sample=self.last_sample, + last_noise=self.last_noise, + this_sample=sample, + order=self.this_corrector_order, + tau=current_tau, + ) + + for i in range(max(self.config.predictor_order, self.config.corrector_order - 1) - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.timestep_list[i] = self.timestep_list[i + 1] + + self.model_outputs[-1] = model_output_convert + self.timestep_list[-1] = timestep + + noise = randn_tensor( + model_output.shape, + generator=generator, + device=model_output.device, + dtype=model_output.dtype, + ) + + if self.config.lower_order_final: + this_predictor_order = min(self.config.predictor_order, len(self.timesteps) - self.step_index) + this_corrector_order = min(self.config.corrector_order, len(self.timesteps) - self.step_index + 1) + else: + this_predictor_order = self.config.predictor_order + this_corrector_order = self.config.corrector_order + + self.this_predictor_order = min(this_predictor_order, self.lower_order_nums + 1) # warmup for multistep + self.this_corrector_order = min(this_corrector_order, self.lower_order_nums + 2) # warmup for multistep + assert self.this_predictor_order > 0 + assert self.this_corrector_order > 0 + + self.last_sample = sample + self.last_noise = noise + + current_tau = self.tau_func(self.timestep_list[-1]) + prev_sample = self.stochastic_adams_bashforth_update( + model_output=model_output_convert, + sample=sample, + noise=noise, + order=self.this_predictor_order, + tau=current_tau, + ) + + if self.lower_order_nums < max(self.config.predictor_order, self.config.corrector_order - 1): + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_scm.py b/diffusers/schedulers/scheduling_scm.py new file mode 100644 index 0000000000000000000000000000000000000000..d3ca01dee507a953a8b77c1db260840137ec7002 --- /dev/null +++ b/diffusers/schedulers/scheduling_scm.py @@ -0,0 +1,305 @@ +# # Copyright 2025 Sana-Sprint Authors and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +from dataclasses import dataclass + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..schedulers.scheduling_utils import SchedulerMixin +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->SCM +class SCMSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +class SCMScheduler(SchedulerMixin, ConfigMixin): + """ + `SCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with + non-Markovian guidance. This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass + documentation for the generic methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + prediction_type (`str`, defaults to `trigflow`): + Prediction type of the scheduler function. Currently only supports "trigflow". + sigma_data (`float`, defaults to `0.5`): + The standard deviation of the noise added during multi-step inference. + """ + + # _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + prediction_type: str = "trigflow", + sigma_data: float = 0.5, + ): + """ + Initialize the SCM scheduler. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + prediction_type (`str`, defaults to `trigflow`): + Prediction type of the scheduler function. Currently only supports "trigflow". + sigma_data (`float`, defaults to `0.5`): + The standard deviation of the noise added during multi-step inference. + """ + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + + self._step_index = None + self._begin_index = None + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase by 1 after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if not initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index for the scheduler, or `None` if not set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0): + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int, + timesteps: list[float] | torch.Tensor | None = None, + device: str | torch.device | None = None, + max_timesteps: float | None = 1.57080, + intermediate_timesteps: float | None = 1.3, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + timesteps (`list[float]` or `torch.Tensor`, *optional*): + Custom timesteps to use for the denoising process. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + max_timesteps (`float`, *optional*, defaults to `1.57080`): + The maximum timestep value used in the SCM scheduler. + intermediate_timesteps (`float`, *optional*, defaults to `1.3`): + The intermediate timestep value used in SCM scheduler (only used when num_inference_steps=2). + """ + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + if timesteps is not None and len(timesteps) != num_inference_steps + 1: + raise ValueError("If providing custom timesteps, `timesteps` must be of length `num_inference_steps + 1`.") + + if timesteps is not None and max_timesteps is not None: + raise ValueError("If providing custom timesteps, `max_timesteps` should not be provided.") + + if timesteps is None and max_timesteps is None: + raise ValueError("Should provide either `timesteps` or `max_timesteps`.") + + if intermediate_timesteps is not None and num_inference_steps != 2: + raise ValueError("Intermediate timesteps for SCM is not supported when num_inference_steps != 2.") + + self.num_inference_steps = num_inference_steps + + if timesteps is not None: + if isinstance(timesteps, list): + self.timesteps = torch.tensor(timesteps, device=device).float() + elif isinstance(timesteps, torch.Tensor): + self.timesteps = timesteps.to(device).float() + else: + raise ValueError(f"Unsupported timesteps type: {type(timesteps)}") + elif intermediate_timesteps is not None: + self.timesteps = torch.tensor([max_timesteps, intermediate_timesteps, 0], device=device).float() + else: + # max_timesteps=arctan(80/0.5)=1.56454 is the default from sCM paper, we choose a different value here + self.timesteps = torch.linspace(max_timesteps, 0, num_inference_steps + 1, device=device).float() + + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + def step( + self, + model_output: torch.FloatTensor, + timestep: float, + sample: torch.FloatTensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> SCMSchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`float`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator for reproducible sampling. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_scm.SCMSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_scm.SCMSchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_scm.SCMSchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + # 2. compute alphas, betas + t = self.timesteps[self.step_index + 1] + s = self.timesteps[self.step_index] + + # 4. Different Parameterization: + parameterization = self.config.prediction_type + + if parameterization == "trigflow": + pred_x0 = torch.cos(s) * sample - torch.sin(s) * model_output + else: + raise ValueError(f"Unsupported parameterization: {parameterization}") + + # 5. Sample z ~ N(0, I), For MultiStep Inference + # Noise is not used for one-step sampling. + if len(self.timesteps) > 1: + noise = ( + randn_tensor(model_output.shape, device=model_output.device, generator=generator) + * self.config.sigma_data + ) + prev_sample = torch.cos(t) * pred_x0 + torch.sin(t) * noise + else: + prev_sample = pred_x0 + + self._step_index += 1 + + if not return_dict: + return (prev_sample, pred_x0) + + return SCMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_x0) + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_sde_ve.py b/diffusers/schedulers/scheduling_sde_ve.py new file mode 100644 index 0000000000000000000000000000000000000000..3c935f3f692caeabcbd528a000737c5b1ae10b76 --- /dev/null +++ b/diffusers/schedulers/scheduling_sde_ve.py @@ -0,0 +1,333 @@ +# Copyright 2025 Google Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch + +import math +from dataclasses import dataclass + +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin, SchedulerOutput + + +@dataclass +class SdeVeOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + prev_sample_mean (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Mean averaged `prev_sample` over previous timesteps. + """ + + prev_sample: torch.Tensor + prev_sample_mean: torch.Tensor + + +class ScoreSdeVeScheduler(SchedulerMixin, ConfigMixin): + """ + `ScoreSdeVeScheduler` is a variance exploding stochastic differential equation (SDE) scheduler. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 2000): + The number of diffusion steps to train the model. + snr (`float`, defaults to 0.15): + A coefficient weighting the step from the `model_output` sample (from the network) to the random noise. + sigma_min (`float`, defaults to 0.01): + The initial noise scale for the sigma sequence in the sampling procedure. The minimum sigma should mirror + the distribution of the data. + sigma_max (`float`, defaults to 1348.0): + The maximum value used for the range of continuous timesteps passed into the model. + sampling_eps (`float`, defaults to 1e-5): + The end value of sampling where timesteps decrease progressively from 1 to epsilon. + correct_steps (`int`, defaults to 1): + The number of correction steps performed on a produced sample. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 2000, + snr: float = 0.15, + sigma_min: float = 0.01, + sigma_max: float = 1348.0, + sampling_eps: float = 1e-5, + correct_steps: int = 1, + ): + # standard deviation of the initial noise distribution + self.init_noise_sigma = sigma_max + + # setable values + self.timesteps = None + + self.set_sigmas(num_train_timesteps, sigma_min, sigma_max, sampling_eps) + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps( + self, num_inference_steps: int, sampling_eps: float | None = None, device: str | torch.device | None = None + ): + """ + Sets the continuous timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + sampling_eps (`float`, *optional*): + The final timestep value (overrides value given during scheduler instantiation). + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + + """ + sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps + + self.timesteps = torch.linspace(1, sampling_eps, num_inference_steps, device=device) + + def set_sigmas( + self, + num_inference_steps: int, + sigma_min: float | None = None, + sigma_max: float | None = None, + sampling_eps: float | None = None, + ): + """ + Sets the noise scales used for the diffusion chain (to be run before inference). The sigmas control the weight + of the `drift` and `diffusion` components of the sample update. + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + sigma_min (`float`, *optional*): + The initial noise scale value (overrides value given during scheduler instantiation). + sigma_max (`float`, *optional*): + The final noise scale value (overrides value given during scheduler instantiation). + sampling_eps (`float`, *optional*): + The final timestep value (overrides value given during scheduler instantiation). + + """ + sigma_min = sigma_min if sigma_min is not None else self.config.sigma_min + sigma_max = sigma_max if sigma_max is not None else self.config.sigma_max + sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps + if self.timesteps is None: + self.set_timesteps(num_inference_steps, sampling_eps) + + self.sigmas = sigma_min * (sigma_max / sigma_min) ** (self.timesteps / sampling_eps) + self.discrete_sigmas = torch.exp(torch.linspace(math.log(sigma_min), math.log(sigma_max), num_inference_steps)) + self.sigmas = torch.tensor([sigma_min * (sigma_max / sigma_min) ** t for t in self.timesteps]) + + def get_adjacent_sigma(self, timesteps: torch.Tensor, t: torch.Tensor) -> torch.Tensor: + """ + Returns the sigma value corresponding to the previous timestep index. When the timestep is zero (no previous + timestep), returns zero. + + Args: + timesteps (`torch.Tensor`): + The tensor of discrete timestep indices. + t (`torch.Tensor`): + The tensor of continuous timestep values. + + Returns: + `torch.Tensor`: + The sigma values for the adjacent (previous) timestep index. + """ + return torch.where( + timesteps == 0, + torch.zeros_like(t.to(timesteps.device)), + self.discrete_sigmas[timesteps - 1].to(timesteps.device), + ) + + def step_pred( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> SdeVeOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_sde_ve.SdeVeOutput`] is returned, otherwise a tuple + is returned where the first element is the sample tensor. + + """ + if self.timesteps is None: + raise ValueError( + "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + timestep = timestep * torch.ones( + sample.shape[0], device=sample.device + ) # torch.repeat_interleave(timestep, sample.shape[0]) + timesteps = (timestep * (len(self.timesteps) - 1)).long() + + # mps requires indices to be in the same device, so we use cpu as is the default with cuda + timesteps = timesteps.to(self.discrete_sigmas.device) + + sigma = self.discrete_sigmas[timesteps].to(sample.device) + adjacent_sigma = self.get_adjacent_sigma(timesteps, timestep).to(sample.device) + drift = torch.zeros_like(sample) + diffusion = (sigma**2 - adjacent_sigma**2) ** 0.5 + + # equation 6 in the paper: the model_output modeled by the network is grad_x log pt(x) + # also equation 47 shows the analog from SDE models to ancestral sampling methods + diffusion = diffusion.flatten() + while len(diffusion.shape) < len(sample.shape): + diffusion = diffusion.unsqueeze(-1) + drift = drift - diffusion**2 * model_output + + # equation 6: sample noise for the diffusion term of + noise = randn_tensor( + sample.shape, layout=sample.layout, generator=generator, device=sample.device, dtype=sample.dtype + ) + prev_sample_mean = sample - drift # subtract because `dt` is a small negative timestep + # TODO is the variable diffusion the correct scaling term for the noise? + prev_sample = prev_sample_mean + diffusion * noise # add impact of diffusion field g + + if not return_dict: + return (prev_sample, prev_sample_mean) + + return SdeVeOutput(prev_sample=prev_sample, prev_sample_mean=prev_sample_mean) + + def step_correct( + self, + model_output: torch.Tensor, + sample: torch.Tensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Correct the predicted sample based on the `model_output` of the network. This is often run repeatedly after + making the prediction for the previous timestep. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.timesteps is None: + raise ValueError( + "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler" + ) + + # For small batch sizes, the paper "suggest replacing norm(z) with sqrt(d), where d is the dim. of z" + # sample noise for correction + noise = randn_tensor(sample.shape, layout=sample.layout, generator=generator).to(sample.device) + + # compute step size from the model_output, the noise, and the snr + grad_norm = torch.norm(model_output.reshape(model_output.shape[0], -1), dim=-1).mean() + noise_norm = torch.norm(noise.reshape(noise.shape[0], -1), dim=-1).mean() + step_size = (self.config.snr * noise_norm / grad_norm) ** 2 * 2 + step_size = step_size * torch.ones(sample.shape[0]).to(sample.device) + # self.repeat_scalar(step_size, sample.shape[0]) + + # compute corrected sample: model_output term and noise term + step_size = step_size.flatten() + while len(step_size.shape) < len(sample.shape): + step_size = step_size.unsqueeze(-1) + prev_sample_mean = sample + step_size * model_output + prev_sample = prev_sample_mean + ((step_size * 2) ** 0.5) * noise + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.Tensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. This is the + forward diffusion process. + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise tensor to add to the original samples. + timesteps (`torch.Tensor`): + The timesteps at which to add noise, determining the noise level from the schedule. + + Returns: + `torch.Tensor`: + The noisy samples with added noise scaled according to the timestep schedule. + """ + timesteps = timesteps.to(original_samples.device) + sigmas = self.discrete_sigmas.to(original_samples.device)[timesteps] + noise = ( + noise * sigmas[:, None, None, None] + if noise is not None + else torch.randn_like(original_samples) * sigmas[:, None, None, None] + ) + noisy_samples = noise + original_samples + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_tcd.py b/diffusers/schedulers/scheduling_tcd.py new file mode 100644 index 0000000000000000000000000000000000000000..951a34b1f1cb0101d9d2c9d2db7ae6047aa0b61e --- /dev/null +++ b/diffusers/schedulers/scheduling_tcd.py @@ -0,0 +1,810 @@ +# Copyright 2025 Stanford University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion +# and https://github.com/hojonathanho/diffusion + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..schedulers.scheduling_utils import SchedulerMixin +from ..utils import BaseOutput, logging +from ..utils.torch_utils import randn_tensor + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class TCDSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_noised_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images, *optional*): + The predicted noised sample `(x_{s})` based on the model output from the current timestep. + """ + + prev_sample: torch.Tensor + pred_noised_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class TCDScheduler(SchedulerMixin, ConfigMixin): + """ + `TCDScheduler` incorporates the `Strategic Stochastic Sampling` introduced by the paper `Trajectory Consistency + Distillation`, extending the original Multistep Consistency Sampling to enable unrestricted trajectory traversal. + + This code is based on the official repo of TCD(https://github.com/jabir-zheng/TCD). + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config + attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be + accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving + functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions. + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to `0.00085`): + The starting `beta` value of inference. + beta_end (`float`, defaults to `0.012`): + The final `beta` value. + beta_schedule (`str`, defaults to `"scaled_linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray` or `list[float]`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + original_inference_steps (`int`, defaults to `50`): + The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we + will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule. + clip_sample (`bool`, defaults to `False`): + Clip the predicted sample for numerical stability. + clip_sample_range (`float`, defaults to `1.0`): + The maximum magnitude for sample clipping. Valid only when `clip_sample=True`. + set_alpha_to_one (`bool`, defaults to `True`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + steps_offset (`int`, defaults to `0`): + An offset added to the inference steps, as required by some model families. + prediction_type (`str`, defaults to `"epsilon"`): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper). + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to `0.995`): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to `1.0`): + The threshold value for dynamic thresholding. Valid only when `thresholding=True`. + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + timestep_scaling (`float`, defaults to `10.0`): + The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions + `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation + error at the default of `10.0` is already pretty small). + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.00085, + beta_end: float = 0.012, + beta_schedule: str = "scaled_linear", + trained_betas: np.ndarray | list[float] | None = None, + original_inference_steps: int = 50, + clip_sample: bool = False, + clip_sample_range: float = 1.0, + set_alpha_to_one: bool = True, + steps_offset: int = 0, + prediction_type: str = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + timestep_spacing: str = "leading", + timestep_scaling: float = 10.0, + rescale_betas_zero_snr: bool = False, + ) -> None: + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + + # Rescale for zero SNR + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + # At every step in ddim, we are looking into the previous alphas_cumprod + # For the final step, there is no previous alphas_cumprod because we are already at 0 + # `set_alpha_to_one` decides whether we set this parameter simply to one or + # whether we use the final alpha of the "non-previous" one. + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64)) + self.custom_timesteps = False + + self._step_index = None + self._begin_index = None + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep + def index_for_timestep( + self, timestep: float | torch.Tensor, schedule_timesteps: torch.Tensor | None = None + ) -> int: + """ + Find the index of a given timestep in the timestep schedule. + + Args: + timestep (`float` or `torch.Tensor`): + The timestep value to find in the schedule. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. For the very first step, returns the second index if + multiple matches exist to avoid skipping a sigma when starting mid-schedule (e.g., for image-to-image). + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + indices = (schedule_timesteps == timestep).nonzero() + + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + pos = 1 if len(indices) > 1 else 0 + + return indices[pos].item() + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index + def _init_step_index(self, timestep: float | torch.Tensor) -> None: + """ + Initialize the step index for the scheduler based on the given timestep. + + Args: + timestep (`float` or `torch.Tensor`): + The current timestep to initialize the step index from. + """ + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + @property + def step_index(self) -> int | None: + """ + The index counter for the current timestep. It increases by one after each scheduler step. + + Returns: + `int` or `None`: + The current step index, or `None` if it has not been initialized. + """ + return self._step_index + + @property + def begin_index(self) -> int | None: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + + Returns: + `int` or `None`: + The begin index for the scheduler, or `None` if it has not been set. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def scale_model_input(self, sample: torch.Tensor, timestep: int | torch.Tensor | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int` or `torch.Tensor`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler._get_variance + def _get_variance(self, timestep: int, prev_timestep: int) -> torch.Tensor: + """ + Computes the variance of the noise added at a given diffusion step. + + For a given `timestep` and its previous step, this method calculates the variance as defined in DDIM/DDPM + literature: + var_t = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + where alpha_prod and beta_prod are cumulative products of alphas and betas, respectively. + + Args: + timestep (`int`): + The current timestep in the diffusion process. + prev_timestep (`int`): + The previous timestep in the diffusion process. If negative, uses `final_alpha_cumprod`. + + Returns: + `torch.Tensor`: + The variance for the current timestep. + """ + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev) + + return variance + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device | None = None, + original_inference_steps: int | None = None, + timesteps: list[int] | None = None, + strength: float = 1.0, + ) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`, *optional*): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + original_inference_steps (`int`, *optional*): + The original number of inference steps, which will be used to generate a linearly-spaced timestep + schedule (which is different from the standard `diffusers` implementation). We will then take + `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as + our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute. + timesteps (`list[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep + schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`. + strength (`float`, defaults to `1.0`): + Used to determine the number of timesteps used for inference when using img2img, inpaint, etc. + """ + # 0. Check inputs + if num_inference_steps is None and timesteps is None: + raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.") + + if num_inference_steps is not None and timesteps is not None: + raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.") + + # 1. Calculate the TCD original training/distillation timestep schedule. + original_steps = ( + original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps + ) + + if original_inference_steps is None: + # default option, timesteps align with discrete inference steps + if original_steps > self.config.num_train_timesteps: + raise ValueError( + f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + # TCD Timesteps Setting + # The skipping step parameter k from the paper. + k = self.config.num_train_timesteps // original_steps + # TCD Training/Distillation Steps Schedule + tcd_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1 + else: + # customised option, sampled timesteps can be any arbitrary value + tcd_origin_timesteps = np.asarray(list(range(0, int(self.config.num_train_timesteps * strength)))) + + # 2. Calculate the TCD inference timestep schedule. + if timesteps is not None: + # 2.1 Handle custom timestep schedules. + train_timesteps = set(tcd_origin_timesteps) + non_train_timesteps = [] + for i in range(1, len(timesteps)): + if timesteps[i] >= timesteps[i - 1]: + raise ValueError("`custom_timesteps` must be in descending order.") + + if timesteps[i] not in train_timesteps: + non_train_timesteps.append(timesteps[i]) + + if timesteps[0] >= self.config.num_train_timesteps: + raise ValueError( + f"`timesteps` must start before `self.config.train_timesteps`: {self.config.num_train_timesteps}." + ) + + # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1 + if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1: + logger.warning( + f"The first timestep on the custom timestep schedule is {timesteps[0]}, not" + f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get" + f" unexpected results when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule + if non_train_timesteps: + logger.warning( + f"The custom timestep schedule contains the following timesteps which are not on the original" + f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results" + f" when using this timestep schedule." + ) + + # Raise warning if custom timestep schedule is longer than original_steps + if original_steps is not None: + if len(timesteps) > original_steps: + logger.warning( + f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the" + f" the length of the timestep schedule used for training: {original_steps}. You may get some" + f" unexpected results when using this timestep schedule." + ) + else: + if len(timesteps) > self.config.num_train_timesteps: + logger.warning( + f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the" + f" the length of the timestep schedule used for training: {self.config.num_train_timesteps}. You may get some" + f" unexpected results when using this timestep schedule." + ) + + timesteps = np.array(timesteps, dtype=np.int64) + self.num_inference_steps = len(timesteps) + self.custom_timesteps = True + + # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps) + init_timestep = min(int(self.num_inference_steps * strength), self.num_inference_steps) + t_start = max(self.num_inference_steps - init_timestep, 0) + timesteps = timesteps[t_start * self.order :] + # TODO: also reset self.num_inference_steps? + else: + # 2.2 Create the "standard" TCD inference timestep schedule. + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:" + f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle" + f" maximal {self.config.num_train_timesteps} timesteps." + ) + + if original_steps is not None: + skipping_step = len(tcd_origin_timesteps) // num_inference_steps + + if skipping_step < 1: + raise ValueError( + f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}." + ) + + self.num_inference_steps = num_inference_steps + + if original_steps is not None: + if num_inference_steps > original_steps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:" + f" {original_steps} because the final timestep schedule will be a subset of the" + f" `original_inference_steps`-sized initial timestep schedule." + ) + else: + if num_inference_steps > self.config.num_train_timesteps: + raise ValueError( + f"`num_inference_steps`: {num_inference_steps} cannot be larger than `num_train_timesteps`:" + f" {self.config.num_train_timesteps} because the final timestep schedule will be a subset of the" + f" `num_train_timesteps`-sized initial timestep schedule." + ) + + # TCD Inference Steps Schedule + tcd_origin_timesteps = tcd_origin_timesteps[::-1].copy() + # Select (approximately) evenly spaced indices from tcd_origin_timesteps. + inference_indices = np.linspace(0, len(tcd_origin_timesteps), num=num_inference_steps, endpoint=False) + inference_indices = np.floor(inference_indices).astype(np.int64) + timesteps = tcd_origin_timesteps[inference_indices] + + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long) + + self._step_index = None + self._begin_index = None + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + eta: float = 0.3, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> TCDSchedulerOutput | tuple[torch.Tensor, torch.Tensor]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + eta (`float`, defaults to `0.3`): + A stochastic parameter (referred to as `gamma` in the paper) used to control the stochasticity in every + step. When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic + sampling. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] or `tuple[torch.Tensor, torch.Tensor]`: + If return_dict is `True`, [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] is returned, otherwise a + tuple is returned containing the previous sample and the predicted noised sample. + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + assert 0 <= eta <= 1.0, "gamma must be less than or equal to 1.0" + + # 1. get previous step value + prev_step_index = self.step_index + 1 + if prev_step_index < len(self.timesteps): + prev_timestep = self.timesteps[prev_step_index] + else: + prev_timestep = torch.tensor(0) + + timestep_s = torch.floor((1 - eta) * prev_timestep).to(dtype=torch.long) + + # 2. compute alphas, betas + alpha_prod_t = self.alphas_cumprod[timestep] + beta_prod_t = 1 - alpha_prod_t + + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + + alpha_prod_s = self.alphas_cumprod[timestep_s] + beta_prod_s = 1 - alpha_prod_s + + # 3. Compute the predicted noised sample x_s based on the model parameterization + if self.config.prediction_type == "epsilon": # noise-prediction + pred_original_sample = (sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt() + pred_epsilon = model_output + pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon + elif self.config.prediction_type == "sample": # x-prediction + pred_original_sample = model_output + pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5) + pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon + elif self.config.prediction_type == "v_prediction": # v-prediction + pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output + pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or" + " `v_prediction` for `TCDScheduler`." + ) + + # 4. Sample and inject noise z ~ N(0, I) for MultiStep Inference + # Noise is not used on the final timestep of the timestep schedule. + # This also means that noise is not used for one-step sampling. + # Eta (referred to as "gamma" in the paper) was introduced to control the stochasticity in every step. + # When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic sampling. + if eta > 0: + if self.step_index != self.num_inference_steps - 1: + noise = randn_tensor( + model_output.shape, generator=generator, device=model_output.device, dtype=pred_noised_sample.dtype + ) + prev_sample = (alpha_prod_t_prev / alpha_prod_s).sqrt() * pred_noised_sample + ( + 1 - alpha_prod_t_prev / alpha_prod_s + ).sqrt() * noise + else: + prev_sample = pred_noised_sample + else: + prev_sample = pred_noised_sample + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample, pred_noised_sample) + + return TCDSchedulerOutput(prev_sample=prev_sample, pred_noised_sample=pred_noised_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity + def get_velocity(self, sample: torch.Tensor, noise: torch.Tensor, timesteps: torch.IntTensor) -> torch.Tensor: + """ + Compute the velocity prediction from the sample and noise according to the velocity formula. + + Args: + sample (`torch.Tensor`): + The input sample. + noise (`torch.Tensor`): + The noise tensor. + timesteps (`torch.IntTensor`): + The timesteps for velocity computation. + + Returns: + `torch.Tensor`: + The computed velocity. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as sample + self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype) + timesteps = timesteps.to(sample.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(sample.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample + return velocity + + def __len__(self) -> int: + """ + Return the number of training timesteps. + + Returns: + `int`: + The number of diffusion steps used to train the model. + """ + return self.config.num_train_timesteps + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep + def previous_timestep(self, timestep: int) -> int | torch.Tensor: + """ + Compute the previous timestep in the diffusion chain. + + Args: + timestep (`int`): + The current timestep. + + Returns: + `int` or `torch.Tensor`: + The previous timestep. + """ + if self.custom_timesteps or self.num_inference_steps: + index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0] + if index == self.timesteps.shape[0] - 1: + prev_t = torch.tensor(-1) + else: + prev_t = self.timesteps[index + 1] + else: + prev_t = timestep - 1 + return prev_t diff --git a/diffusers/schedulers/scheduling_unclip.py b/diffusers/schedulers/scheduling_unclip.py new file mode 100644 index 0000000000000000000000000000000000000000..747453eb2d405b086f2dc31da1ebe5d8edb4bdd8 --- /dev/null +++ b/diffusers/schedulers/scheduling_unclip.py @@ -0,0 +1,396 @@ +# Copyright 2025 Kakao Brain and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import math +from dataclasses import dataclass +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from ..utils.torch_utils import randn_tensor +from .scheduling_utils import SchedulerMixin + + +@dataclass +# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->UnCLIP +class UnCLIPSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's `step` function output. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + pred_original_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + The predicted denoised sample `(x_{0})` based on the model output from the current timestep. + `pred_original_sample` can be used to preview progress or for guidance. + """ + + prev_sample: torch.Tensor + pred_original_sample: torch.Tensor | None = None + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class UnCLIPScheduler(SchedulerMixin, ConfigMixin): + """ + NOTE: do not use this scheduler. The DDPM scheduler has been updated to support the changes made here. This + scheduler will be removed and replaced with DDPM. + + This is a modified DDPM Scheduler specifically for the karlo unCLIP model. + + This scheduler has some minor variations in how it calculates the learned range variance and dynamically + re-calculates betas based off the timesteps it is skipping. + + The scheduler also uses a slightly different step ratio when computing timesteps to use for inference. + + See [`~DDPMScheduler`] for more information on DDPM scheduling + + Args: + num_train_timesteps (`int`, defaults to `1000`): + The number of diffusion steps used to train the model. + variance_type (`"fixed_small_log"` or `"learned_range"`, defaults to `"fixed_small_log"`): + The variance type used when adding noise to the denoised sample. + clip_sample (`bool`, defaults to `True`): + Whether to clip the predicted sample between `-clip_sample_range` and `clip_sample_range` for numerical + stability. + clip_sample_range (`float`, defaults to `1.0`): + The range to clip the sample between. See `clip_sample`. + prediction_type (`"epsilon"` or `"sample"`, defaults to `"epsilon"`): + The prediction type of the scheduler function. Choose `epsilon` to predict the noise of the diffusion + process or `sample` to directly predict the noisy sample. + beta_schedule (`"squaredcos_cap_v2"`, defaults to `"squaredcos_cap_v2"`): + The beta schedule used to generate the beta values. Only `squaredcos_cap_v2` is supported. + """ + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + variance_type: Literal["fixed_small_log", "learned_range"] = "fixed_small_log", + clip_sample: bool = True, + clip_sample_range: float = 1.0, + prediction_type: Literal["epsilon", "sample"] = "epsilon", + beta_schedule: Literal["squaredcos_cap_v2"] = "squaredcos_cap_v2", + ) -> None: + if beta_schedule != "squaredcos_cap_v2": + raise ValueError("UnCLIPScheduler only supports `beta_schedule`: 'squaredcos_cap_v2'") + + self.betas = betas_for_alpha_bar(num_train_timesteps) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + self.one = torch.tensor(1.0) + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + self.variance_type = variance_type + + def scale_model_input(self, sample: torch.Tensor, timestep: int | None = None) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + timestep (`int`, *optional*): + The current timestep in the diffusion chain. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device | None = None) -> None: + """ + Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference. + + Note that this scheduler uses a slightly different step ratio than the other diffusers schedulers. The + different step ratio is to mimic the original karlo implementation and does not affect the quality or accuracy + of the results. + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps are moved. If `None`, the timesteps are not moved. + """ + self.num_inference_steps = num_inference_steps + step_ratio = (self.config.num_train_timesteps - 1) / (self.num_inference_steps - 1) + timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + def _get_variance( + self, + t: int, + prev_timestep: int | None = None, + predicted_variance: torch.Tensor | None = None, + variance_type: Literal["fixed_small_log", "learned_range"] | None = None, + ) -> torch.Tensor: + if prev_timestep is None: + prev_timestep = t - 1 + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if prev_timestep == t - 1: + beta = self.betas[t] + else: + beta = 1 - alpha_prod_t / alpha_prod_t_prev + + # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://huggingface.co/papers/2006.11239) + # and sample from it to get previous sample + # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample + variance = beta_prod_t_prev / beta_prod_t * beta + + if variance_type is None: + variance_type = self.config.variance_type + + # hacks - were probably added for training stability + if variance_type == "fixed_small_log": + variance = torch.log(torch.clamp(variance, min=1e-20)) + variance = torch.exp(0.5 * variance) + elif variance_type == "learned_range": + # NOTE difference with DDPM scheduler + min_log = variance.log() + max_log = beta.log() + + frac = (predicted_variance + 1) / 2 + variance = frac * max_log + (1 - frac) * min_log + + return variance + + def step( + self, + model_output: torch.Tensor, + timestep: int, + sample: torch.Tensor, + prev_timestep: int | None = None, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> UnCLIPSchedulerOutput | tuple: + """ + Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion + process from the learned model outputs (most often the predicted noise). + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + prev_timestep (`int`, *optional*): + The previous timestep for which to predict the previous sample. It is used to dynamically compute beta. + If `None`, `t-1` and the precomputed beta are used. + generator (`torch.Generator`, *optional*): + A random number generator. + return_dict (`bool`, defaults to `True`): + Whether to return an [`~schedulers.scheduling_unclip.UnCLIPSchedulerOutput`] or a `tuple`. + + Returns: + [`~schedulers.scheduling_unclip.UnCLIPSchedulerOutput`] or `tuple`: + If `return_dict` is `True`, an [`~schedulers.scheduling_unclip.UnCLIPSchedulerOutput`] is returned; + otherwise, a tuple is returned where the first element is the sample tensor. + """ + t = timestep + + if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type == "learned_range": + model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1) + else: + predicted_variance = None + + # 1. compute alphas, betas + if prev_timestep is None: + prev_timestep = t - 1 + + alpha_prod_t = self.alphas_cumprod[t] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.one + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if prev_timestep == t - 1: + beta = self.betas[t] + alpha = self.alphas[t] + else: + beta = 1 - alpha_prod_t / alpha_prod_t_prev + alpha = 1 - beta + + # 2. compute predicted original sample from predicted noise also called + # "predicted x_0" of formula (15) from https://huggingface.co/papers/2006.11239 + if self.config.prediction_type == "epsilon": + pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) + elif self.config.prediction_type == "sample": + pred_original_sample = model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `sample`" + " for the UnCLIPScheduler." + ) + + # 3. Clip "predicted x_0" + if self.config.clip_sample: + pred_original_sample = torch.clamp( + pred_original_sample, -self.config.clip_sample_range, self.config.clip_sample_range + ) + + # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * beta) / beta_prod_t + current_sample_coeff = alpha ** (0.5) * beta_prod_t_prev / beta_prod_t + + # 5. Compute predicted previous sample µ_t + # See formula (7) from https://huggingface.co/papers/2006.11239 + pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample + + # 6. Add noise + variance = 0 + if t > 0: + variance_noise = randn_tensor( + model_output.shape, dtype=model_output.dtype, generator=generator, device=model_output.device + ) + + variance = self._get_variance( + t, + predicted_variance=predicted_variance, + prev_timestep=prev_timestep, + ) + + if self.variance_type == "fixed_small_log": + variance = variance + elif self.variance_type == "learned_range": + variance = (0.5 * variance).exp() + else: + raise ValueError( + f"variance_type given as {self.variance_type} must be one of `fixed_small_log` or `learned_range`" + " for the UnCLIPScheduler." + ) + + variance = variance * variance_noise + + pred_prev_sample = pred_prev_sample + variance + + if not return_dict: + return ( + pred_prev_sample, + pred_original_sample, + ) + + return UnCLIPSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise magnitude at each timestep (this is the forward + diffusion process). + + Args: + original_samples (`torch.Tensor`): + The original samples to which noise will be added. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps indicating the noise level for each sample. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples diff --git a/diffusers/schedulers/scheduling_unipc_multistep.py b/diffusers/schedulers/scheduling_unipc_multistep.py new file mode 100644 index 0000000000000000000000000000000000000000..5c2cbcc13ff1139fa49b88cfde739013153e93b7 --- /dev/null +++ b/diffusers/schedulers/scheduling_unipc_multistep.py @@ -0,0 +1,1300 @@ +# Copyright 2025 TSAIL Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: check https://huggingface.co/papers/2302.04867 and https://github.com/wl-zhao/UniPC for more info +# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py + +import math +from typing import Literal + +import numpy as np +import torch + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import deprecate, is_scipy_available +from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +if is_scipy_available(): + import scipy.stats + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps: int, + max_beta: float = 0.999, + alpha_transform_type: Literal["cosine", "exp", "laplace"] = "cosine", +) -> torch.Tensor: + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + Args: + num_diffusion_timesteps (`int`): + The number of betas to produce. + max_beta (`float`, defaults to `0.999`): + The maximum beta to use; use values lower than 1 to avoid numerical instability. + alpha_transform_type (`str`, defaults to `"cosine"`): + The type of noise schedule for `alpha_bar`. Choose from `cosine`, `exp`, or `laplace`. + + Returns: + `torch.Tensor`: + The betas used by the scheduler to step the model outputs. + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "laplace": + + def alpha_bar_fn(t): + lmb = -0.5 * math.copysign(1, 0.5 - t) * math.log(1 - 2 * math.fabs(0.5 - t) + 1e-6) + snr = math.exp(lmb) + return math.sqrt(snr / (1 + snr)) + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_transform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr +def rescale_zero_terminal_snr(betas: torch.Tensor) -> torch.Tensor: + """ + Rescales betas to have zero terminal SNR Based on https://huggingface.co/papers/2305.08891 (Algorithm 1) + + Args: + betas (`torch.Tensor`): + The betas that the scheduler is being initialized with. + + Returns: + `torch.Tensor`: + Rescaled betas with zero terminal SNR. + """ + # Convert betas to alphas_bar_sqrt + alphas = 1.0 - betas + alphas_cumprod = torch.cumprod(alphas, dim=0) + alphas_bar_sqrt = alphas_cumprod.sqrt() + + # Store old values. + alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone() + alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone() + + # Shift so the last timestep is zero. + alphas_bar_sqrt -= alphas_bar_sqrt_T + + # Scale so the first timestep is back to the old value. + alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T) + + # Convert alphas_bar_sqrt to betas + alphas_bar = alphas_bar_sqrt**2 # Revert sqrt + alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod + alphas = torch.cat([alphas_bar[0:1], alphas]) + betas = 1 - alphas + + return betas + + +class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin): + """ + `UniPCMultistepScheduler` is a training-free framework designed for the fast sampling of diffusion models. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`"linear"`, `"scaled_linear"`, or `"squaredcos_cap_v2"`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + solver_order (`int`, defaults to `2`): + The UniPC order which can be any positive integer. The effective order of accuracy is `solver_order + 1` + due to the UniC. It is recommended to use `solver_order=2` for guided sampling, and `solver_order=3` for + unconditional sampling. + prediction_type (`"epsilon"`, `"sample"`, `"v_prediction"`, or `"flow_prediction"`, defaults to `"epsilon"`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process), + `sample` (directly predicts the noisy sample`), `v_prediction` (see section 2.4 of [Imagen + Video](https://huggingface.co/papers/2210.02303) paper), or `flow_prediction`. + thresholding (`bool`, defaults to `False`): + Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such + as Stable Diffusion. + dynamic_thresholding_ratio (`float`, defaults to 0.995): + The ratio for the dynamic thresholding method. Valid only when `thresholding=True`. + sample_max_value (`float`, defaults to 1.0): + The threshold value for dynamic thresholding. Valid only when `thresholding=True` and `predict_x0=True`. + predict_x0 (`bool`, defaults to `True`): + Whether to use the updating algorithm on the predicted x0. + solver_type (`"bh1"` or `"bh2"`, defaults to `"bh2"`): + Solver type for UniPC. It is recommended to use `bh1` for unconditional sampling when steps < 10, and `bh2` + otherwise. + lower_order_final (`bool`, default `True`): + Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can + stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10. + disable_corrector (`list`, default `[]`): + Decides which step to disable the corrector to mitigate the misalignment between `epsilon_theta(x_t, c)` + and `epsilon_theta(x_t^c, c)` which can influence convergence for a large guidance scale. Corrector is + usually disabled during the first few steps. + solver_p (`SchedulerMixin`, default `None`): + Any other scheduler that if specified, the algorithm becomes `solver_p + UniC`. + use_karras_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`, + the sigmas are determined according to a sequence of noise levels {σi}. + use_exponential_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use exponential sigmas for step sizes in the noise schedule during the sampling process. + use_beta_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use beta sigmas for step sizes in the noise schedule during the sampling process. Refer to [Beta + Sampling is All You Need](https://huggingface.co/papers/2407.12173) for more information. + use_flow_sigmas (`bool`, *optional*, defaults to `False`): + Whether to use flow sigmas for step sizes in the noise schedule during the sampling process. + timestep_spacing (`"linspace"`, `"leading"`, or `"trailing"`, defaults to `"linspace"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps, as required by some model families. + final_sigmas_type (`"zero"` or `"sigma_min"`, defaults to `"zero"`): + The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final + sigma is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0. + rescale_betas_zero_snr (`bool`, defaults to `False`): + Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and + dark samples instead of limiting it to samples with medium brightness. Loosely related to + [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506). + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: np.ndarray | list[float] | None = None, + solver_order: int = 2, + prediction_type: Literal["epsilon", "sample", "v_prediction", "flow_prediction"] = "epsilon", + thresholding: bool = False, + dynamic_thresholding_ratio: float = 0.995, + sample_max_value: float = 1.0, + predict_x0: bool = True, + solver_type: Literal["bh1", "bh2"] = "bh2", + lower_order_final: bool = True, + disable_corrector: list[int] = [], + solver_p: SchedulerMixin = None, + use_karras_sigmas: bool = False, + use_exponential_sigmas: bool = False, + use_beta_sigmas: bool = False, + use_flow_sigmas: bool = False, + flow_shift: float = 1.0, + timestep_spacing: Literal["linspace", "leading", "trailing"] = "linspace", + steps_offset: int = 0, + final_sigmas_type: Literal["zero", "sigma_min"] = "zero", + rescale_betas_zero_snr: bool = False, + use_dynamic_shifting: bool = False, + time_shift_type: Literal["exponential"] = "exponential", + sigma_min: float | None = None, + sigma_max: bool | None = None, + shift_terminal: bool | None = None, + ) -> None: + if self.config.use_beta_sigmas and not is_scipy_available(): + raise ImportError("Make sure to install scipy if you want to use beta sigmas.") + if sum([self.config.use_beta_sigmas, self.config.use_exponential_sigmas, self.config.use_karras_sigmas]) > 1: + raise ValueError( + "Only one of `config.use_beta_sigmas`, `config.use_exponential_sigmas`, `config.use_karras_sigmas` can be used." + ) + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} is not implemented for {self.__class__}") + if shift_terminal is not None and not use_flow_sigmas: + raise ValueError("`shift_terminal` is only supported when `use_flow_sigmas=True`.") + + if rescale_betas_zero_snr: + self.betas = rescale_zero_terminal_snr(self.betas) + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + if rescale_betas_zero_snr: + # Close to 0 without being 0 so first sigma is not inf + # FP16 smallest positive subnormal works well here + self.alphas_cumprod[-1] = 2**-24 + + # Currently we only support VP-type noise schedule + self.alpha_t = torch.sqrt(self.alphas_cumprod) + self.sigma_t = torch.sqrt(1 - self.alphas_cumprod) + self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t) + self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5 + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + if solver_type not in ["bh1", "bh2"]: + if solver_type in ["midpoint", "heun", "logrho"]: + self.register_to_config(solver_type="bh2") + else: + raise NotImplementedError(f"{solver_type} is not implemented for {self.__class__}") + + self.predict_x0 = predict_x0 + # setable values + self.num_inference_steps = None + timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps) + self.model_outputs = [None] * solver_order + self.timestep_list = [None] * solver_order + self.lower_order_nums = 0 + self.disable_corrector = disable_corrector + self.solver_p = solver_p + self.last_sample = None + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + @property + def step_index(self) -> int: + """ + The index counter for current timestep. It will increase 1 after each scheduler step. + """ + return self._step_index + + @property + def begin_index(self) -> int: + """ + The index for the first timestep. It should be set from pipeline with `set_begin_index` method. + """ + return self._begin_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index + def set_begin_index(self, begin_index: int = 0) -> None: + """ + Sets the begin index for the scheduler. This function should be run from pipeline before the inference. + + Args: + begin_index (`int`, defaults to `0`): + The begin index for the scheduler. + """ + self._begin_index = begin_index + + def set_timesteps( + self, + num_inference_steps: int | None = None, + device: str | torch.device = None, + sigmas: list[float] | None = None, + mu: bool | None = None, + ): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + sigmas (`List[float]`, *optional*): + Custom values for sigmas to be used for each diffusion step. If `None`, the sigmas are computed + automatically. + mu (`float`, *optional*): + Optional mu parameter for dynamic shifting when using exponential time shift type. + """ + if self.config.use_dynamic_shifting and mu is None: + raise ValueError("`mu` must be passed when `use_dynamic_shifting` is set to be `True`") + + if sigmas is not None: + if not self.config.use_flow_sigmas: + raise ValueError( + "Passing `sigmas` is only supported when `use_flow_sigmas=True`. " + "Please set `use_flow_sigmas=True` during scheduler initialization." + ) + num_inference_steps = len(sigmas) + + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://huggingface.co/papers/2305.08891 + if self.config.timestep_spacing == "linspace": + timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1) + .round()[::-1][:-1] + .copy() + .astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1) + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64) + timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64) + timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + if self.config.use_karras_sigmas: + if sigmas is None: + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + if self.config.use_flow_sigmas: + sigmas = sigmas / (sigmas + 1) + timesteps = (sigmas * self.config.num_train_timesteps).copy() + else: + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round() + + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + elif self.config.use_exponential_sigmas: + if sigmas is None: + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_exponential(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + elif self.config.use_beta_sigmas: + if sigmas is None: + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + log_sigmas = np.log(sigmas) + sigmas = np.flip(sigmas).copy() + sigmas = self._convert_to_beta(in_sigmas=sigmas, num_inference_steps=num_inference_steps) + timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]) + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + elif self.config.use_flow_sigmas: + if sigmas is None: + sigmas = np.linspace(1, 1 / self.config.num_train_timesteps, num_inference_steps + 1)[:-1] + if self.config.use_dynamic_shifting: + sigmas = self.time_shift(mu, 1.0, sigmas) + else: + sigmas = self.config.flow_shift * sigmas / (1 + (self.config.flow_shift - 1) * sigmas) + if self.config.shift_terminal: + sigmas = self.stretch_shift_to_terminal(sigmas) + eps = 1e-6 + if np.fabs(sigmas[0] - 1) < eps: + # to avoid inf torch.log(alpha_si) in multistep_uni_p_bh_update during first/second update + sigmas[0] -= eps + timesteps = (sigmas * self.config.num_train_timesteps).copy() + if self.config.final_sigmas_type == "sigma_min": + sigma_last = sigmas[-1] + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + else: + if sigmas is None: + sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5) + sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas) + if self.config.final_sigmas_type == "sigma_min": + sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5 + elif self.config.final_sigmas_type == "zero": + sigma_last = 0 + else: + raise ValueError( + f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}" + ) + sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32) + + self.sigmas = torch.from_numpy(sigmas) + self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64) + + self.num_inference_steps = len(timesteps) + + self.model_outputs = [ + None, + ] * self.config.solver_order + self.lower_order_nums = 0 + self.last_sample = None + if self.solver_p: + self.solver_p.set_timesteps(self.num_inference_steps, device=device) + + # add an index counter for schedulers that allow duplicated timesteps + self._step_index = None + self._begin_index = None + self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler.time_shift + def time_shift(self, mu: float, sigma: float, t: torch.Tensor): + """ + Apply time shifting to the sigmas. + + Args: + mu (`float`): + The mu parameter for the time shift. + sigma (`float`): + The sigma parameter for the time shift. + t (`torch.Tensor`): + The input timesteps. + + Returns: + `torch.Tensor`: + The time-shifted timesteps. + """ + if self.config.time_shift_type == "exponential": + return self._time_shift_exponential(mu, sigma, t) + elif self.config.time_shift_type == "linear": + return self._time_shift_linear(mu, sigma, t) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler.stretch_shift_to_terminal + def stretch_shift_to_terminal(self, t: torch.Tensor) -> torch.Tensor: + r""" + Stretches and shifts the timestep schedule to ensure it terminates at the configured `shift_terminal` config + value. + + Reference: + https://github.com/Lightricks/LTX-Video/blob/a01a171f8fe3d99dce2728d60a73fecf4d4238ae/ltx_video/schedulers/rf.py#L51 + + Args: + t (`torch.Tensor`): + A tensor of timesteps to be stretched and shifted. + + Returns: + `torch.Tensor`: + A tensor of adjusted timesteps such that the final value equals `self.config.shift_terminal`. + """ + one_minus_z = 1 - t + scale_factor = one_minus_z[-1] / (1 - self.config.shift_terminal) + stretched_t = 1 - (one_minus_z / scale_factor) + return stretched_t + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler._time_shift_exponential + def _time_shift_exponential(self, mu, sigma, t): + return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma) + + # Copied from diffusers.schedulers.scheduling_flow_match_euler_discrete.FlowMatchEulerDiscreteScheduler._time_shift_linear + def _time_shift_linear(self, mu, sigma, t): + return mu / (mu + (1 / t - 1) ** sigma) + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample + def _threshold_sample(self, sample: torch.Tensor) -> torch.Tensor: + """ + Apply dynamic thresholding to the predicted sample. + + "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the + prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by + s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing + pixels from saturation at each step. We find that dynamic thresholding results in significantly better + photorealism as well as better image-text alignment, especially when using very large guidance weights." + + https://huggingface.co/papers/2205.11487 + + Args: + sample (`torch.Tensor`): + The predicted sample to be thresholded. + + Returns: + `torch.Tensor`: + The thresholded sample. + """ + dtype = sample.dtype + batch_size, channels, *remaining_dims = sample.shape + + if dtype not in (torch.float32, torch.float64): + sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half + + # Flatten sample for doing quantile calculation along each image + sample = sample.reshape(batch_size, channels * np.prod(remaining_dims)) + + abs_sample = sample.abs() # "a certain percentile absolute pixel value" + + s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1) + s = torch.clamp( + s, min=1, max=self.config.sample_max_value + ) # When clamped to min=1, equivalent to standard clipping to [-1, 1] + s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0 + sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s" + + sample = sample.reshape(batch_size, channels, *remaining_dims) + sample = sample.to(dtype) + + return sample + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t + def _sigma_to_t(self, sigma: np.ndarray, log_sigmas: np.ndarray) -> np.ndarray: + """ + Convert sigma values to corresponding timestep values through interpolation. + + Args: + sigma (`np.ndarray`): + The sigma value(s) to convert to timestep(s). + log_sigmas (`np.ndarray`): + The logarithm of the sigma schedule used for interpolation. + + Returns: + `np.ndarray`: + The interpolated timestep value(s) corresponding to the input sigma(s). + """ + # get log sigma + log_sigma = np.log(np.maximum(sigma, 1e-10)) + + # get distribution + dists = log_sigma - log_sigmas[:, np.newaxis] + + # get sigmas range + low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2) + high_idx = low_idx + 1 + + low = log_sigmas[low_idx] + high = log_sigmas[high_idx] + + # interpolate sigmas + w = (low - log_sigma) / (low - high) + w = np.clip(w, 0, 1) + + # transform interpolation to time range + t = (1 - w) * low_idx + w * high_idx + t = t.reshape(sigma.shape) + return t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t + def _sigma_to_alpha_sigma_t(self, sigma: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: + """ + Convert sigma values to alpha_t and sigma_t values. + + Args: + sigma (`torch.Tensor`): + The sigma value(s) to convert. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: + A tuple containing (alpha_t, sigma_t) values. + """ + if self.config.use_flow_sigmas: + alpha_t = 1 - sigma + sigma_t = sigma + else: + alpha_t = 1 / ((sigma**2 + 1) ** 0.5) + sigma_t = sigma * alpha_t + + return alpha_t, sigma_t + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras + def _convert_to_karras(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct the noise schedule as proposed in [Elucidating the Design Space of Diffusion-Based Generative + Models](https://huggingface.co/papers/2206.00364). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following the Karras noise schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + rho = 7.0 # 7.0 is the value used in the paper + ramp = np.linspace(0, 1, num_inference_steps) + min_inv_rho = sigma_min ** (1 / rho) + max_inv_rho = sigma_max ** (1 / rho) + sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_exponential + def _convert_to_exponential(self, in_sigmas: torch.Tensor, num_inference_steps: int) -> torch.Tensor: + """ + Construct an exponential noise schedule. + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + + Returns: + `torch.Tensor`: + The converted sigma values following an exponential schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.exp(np.linspace(math.log(sigma_max), math.log(sigma_min), num_inference_steps)) + return sigmas + + # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_beta + def _convert_to_beta( + self, in_sigmas: torch.Tensor, num_inference_steps: int, alpha: float = 0.6, beta: float = 0.6 + ) -> torch.Tensor: + """ + Construct a beta noise schedule as proposed in [Beta Sampling is All You + Need](https://huggingface.co/papers/2407.12173). + + Args: + in_sigmas (`torch.Tensor`): + The input sigma values to be converted. + num_inference_steps (`int`): + The number of inference steps to generate the noise schedule for. + alpha (`float`, *optional*, defaults to `0.6`): + The alpha parameter for the beta distribution. + beta (`float`, *optional*, defaults to `0.6`): + The beta parameter for the beta distribution. + + Returns: + `torch.Tensor`: + The converted sigma values following a beta distribution schedule. + """ + + # Hack to make sure that other schedulers which copy this function don't break + # TODO: Add this logic to the other schedulers + if hasattr(self.config, "sigma_min"): + sigma_min = self.config.sigma_min + else: + sigma_min = None + + if hasattr(self.config, "sigma_max"): + sigma_max = self.config.sigma_max + else: + sigma_max = None + + sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item() + sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item() + + sigmas = np.array( + [ + sigma_min + (ppf * (sigma_max - sigma_min)) + for ppf in [ + scipy.stats.beta.ppf(timestep, alpha, beta) + for timestep in 1 - np.linspace(0, 1, num_inference_steps) + ] + ] + ) + return sigmas + + def convert_model_output( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + **kwargs, + ) -> torch.Tensor: + r""" + Convert the model output to the corresponding type the UniPC algorithm needs. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + + Returns: + `torch.Tensor`: + The converted model output. + """ + timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if timestep is not None: + deprecate( + "timesteps", + "1.0.0", + "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + sigma = self.sigmas[self.step_index] + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + + if self.predict_x0: + if self.config.prediction_type == "epsilon": + x0_pred = (sample - sigma_t * model_output) / alpha_t + elif self.config.prediction_type == "sample": + x0_pred = model_output + elif self.config.prediction_type == "v_prediction": + x0_pred = alpha_t * sample - sigma_t * model_output + elif self.config.prediction_type == "flow_prediction": + sigma_t = self.sigmas[self.step_index] + x0_pred = sample - sigma_t * model_output + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, " + "`v_prediction`, or `flow_prediction` for the UniPCMultistepScheduler." + ) + + if self.config.thresholding: + x0_pred = self._threshold_sample(x0_pred) + + return x0_pred + else: + if self.config.prediction_type == "epsilon": + return model_output + elif self.config.prediction_type == "sample": + epsilon = (sample - alpha_t * model_output) / sigma_t + return epsilon + elif self.config.prediction_type == "v_prediction": + epsilon = alpha_t * model_output + sigma_t * sample + return epsilon + else: + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or" + " `v_prediction` for the UniPCMultistepScheduler." + ) + + def multistep_uni_p_bh_update( + self, + model_output: torch.Tensor, + *args, + sample: torch.Tensor = None, + order: int = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the UniP (B(h) version). Alternatively, `self.solver_p` is used if is specified. + + Args: + model_output (`torch.Tensor`): + The direct output from the learned diffusion model at the current timestep. + prev_timestep (`int`): + The previous discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + order (`int`): + The order of UniP at this timestep (corresponds to the *p* in UniPC-p). + + Returns: + `torch.Tensor`: + The sample tensor at the previous timestep. + """ + prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None) + if sample is None: + if len(args) > 1: + sample = args[1] + else: + raise ValueError("missing `sample` as a required keyword argument") + if order is None: + if len(args) > 2: + order = args[2] + else: + raise ValueError("missing `order` as a required keyword argument") + if prev_timestep is not None: + deprecate( + "prev_timestep", + "1.0.0", + "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + model_output_list = self.model_outputs + + s0 = self.timestep_list[-1] + m0 = model_output_list[-1] + x = sample + + if self.solver_p: + x_t = self.solver_p.step(model_output, s0, x).prev_sample + return x_t + + device = sample.device + sigma_t, sigma_s0 = self.sigmas[self.step_index + 1].to(device), self.sigmas[self.step_index].to(device) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + h = lambda_t - lambda_s0 + + rks = [] + D1s = [] + for i in range(1, order): + si = self.step_index - i + mi = model_output_list[-(i + 1)] + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si].to(device)) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + rk = (lambda_si - lambda_s0) / h + rks.append(rk) + D1s.append((mi - m0) / rk) + + rks.append(torch.ones((), device=device)) + rks = torch.stack(rks) + + R = [] + b = [] + + hh = -h if self.predict_x0 else h + h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 + h_phi_k = h_phi_1 / hh - 1 + + factorial_i = 1 + + if self.config.solver_type == "bh1": + B_h = hh + elif self.config.solver_type == "bh2": + B_h = torch.expm1(hh) + else: + raise NotImplementedError() + + for i in range(1, order + 1): + R.append(torch.pow(rks, i - 1)) + b.append(h_phi_k * factorial_i / B_h) + factorial_i *= i + 1 + h_phi_k = h_phi_k / hh - 1 / factorial_i + + R = torch.stack(R) + b = torch.stack(b) if len(b) > 0 else torch.tensor(b, device=device) + + if len(D1s) > 0: + D1s = torch.stack(D1s, dim=1) # (B, K) + # for order 2, we use a simplified version + if order == 2: + rhos_p = torch.ones(1, dtype=x.dtype, device=device) * 0.5 + else: + rhos_p = torch.linalg.solve(R[:-1, :-1], b[:-1]).to(device).to(x.dtype) + else: + D1s = None + + if self.predict_x0: + x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0 + if D1s is not None: + pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s) + else: + pred_res = 0 + x_t = x_t_ - alpha_t * B_h * pred_res + else: + x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0 + if D1s is not None: + pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s) + else: + pred_res = 0 + x_t = x_t_ - sigma_t * B_h * pred_res + + x_t = x_t.to(x.dtype) + return x_t + + def multistep_uni_c_bh_update( + self, + this_model_output: torch.Tensor, + *args, + last_sample: torch.Tensor = None, + this_sample: torch.Tensor = None, + order: int = None, + **kwargs, + ) -> torch.Tensor: + """ + One step for the UniC (B(h) version). + + Args: + this_model_output (`torch.Tensor`): + The model outputs at `x_t`. + this_timestep (`int`): + The current timestep `t`. + last_sample (`torch.Tensor`): + The generated sample before the last predictor `x_{t-1}`. + this_sample (`torch.Tensor`): + The generated sample after the last predictor `x_{t}`. + order (`int`): + The `p` of UniC-p at this step. The effective order of accuracy should be `order + 1`. + + Returns: + `torch.Tensor`: + The corrected sample tensor at the current timestep. + """ + this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None) + if last_sample is None: + if len(args) > 1: + last_sample = args[1] + else: + raise ValueError("missing `last_sample` as a required keyword argument") + if this_sample is None: + if len(args) > 2: + this_sample = args[2] + else: + raise ValueError("missing `this_sample` as a required keyword argument") + if order is None: + if len(args) > 3: + order = args[3] + else: + raise ValueError("missing `order` as a required keyword argument") + if this_timestep is not None: + deprecate( + "this_timestep", + "1.0.0", + "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`", + ) + + model_output_list = self.model_outputs + + m0 = model_output_list[-1] + x = last_sample + x_t = this_sample + model_t = this_model_output + + device = this_sample.device + sigma_t, sigma_s0 = self.sigmas[self.step_index].to(device), self.sigmas[self.step_index - 1].to(device) + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t) + alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0) + + lambda_t = torch.log(alpha_t) - torch.log(sigma_t) + lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0) + + h = lambda_t - lambda_s0 + + rks = [] + D1s = [] + for i in range(1, order): + si = self.step_index - (i + 1) + mi = model_output_list[-(i + 1)] + alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si].to(device)) + lambda_si = torch.log(alpha_si) - torch.log(sigma_si) + rk = (lambda_si - lambda_s0) / h + rks.append(rk) + D1s.append((mi - m0) / rk) + + rks.append(torch.ones((), device=device)) + rks = torch.stack(rks) + + R = [] + b = [] + + hh = -h if self.predict_x0 else h + h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1 + h_phi_k = h_phi_1 / hh - 1 + + factorial_i = 1 + + if self.config.solver_type == "bh1": + B_h = hh + elif self.config.solver_type == "bh2": + B_h = torch.expm1(hh) + else: + raise NotImplementedError() + + for i in range(1, order + 1): + R.append(torch.pow(rks, i - 1)) + b.append(h_phi_k * factorial_i / B_h) + factorial_i *= i + 1 + h_phi_k = h_phi_k / hh - 1 / factorial_i + + R = torch.stack(R) + b = torch.stack(b) if len(b) > 0 else torch.tensor(b, device=device) + + if len(D1s) > 0: + D1s = torch.stack(D1s, dim=1) + else: + D1s = None + + # for order 1, we use a simplified version + if order == 1: + rhos_c = torch.ones(1, dtype=x.dtype, device=device) * 0.5 + else: + rhos_c = torch.linalg.solve(R, b).to(device).to(x.dtype) + + if self.predict_x0: + x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0 + if D1s is not None: + corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s) + else: + corr_res = 0 + D1_t = model_t - m0 + x_t = x_t_ - alpha_t * B_h * (corr_res + rhos_c[-1] * D1_t) + else: + x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0 + if D1s is not None: + corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s) + else: + corr_res = 0 + D1_t = model_t - m0 + x_t = x_t_ - sigma_t * B_h * (corr_res + rhos_c[-1] * D1_t) + x_t = x_t.to(x.dtype) + return x_t + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep + def index_for_timestep( + self, + timestep: int | torch.Tensor, + schedule_timesteps: torch.Tensor | None = None, + ) -> int: + """ + Find the index for a given timestep in the schedule. + + Args: + timestep (`int` or `torch.Tensor`): + The timestep for which to find the index. + schedule_timesteps (`torch.Tensor`, *optional*): + The timestep schedule to search in. If `None`, uses `self.timesteps`. + + Returns: + `int`: + The index of the timestep in the schedule. + """ + if schedule_timesteps is None: + schedule_timesteps = self.timesteps + + index_candidates = (schedule_timesteps == timestep).nonzero() + + if len(index_candidates) == 0: + step_index = len(self.timesteps) - 1 + # The sigma index that is taken for the **very** first `step` + # is always the second index (or the last index if there is only 1) + # This way we can ensure we don't accidentally skip a sigma in + # case we start in the middle of the denoising schedule (e.g. for image-to-image) + elif len(index_candidates) > 1: + step_index = index_candidates[1].item() + else: + step_index = index_candidates[0].item() + + return step_index + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index + def _init_step_index(self, timestep: int | torch.Tensor) -> None: + """ + Initialize the step_index counter for the scheduler. + + Args: + timestep (`int` or `torch.Tensor`): + The current timestep for which to initialize the step index. + """ + + if self.begin_index is None: + if isinstance(timestep, torch.Tensor): + timestep = timestep.to(self.timesteps.device) + self._step_index = self.index_for_timestep(timestep) + else: + self._step_index = self._begin_index + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.Tensor, + return_dict: bool = True, + ) -> SchedulerOutput | tuple: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the multistep UniPC. + + Args: + model_output (`torch.Tensor`): + The direct output from learned diffusion model. + timestep (`int` or `torch.Tensor`): + The current discrete timestep in the diffusion chain. + sample (`torch.Tensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if self.step_index is None: + self._init_step_index(timestep) + + use_corrector = ( + self.step_index > 0 and self.step_index - 1 not in self.disable_corrector and self.last_sample is not None + ) + + model_output_convert = self.convert_model_output(model_output, sample=sample) + if use_corrector: + sample = self.multistep_uni_c_bh_update( + this_model_output=model_output_convert, + last_sample=self.last_sample, + this_sample=sample, + order=self.this_order, + ) + + for i in range(self.config.solver_order - 1): + self.model_outputs[i] = self.model_outputs[i + 1] + self.timestep_list[i] = self.timestep_list[i + 1] + + self.model_outputs[-1] = model_output_convert + self.timestep_list[-1] = timestep + + if self.config.lower_order_final: + this_order = min(self.config.solver_order, len(self.timesteps) - self.step_index) + else: + this_order = self.config.solver_order + + self.this_order = min(this_order, self.lower_order_nums + 1) # warmup for multistep + assert self.this_order > 0 + + self.last_sample = sample + prev_sample = self.multistep_uni_p_bh_update( + model_output=model_output, # pass the original non-converted model output, in case solver-p is used + sample=sample, + order=self.this_order, + ) + + if self.lower_order_nums < self.config.solver_order: + self.lower_order_nums += 1 + + # upon completion increase step index by one + self._step_index += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.Tensor, *args, **kwargs) -> torch.Tensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.Tensor`): + The input sample. + + Returns: + `torch.Tensor`: + A scaled input sample. + """ + return sample + + # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise + def add_noise( + self, + original_samples: torch.Tensor, + noise: torch.Tensor, + timesteps: torch.IntTensor, + ) -> torch.Tensor: + """ + Add noise to the original samples according to the noise schedule at the specified timesteps. + + Args: + original_samples (`torch.Tensor`): + The original samples without noise. + noise (`torch.Tensor`): + The noise to add to the samples. + timesteps (`torch.IntTensor`): + The timesteps at which to add noise to the samples. + + Returns: + `torch.Tensor`: + The noisy samples. + """ + # Make sure sigmas and timesteps have the same device and dtype as original_samples + sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype) + if original_samples.device.type == "mps" and torch.is_floating_point(timesteps): + # mps does not support float64 + schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32) + timesteps = timesteps.to(original_samples.device, dtype=torch.float32) + else: + schedule_timesteps = self.timesteps.to(original_samples.device) + timesteps = timesteps.to(original_samples.device) + + # begin_index is None when the scheduler is used for training or pipeline does not implement set_begin_index + if self.begin_index is None: + step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps] + elif self.step_index is not None: + # add_noise is called after first denoising step (for inpainting) + step_indices = [self.step_index] * timesteps.shape[0] + else: + # add noise is called before first denoising step to create initial latent(img2img) + step_indices = [self.begin_index] * timesteps.shape[0] + + sigma = sigmas[step_indices].flatten() + while len(sigma.shape) < len(original_samples.shape): + sigma = sigma.unsqueeze(-1) + + alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma) + noisy_samples = alpha_t * original_samples + sigma_t * noise + return noisy_samples + + def __len__(self) -> int: + return self.config.num_train_timesteps diff --git a/diffusers/schedulers/scheduling_utils.py b/diffusers/schedulers/scheduling_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..8cdb21c2f0119907ee96867da08e4e9b78de93c9 --- /dev/null +++ b/diffusers/schedulers/scheduling_utils.py @@ -0,0 +1,199 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import importlib +import os +from dataclasses import dataclass +from enum import Enum +from typing import Any + +import torch +from huggingface_hub.utils import validate_hf_hub_args +from typing_extensions import Self + +from ..utils import BaseOutput, PushToHubMixin + + +SCHEDULER_CONFIG_NAME = "scheduler_config.json" + + +# NOTE: We make this type an enum because it simplifies usage in docs and prevents +# circular imports when used for `_compatibles` within the schedulers module. +# When it's used as a type in pipelines, it really is a Union because the actual +# scheduler instance is passed in. +class KarrasDiffusionSchedulers(Enum): + """ + Enumeration of schedulers compatible with Karras diffusion pipelines. + """ + + DDIMScheduler = 1 + DDPMScheduler = 2 + PNDMScheduler = 3 + LMSDiscreteScheduler = 4 + EulerDiscreteScheduler = 5 + HeunDiscreteScheduler = 6 + EulerAncestralDiscreteScheduler = 7 + DPMSolverMultistepScheduler = 8 + DPMSolverSinglestepScheduler = 9 + KDPM2DiscreteScheduler = 10 + KDPM2AncestralDiscreteScheduler = 11 + DEISMultistepScheduler = 12 + UniPCMultistepScheduler = 13 + DPMSolverSDEScheduler = 14 + EDMEulerScheduler = 15 + + +AysSchedules = { + "StableDiffusionTimesteps": [999, 850, 736, 645, 545, 455, 343, 233, 124, 24], + "StableDiffusionSigmas": [14.615, 6.475, 3.861, 2.697, 1.886, 1.396, 0.963, 0.652, 0.399, 0.152, 0.0], + "StableDiffusionXLTimesteps": [999, 845, 730, 587, 443, 310, 193, 116, 53, 13], + "StableDiffusionXLSigmas": [14.615, 6.315, 3.771, 2.181, 1.342, 0.862, 0.555, 0.380, 0.234, 0.113, 0.0], + "StableDiffusionVideoSigmas": [700.00, 54.5, 15.886, 7.977, 4.248, 1.789, 0.981, 0.403, 0.173, 0.034, 0.0], +} + + +@dataclass +class SchedulerOutput(BaseOutput): + """ + Base class for the output of a scheduler's `step` function. + + Args: + prev_sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)` for images): + Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the + denoising loop. + """ + + prev_sample: torch.Tensor + + +class SchedulerMixin(PushToHubMixin): + """ + Base class for all schedulers. + + [`SchedulerMixin`] contains common functions shared by all schedulers such as general loading and saving + functionalities. + + [`ConfigMixin`] takes care of storing the configuration attributes (like `num_train_timesteps`) that are passed to + the scheduler's `__init__` function, and the attributes can be accessed by `scheduler.config.num_train_timesteps`. + + Class attributes: + - **_compatibles** (`list[str]`) -- A list of scheduler classes that are compatible with the parent scheduler + class. Use [`~ConfigMixin.from_config`] to load a different compatible scheduler class (should be overridden + by parent class). + """ + + config_name = SCHEDULER_CONFIG_NAME + _compatibles = [] + has_compatibles = True + + @classmethod + @validate_hf_hub_args + def from_pretrained( + cls, + pretrained_model_name_or_path: str | os.PathLike | None = None, + subfolder: str | None = None, + return_unused_kwargs: bool = False, + **kwargs: Any, + ) -> Self | tuple[Self, dict[str, Any]]: + r""" + Instantiate a scheduler from a pre-defined JSON configuration file in a local directory or Hub repository. + + Args: + pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*): + Can be either: + + - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on + the Hub. + - A path to a *directory* (for example `./my_model_directory`) containing the scheduler + configuration saved with [`~SchedulerMixin.save_pretrained`]. + subfolder (`str`, *optional*): + The subfolder location of a model file within a larger model repository on the Hub or locally. + return_unused_kwargs (`bool`, *optional*, defaults to `False`): + Whether kwargs that are not consumed by the Python class should be returned or not. + cache_dir (`str` or `os.PathLike`, *optional*): + Path to a directory where a downloaded pretrained model configuration is cached if the standard cache + is not used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force the (re-)download of the model weights and configuration files, overriding the + cached versions if they exist. + proxies (`dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request. + output_loading_info (`bool`, *optional*, defaults to `False`): + Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages. + local_files_only (`bool`, *optional*, defaults to `False`): + Whether to only load local model weights and configuration files or not. If set to `True`, the model + won't be downloaded from the Hub. + token (`str` or `bool`, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from + `diffusers-cli login` (stored in `~/.huggingface`) is used. + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier + allowed by Git. + + Returns: + [`~schedulers.scheduling_utils.SchedulerMixin`] or `tuple[SchedulerMixin, dict[str, Any]]`: + The scheduler instantiated from the configuration. If `return_unused_kwargs` is `True`, a tuple is + returned where the second element is a dictionary of unused keyword arguments. + + > [!TIP] > To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in + with `hf > auth login`. You can also activate the special > + ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a > + firewalled environment. + + """ + config, kwargs, commit_hash = cls.load_config( + pretrained_model_name_or_path=pretrained_model_name_or_path, + subfolder=subfolder, + return_unused_kwargs=True, + return_commit_hash=True, + **kwargs, + ) + return cls.from_config(config, return_unused_kwargs=return_unused_kwargs, **kwargs) + + def save_pretrained(self, save_directory: str | os.PathLike, push_to_hub: bool = False, **kwargs: Any) -> None: + """ + Save a scheduler configuration object to a directory so that it can be reloaded using the + [`~SchedulerMixin.from_pretrained`] class method. + + Args: + save_directory (`str` or `os.PathLike`): + Directory where the configuration JSON file will be saved (will be created if it does not exist). + push_to_hub (`bool`, *optional*, defaults to `False`): + Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the + repository you want to push to with `repo_id` (will default to the name of `save_directory` in your + namespace). + kwargs (`dict[str, Any]`, *optional*): + Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method. + """ + self.save_config(save_directory=save_directory, push_to_hub=push_to_hub, **kwargs) + + @property + def compatibles(self) -> list[type[Self]]: + """ + Return all scheduler classes that are compatible with this scheduler. + + Returns: + `list[type[SchedulerMixin]]`: + A list of compatible scheduler classes. + """ + return self._get_compatibles() + + @classmethod + def _get_compatibles(cls) -> list[type[Self]]: + compatible_classes_str = list(set([cls.__name__] + cls._compatibles)) + diffusers_library = importlib.import_module(__name__.split(".")[0]) + compatible_classes = [ + getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c) + ] + return compatible_classes diff --git a/diffusers/schedulers/scheduling_vq_diffusion.py b/diffusers/schedulers/scheduling_vq_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..bb3f0fc6f697a52de34dc7ef6b20aab31e89a2de --- /dev/null +++ b/diffusers/schedulers/scheduling_vq_diffusion.py @@ -0,0 +1,516 @@ +# Copyright 2025 Microsoft and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from dataclasses import dataclass + +import numpy as np +import torch +import torch.nn.functional as F + +from ..configuration_utils import ConfigMixin, register_to_config +from ..utils import BaseOutput +from .scheduling_utils import SchedulerMixin + + +@dataclass +class VQDiffusionSchedulerOutput(BaseOutput): + """ + Output class for the scheduler's step function output. + + Args: + prev_sample (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The computed sample at the previous timestep. `prev_sample` should be used as the next model input in the + denoising loop. + """ + + prev_sample: torch.LongTensor + + +def index_to_log_onehot(x: torch.LongTensor, num_classes: int) -> torch.Tensor: + """ + Convert a batch of class index vectors into a batch of log one-hot vectors. + + Args: + x (`torch.LongTensor` of shape `(batch size, vector length)`): + The batch of class indices. + num_classes (`int`): + The number of classes used for the one-hot vectors. + + Returns: + `torch.Tensor` of shape `(batch size, num classes, vector length)`: + The log one-hot vectors. + """ + x_onehot = F.one_hot(x, num_classes) + x_onehot = x_onehot.permute(0, 2, 1) + log_x = torch.log(x_onehot.float().clamp(min=1e-30)) + return log_x + + +def gumbel_noised(logits: torch.Tensor, generator: torch.Generator | None) -> torch.Tensor: + """ + Apply Gumbel noise to the input logits. + + Args: + logits (`torch.Tensor`): + The logits to which Gumbel noise is added. + generator (`torch.Generator` or `None`): + A random number generator for reproducible noise generation, or `None` to use the default generator. + + Returns: + `torch.Tensor`: + The logits with Gumbel noise added. + """ + uniform = torch.rand(logits.shape, device=logits.device, generator=generator) + gumbel_noise = -torch.log(-torch.log(uniform + 1e-30) + 1e-30) + noised = gumbel_noise + logits + return noised + + +def alpha_schedules( + num_diffusion_timesteps: int, alpha_cum_start: float = 0.99999, alpha_cum_end: float = 0.000009 +) -> tuple[np.ndarray, np.ndarray]: + """ + Create cumulative and non-cumulative alpha schedules for the diffusion process. + + The schedules follow the formulation in section 4.1 of the VQ-Diffusion paper. + + Args: + num_diffusion_timesteps (`int`): + The number of diffusion timesteps. + alpha_cum_start (`float`, defaults to `0.99999`): + The starting cumulative alpha value. + alpha_cum_end (`float`, defaults to `0.000009`): + The ending cumulative alpha value. + + Returns: + `tuple[np.ndarray, np.ndarray]`: + A tuple containing the non-cumulative and cumulative alpha schedules, respectively. + """ + att = ( + np.arange(0, num_diffusion_timesteps) / (num_diffusion_timesteps - 1) * (alpha_cum_end - alpha_cum_start) + + alpha_cum_start + ) + att = np.concatenate(([1], att)) + at = att[1:] / att[:-1] + att = np.concatenate((att[1:], [1])) + return at, att + + +def gamma_schedules( + num_diffusion_timesteps: int, gamma_cum_start: float = 0.000009, gamma_cum_end: float = 0.99999 +) -> tuple[np.ndarray, np.ndarray]: + """ + Create cumulative and non-cumulative gamma schedules for the diffusion process. + + The schedules follow the formulation in section 4.1 of the VQ-Diffusion paper. + + Args: + num_diffusion_timesteps (`int`): + The number of diffusion timesteps. + gamma_cum_start (`float`, defaults to `0.000009`): + The starting cumulative gamma value. + gamma_cum_end (`float`, defaults to `0.99999`): + The ending cumulative gamma value. + + Returns: + `tuple[np.ndarray, np.ndarray]`: + A tuple containing the non-cumulative and cumulative gamma schedules, respectively. + """ + ctt = ( + np.arange(0, num_diffusion_timesteps) / (num_diffusion_timesteps - 1) * (gamma_cum_end - gamma_cum_start) + + gamma_cum_start + ) + ctt = np.concatenate(([0], ctt)) + one_minus_ctt = 1 - ctt + one_minus_ct = one_minus_ctt[1:] / one_minus_ctt[:-1] + ct = 1 - one_minus_ct + ctt = np.concatenate((ctt[1:], [0])) + return ct, ctt + + +class VQDiffusionScheduler(SchedulerMixin, ConfigMixin): + """ + A scheduler for vector quantized diffusion. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_vec_classes (`int`): + The number of classes of the vector embeddings of the latent pixels. Includes the class for the masked + latent pixel. + num_train_timesteps (`int`, defaults to `100`): + The number of diffusion steps to train the model. + alpha_cum_start (`float`, defaults to `0.99999`): + The starting cumulative alpha value. + alpha_cum_end (`float`, defaults to `0.000009`): + The ending cumulative alpha value. + gamma_cum_start (`float`, defaults to `0.000009`): + The starting cumulative gamma value. + gamma_cum_end (`float`, defaults to `0.99999`): + The ending cumulative gamma value. + """ + + order = 1 + + @register_to_config + def __init__( + self, + num_vec_classes: int, + num_train_timesteps: int = 100, + alpha_cum_start: float = 0.99999, + alpha_cum_end: float = 0.000009, + gamma_cum_start: float = 0.000009, + gamma_cum_end: float = 0.99999, + ) -> None: + self.num_embed = num_vec_classes + + # By convention, the index for the mask class is the last class index + self.mask_class = self.num_embed - 1 + + at, att = alpha_schedules(num_train_timesteps, alpha_cum_start=alpha_cum_start, alpha_cum_end=alpha_cum_end) + ct, ctt = gamma_schedules(num_train_timesteps, gamma_cum_start=gamma_cum_start, gamma_cum_end=gamma_cum_end) + + num_non_mask_classes = self.num_embed - 1 + bt = (1 - at - ct) / num_non_mask_classes + btt = (1 - att - ctt) / num_non_mask_classes + + at = torch.tensor(at.astype("float64")) + bt = torch.tensor(bt.astype("float64")) + ct = torch.tensor(ct.astype("float64")) + log_at = torch.log(at) + log_bt = torch.log(bt) + log_ct = torch.log(ct) + + att = torch.tensor(att.astype("float64")) + btt = torch.tensor(btt.astype("float64")) + ctt = torch.tensor(ctt.astype("float64")) + log_cumprod_at = torch.log(att) + log_cumprod_bt = torch.log(btt) + log_cumprod_ct = torch.log(ctt) + + self.log_at = log_at.float() + self.log_bt = log_bt.float() + self.log_ct = log_ct.float() + self.log_cumprod_at = log_cumprod_at.float() + self.log_cumprod_bt = log_cumprod_bt.float() + self.log_cumprod_ct = log_cumprod_ct.float() + + # setable values + self.num_inference_steps = None + self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy()) + + def set_timesteps(self, num_inference_steps: int, device: str | torch.device | None = None) -> None: + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps and diffusion process parameters (alpha, beta, gamma) should be moved + to. + """ + self.num_inference_steps = num_inference_steps + timesteps = np.arange(0, self.num_inference_steps)[::-1].copy() + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.log_at = self.log_at.to(device) + self.log_bt = self.log_bt.to(device) + self.log_ct = self.log_ct.to(device) + self.log_cumprod_at = self.log_cumprod_at.to(device) + self.log_cumprod_bt = self.log_cumprod_bt.to(device) + self.log_cumprod_ct = self.log_cumprod_ct.to(device) + + def step( + self, + model_output: torch.Tensor, + timestep: int | torch.Tensor, + sample: torch.LongTensor, + generator: torch.Generator | None = None, + return_dict: bool = True, + ) -> VQDiffusionSchedulerOutput | tuple[torch.LongTensor]: + """ + Predict the sample from the previous timestep by the reverse transition distribution. See + [`~VQDiffusionScheduler.q_posterior`] for more details about how the distribution is computed. + + Args: + model_output (`torch.Tensor` of shape `(batch size, num classes - 1, num latent pixels)`): + The log probabilities for the predicted classes of the initial latent pixels. Does not include a + prediction for the masked class as the initial unnoised image cannot be masked. + timestep (`int` or `torch.Tensor`): + The timestep that determines which transition matrices are used. + sample (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The classes of each latent pixel at the current timestep. + generator (`torch.Generator`, *optional*): + A random number generator for the noise applied to `p(x_{t-1} | x_t)` before it is sampled from. + return_dict (`bool`, defaults to `True`): + Whether or not to return a [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] or + `tuple`. + + Returns: + [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] or `tuple[torch.LongTensor]`: + If return_dict is `True`, [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] is + returned, otherwise a tuple is returned where the first element is the sample tensor. + """ + if timestep == 0: + log_p_x_t_min_1 = model_output + else: + log_p_x_t_min_1 = self.q_posterior(model_output, sample, timestep) + + log_p_x_t_min_1 = gumbel_noised(log_p_x_t_min_1, generator) + + x_t_min_1 = log_p_x_t_min_1.argmax(dim=1) + + if not return_dict: + return (x_t_min_1,) + + return VQDiffusionSchedulerOutput(prev_sample=x_t_min_1) + + def q_posterior(self, log_p_x_0: torch.Tensor, x_t: torch.LongTensor, t: int | torch.Tensor) -> torch.Tensor: + """ + Calculates the log probabilities for the predicted classes of the image at timestep `t-1`: + + ``` + p(x_{t-1} | x_t) = sum( q(x_t | x_{t-1}) * q(x_{t-1} | x_0) * p(x_0) / q(x_t | x_0) ) + ``` + + Args: + log_p_x_0 (`torch.Tensor` of shape `(batch size, num classes - 1, num latent pixels)`): + The log probabilities for the predicted classes of the initial latent pixels. Does not include a + prediction for the masked class as the initial unnoised image cannot be masked. + x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The classes of each latent pixel at time `t`. + t (`int` or `torch.Tensor`): + The timestep that determines which transition matrix is used. + + Returns: + `torch.Tensor` of shape `(batch size, num classes, num latent pixels)`: + The log probabilities for the predicted classes of the image at timestep `t-1`. + """ + log_onehot_x_t = index_to_log_onehot(x_t, self.num_embed) + + log_q_x_t_given_x_0 = self.log_Q_t_transitioning_to_known_class( + t=t, x_t=x_t, log_onehot_x_t=log_onehot_x_t, cumulative=True + ) + + log_q_t_given_x_t_min_1 = self.log_Q_t_transitioning_to_known_class( + t=t, x_t=x_t, log_onehot_x_t=log_onehot_x_t, cumulative=False + ) + + # p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) ... p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + # . . . + # . . . + # . . . + # p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) ... p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) + q = log_p_x_0 - log_q_x_t_given_x_0 + + # sum_0 = p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + ... + p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}), ... , + # sum_n = p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + ... + p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) + q_log_sum_exp = torch.logsumexp(q, dim=1, keepdim=True) + + # p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0 ... p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n + # . . . + # . . . + # . . . + # p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0 ... p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n + q = q - q_log_sum_exp + + # (p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1} ... (p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1} + # . . . + # . . . + # . . . + # (p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1} ... (p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1} + # c_cumulative_{t-1} ... c_cumulative_{t-1} + q = self.apply_cumulative_transitions(q, t - 1) + + # ((p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_0) * sum_0 ... ((p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_0) * sum_n + # . . . + # . . . + # . . . + # ((p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_{k-1}) * sum_0 ... ((p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_{k-1}) * sum_n + # c_cumulative_{t-1} * q(x_t | x_{t-1}=C_k) * sum_0 ... c_cumulative_{t-1} * q(x_t | x_{t-1}=C_k) * sum_0 + log_p_x_t_min_1 = q + log_q_t_given_x_t_min_1 + q_log_sum_exp + + # For each column, there are two possible cases. + # + # Where: + # - sum(p_n(x_0))) is summing over all classes for x_0 + # - C_i is the class transitioning from (not to be confused with c_t and c_cumulative_t being used for gamma's) + # - C_j is the class transitioning to + # + # 1. x_t is masked i.e. x_t = c_k + # + # Simplifying the expression, the column vector is: + # . + # . + # . + # (c_t / c_cumulative_t) * (a_cumulative_{t-1} * p_n(x_0 = C_i | x_t) + b_cumulative_{t-1} * sum(p_n(x_0))) + # . + # . + # . + # (c_cumulative_{t-1} / c_cumulative_t) * sum(p_n(x_0)) + # + # From equation (11) stated in terms of forward probabilities, the last row is trivially verified. + # + # For the other rows, we can state the equation as ... + # + # (c_t / c_cumulative_t) * [b_cumulative_{t-1} * p(x_0=c_0) + ... + (a_cumulative_{t-1} + b_cumulative_{t-1}) * p(x_0=C_i) + ... + b_cumulative_{k-1} * p(x_0=c_{k-1})] + # + # This verifies the other rows. + # + # 2. x_t is not masked + # + # Simplifying the expression, there are two cases for the rows of the column vector, where C_j = C_i and where C_j != C_i: + # . + # . + # . + # C_j != C_i: b_t * ((b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_0) + ... + ((a_cumulative_{t-1} + b_cumulative_{t-1}) / b_cumulative_t) * p_n(x_0 = C_i) + ... + (b_cumulative_{t-1} / (a_cumulative_t + b_cumulative_t)) * p_n(c_0=C_j) + ... + (b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_{k-1})) + # . + # . + # . + # C_j = C_i: (a_t + b_t) * ((b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_0) + ... + ((a_cumulative_{t-1} + b_cumulative_{t-1}) / (a_cumulative_t + b_cumulative_t)) * p_n(x_0 = C_i = C_j) + ... + (b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_{k-1})) + # . + # . + # . + # 0 + # + # The last row is trivially verified. The other rows can be verified by directly expanding equation (11) stated in terms of forward probabilities. + return log_p_x_t_min_1 + + def log_Q_t_transitioning_to_known_class( + self, *, t: int | torch.Tensor, x_t: torch.LongTensor, log_onehot_x_t: torch.Tensor, cumulative: bool + ) -> torch.Tensor: + """ + Calculates the log probabilities of the rows from the (cumulative or non-cumulative) transition matrix for each + latent pixel in `x_t`. + + Args: + t (`int` or `torch.Tensor`): + The timestep that determines which transition matrix is used. + x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`): + The classes of each latent pixel at time `t`. + log_onehot_x_t (`torch.Tensor` of shape `(batch size, num classes, num latent pixels)`): + The log one-hot vectors of `x_t`. + cumulative (`bool`): + If cumulative is `False`, the single step transition matrix `t-1`->`t` is used. If cumulative is + `True`, the cumulative transition matrix `0`->`t` is used. + + Returns: + `torch.Tensor`: + Each _column_ of the returned matrix is a _row_ of log probabilities of the complete probability + transition matrix. The tensor has shape `(batch size, num classes - 1, num latent pixels)` when + `cumulative` is `True` and shape `(batch size, num classes, num latent pixels)` otherwise. + + When cumulative, the tensor has one fewer row because the initial latent pixel cannot be masked. + + Where: + - `q_n` is the probability distribution for the forward process of the `n`th latent pixel. + - C_0 is a class of a latent pixel embedding + - C_k is the class of the masked latent pixel + + non-cumulative result (omitting logarithms): + ``` + q_0(x_t | x_{t-1} = C_0) ... q_n(x_t | x_{t-1} = C_0) + . . . + . . . + . . . + q_0(x_t | x_{t-1} = C_k) ... q_n(x_t | x_{t-1} = C_k) + ``` + + cumulative result (omitting logarithms): + ``` + q_0_cumulative(x_t | x_0 = C_0) ... q_n_cumulative(x_t | x_0 = C_0) + . . . + . . . + . . . + q_0_cumulative(x_t | x_0 = C_{k-1}) ... q_n_cumulative(x_t | x_0 = C_{k-1}) + ``` + """ + if cumulative: + a = self.log_cumprod_at[t] + b = self.log_cumprod_bt[t] + c = self.log_cumprod_ct[t] + else: + a = self.log_at[t] + b = self.log_bt[t] + c = self.log_ct[t] + + if not cumulative: + # The values in the onehot vector can also be used as the logprobs for transitioning + # from masked latent pixels. If we are not calculating the cumulative transitions, + # we need to save these vectors to be re-appended to the final matrix so the values + # aren't overwritten. + # + # `P(x_t!=mask|x_{t-1=mask}) = 0` and 0 will be the value of the last row of the onehot vector + # if x_t is not masked + # + # `P(x_t=mask|x_{t-1=mask}) = 1` and 1 will be the value of the last row of the onehot vector + # if x_t is masked + log_onehot_x_t_transitioning_from_masked = log_onehot_x_t[:, -1, :].unsqueeze(1) + + # `index_to_log_onehot` will add onehot vectors for masked pixels, + # so the default one hot matrix has one too many rows. See the doc string + # for an explanation of the dimensionality of the returned matrix. + log_onehot_x_t = log_onehot_x_t[:, :-1, :] + + # this is a cheeky trick to produce the transition probabilities using log one-hot vectors. + # + # Don't worry about what values this sets in the columns that mark transitions + # to masked latent pixels. They are overwrote later with the `mask_class_mask`. + # + # Looking at the below logspace formula in non-logspace, each value will evaluate to either + # `1 * a + b = a + b` where `log_Q_t` has the one hot value in the column + # or + # `0 * a + b = b` where `log_Q_t` has the 0 values in the column. + # + # See equation 7 for more details. + log_Q_t = (log_onehot_x_t + a).logaddexp(b) + + # The whole column of each masked pixel is `c` + mask_class_mask = x_t == self.mask_class + mask_class_mask = mask_class_mask.unsqueeze(1).expand(-1, self.num_embed - 1, -1) + log_Q_t[mask_class_mask] = c + + if not cumulative: + log_Q_t = torch.cat((log_Q_t, log_onehot_x_t_transitioning_from_masked), dim=1) + + return log_Q_t + + def apply_cumulative_transitions(self, q: torch.Tensor, t: int | torch.Tensor) -> torch.Tensor: + """ + Apply the cumulative transition matrix at a timestep to log probabilities over non-mask classes. + + Args: + q (`torch.Tensor` of shape `(batch size, num classes - 1, num latent pixels)`): + The log probabilities over the non-mask latent pixel classes. + t (`int` or `torch.Tensor`): + The timestep that determines which cumulative transition matrix is used. + + Returns: + `torch.Tensor` of shape `(batch size, num classes, num latent pixels)`: + The transitioned log probabilities with the mask class probabilities appended. + """ + bsz = q.shape[0] + a = self.log_cumprod_at[t] + b = self.log_cumprod_bt[t] + c = self.log_cumprod_ct[t] + + num_latent_pixels = q.shape[2] + c = c.expand(bsz, 1, num_latent_pixels) + + q = (q + a).logaddexp(b) + q = torch.cat((q, c), dim=1) + + return q diff --git a/diffusers/training_utils.py b/diffusers/training_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..68c0c5254234dc012071a90c17e8418de9c3404b --- /dev/null +++ b/diffusers/training_utils.py @@ -0,0 +1,963 @@ +import contextlib +import copy +import gc +import math +import random +import re +import warnings +from contextlib import contextmanager +from functools import partial +from typing import Any, Iterable + +import numpy as np +import torch +import torch.nn.functional as F + + +if torch.distributed.is_available(): + from torch.distributed.fsdp import CPUOffload, ShardingStrategy + from torch.distributed.fsdp import FullyShardedDataParallel as FSDP + from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy +else: + CPUOffload = None + ShardingStrategy = None + FSDP = None + transformer_auto_wrap_policy = None + +from .models import UNet2DConditionModel +from .pipelines import DiffusionPipeline +from .schedulers import SchedulerMixin +from .utils import ( + convert_state_dict_to_diffusers, + convert_state_dict_to_peft, + deprecate, + is_accelerate_available, + is_peft_available, + is_torch_npu_available, + is_torchvision_available, + is_transformers_available, +) + + +if is_transformers_available(): + import transformers + + if transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): + import deepspeed + +if is_accelerate_available(): + from accelerate.logging import get_logger + +if is_peft_available(): + from peft import set_peft_model_state_dict + +if is_torchvision_available(): + from torchvision import transforms + +if is_torch_npu_available(): + import torch_npu # noqa: F401 + + +def set_seed(seed: int): + """ + Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`. + + Args: + seed (`int`): The seed to set. + + Returns: + `None` + """ + random.seed(seed) + np.random.seed(seed) + torch.manual_seed(seed) + if is_torch_npu_available(): + torch.npu.manual_seed_all(seed) + else: + torch.cuda.manual_seed_all(seed) + # ^^ safe to call this function even if cuda is not available + + +def compute_snr(noise_scheduler, timesteps): + """ + Computes SNR as per + https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849 + for the given timesteps using the provided noise scheduler. + + Args: + noise_scheduler (`NoiseScheduler`): + An object containing the noise schedule parameters, specifically `alphas_cumprod`, which is used to compute + the SNR values. + timesteps (`torch.Tensor`): + A tensor of timesteps for which the SNR is computed. + + Returns: + `torch.Tensor`: A tensor containing the computed SNR values for each timestep. + """ + alphas_cumprod = noise_scheduler.alphas_cumprod + sqrt_alphas_cumprod = alphas_cumprod**0.5 + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # Expand the tensors. + # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026 + sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None] + alpha = sqrt_alphas_cumprod.expand(timesteps.shape) + + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float() + while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape): + sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None] + sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape) + + # Compute SNR. + snr = (alpha / sigma) ** 2 + return snr + + +def compute_confidence_aware_loss( + logits: torch.Tensor, + labels: torch.Tensor, + *, + lambda_conf: float = 0.0, + temperature: float = 1.0, + per_token_weights: torch.Tensor | None = None, + ignore_index: int = -100, +) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Computes a confidence-aware training loss for token classification-style heads. + + This loss combines: + - `loss_sft`: standard supervised cross-entropy on all non-ignored labels. + - `loss_conf`: an entropy penalty applied only on tokens that are already predicted correctly. + + Args: + logits (`torch.Tensor`): Logits of shape `(..., vocab_size)`. + labels (`torch.Tensor`): Labels of shape `(...)`, matching `logits.shape[:-1]`. Values set to `ignore_index` + are excluded from both losses. + lambda_conf (`float`, *optional*, defaults to `0.0`): Weight for the confidence term. + temperature (`float`, *optional*, defaults to `1.0`): Temperature used for the entropy term only. Lower values + sharpen the distribution and change the strength of the confidence gradients. + per_token_weights (`torch.Tensor`, *optional*): Optional weights of shape `(...)` to reweight both losses per + token (e.g. schedule-aware weights). Tokens with weight `0` contribute nothing. + ignore_index (`int`, *optional*, defaults to `-100`): Ignore index for labels. + + Returns: + `Tuple[torch.Tensor, torch.Tensor, torch.Tensor]`: `(loss, loss_sft, loss_conf)`. + """ + if logits.ndim < 2: + raise ValueError(f"`logits` must have at least 2 dims, got shape {tuple(logits.shape)}.") + if labels.shape != logits.shape[:-1]: + raise ValueError( + f"`labels` shape must match `logits.shape[:-1]`, got labels={tuple(labels.shape)} logits={tuple(logits.shape)}." + ) + if temperature <= 0: + raise ValueError(f"`temperature` must be > 0, got {temperature}.") + + valid = labels.ne(ignore_index) + if per_token_weights is None: + weights = torch.ones_like(labels, dtype=logits.dtype) + else: + if per_token_weights.shape != labels.shape: + raise ValueError( + f"`per_token_weights` shape must match `labels` shape, got {tuple(per_token_weights.shape)} != {tuple(labels.shape)}." + ) + weights = per_token_weights.to(dtype=logits.dtype) + + # Supervised CE (optionally weighted). + vocab_size = logits.shape[-1] + per_token_nll = F.cross_entropy( + logits.reshape(-1, vocab_size), + labels.reshape(-1), + reduction="none", + ignore_index=ignore_index, + ).reshape_as(labels) + + denom_sft = (weights * valid.to(weights.dtype)).sum().clamp_min(1) + loss_sft = (per_token_nll * weights * valid.to(per_token_nll.dtype)).sum() / denom_sft + + # Confidence loss: penalize entropy only where prediction is already correct. + if lambda_conf == 0.0: + loss_conf = torch.zeros((), device=logits.device, dtype=loss_sft.dtype) + return loss_sft, loss_sft, loss_conf + + with torch.no_grad(): + pred = logits.argmax(dim=-1) + correct = valid & pred.eq(labels) + + scaled_logits = logits.float() + if temperature != 1.0: + scaled_logits = scaled_logits / float(temperature) + + probs = torch.softmax(scaled_logits, dim=-1) + eps = torch.finfo(probs.dtype).tiny + log_probs = torch.log(probs.clamp_min(eps)) + entropy = -(probs * log_probs).sum(dim=-1).to(dtype=logits.dtype) + + denom_conf = (weights * correct.to(weights.dtype)).sum().clamp_min(1) + loss_conf = (entropy * weights * correct.to(entropy.dtype)).sum() / denom_conf + + loss = loss_sft + float(lambda_conf) * loss_conf + return loss, loss_sft, loss_conf + + +def resolve_interpolation_mode(interpolation_type: str): + """ + Maps a string describing an interpolation function to the corresponding torchvision `InterpolationMode` enum. The + full list of supported enums is documented at + https://pytorch.org/vision/0.9/transforms.html#torchvision.transforms.functional.InterpolationMode. + + Args: + interpolation_type (`str`): + A string describing an interpolation method. Currently, `bilinear`, `bicubic`, `box`, `nearest`, + `nearest_exact`, `hamming`, and `lanczos` are supported, corresponding to the supported interpolation modes + in torchvision. + + Returns: + `torchvision.transforms.InterpolationMode`: an `InterpolationMode` enum used by torchvision's `resize` + transform. + """ + if not is_torchvision_available(): + raise ImportError( + "Please make sure to install `torchvision` to be able to use the `resolve_interpolation_mode()` function." + ) + + if interpolation_type == "bilinear": + interpolation_mode = transforms.InterpolationMode.BILINEAR + elif interpolation_type == "bicubic": + interpolation_mode = transforms.InterpolationMode.BICUBIC + elif interpolation_type == "box": + interpolation_mode = transforms.InterpolationMode.BOX + elif interpolation_type == "nearest": + interpolation_mode = transforms.InterpolationMode.NEAREST + elif interpolation_type == "nearest_exact": + interpolation_mode = transforms.InterpolationMode.NEAREST_EXACT + elif interpolation_type == "hamming": + interpolation_mode = transforms.InterpolationMode.HAMMING + elif interpolation_type == "lanczos": + interpolation_mode = transforms.InterpolationMode.LANCZOS + else: + raise ValueError( + f"The given interpolation mode {interpolation_type} is not supported. Currently supported interpolation" + f" modes are `bilinear`, `bicubic`, `box`, `nearest`, `nearest_exact`, `hamming`, and `lanczos`." + ) + + return interpolation_mode + + +def compute_dream_and_update_latents( + unet: UNet2DConditionModel, + noise_scheduler: SchedulerMixin, + timesteps: torch.Tensor, + noise: torch.Tensor, + noisy_latents: torch.Tensor, + target: torch.Tensor, + encoder_hidden_states: torch.Tensor, + dream_detail_preservation: float = 1.0, +) -> tuple[torch.Tensor, torch.Tensor]: + """ + Implements "DREAM (Diffusion Rectification and Estimation-Adaptive Models)" from + https://huggingface.co/papers/2312.00210. DREAM helps align training with sampling to help training be more + efficient and accurate at the cost of an extra forward step without gradients. + + Args: + `unet`: The state unet to use to make a prediction. + `noise_scheduler`: The noise scheduler used to add noise for the given timestep. + `timesteps`: The timesteps for the noise_scheduler to user. + `noise`: A tensor of noise in the shape of noisy_latents. + `noisy_latents`: Previously noise latents from the training loop. + `target`: The ground-truth tensor to predict after eps is removed. + `encoder_hidden_states`: Text embeddings from the text model. + `dream_detail_preservation`: A float value that indicates detail preservation level. + See reference. + + Returns: + `tuple[torch.Tensor, torch.Tensor]`: Adjusted noisy_latents and target. + """ + alphas_cumprod = noise_scheduler.alphas_cumprod.to(timesteps.device)[timesteps, None, None, None] + sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5 + + # The paper uses lambda = sqrt(1 - alpha) ** p, with p = 1 in their experiments. + dream_lambda = sqrt_one_minus_alphas_cumprod**dream_detail_preservation + + pred = None + with torch.no_grad(): + pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample + + _noisy_latents, _target = (None, None) + if noise_scheduler.config.prediction_type == "epsilon": + predicted_noise = pred + delta_noise = (noise - predicted_noise).detach() + delta_noise.mul_(dream_lambda) + _noisy_latents = noisy_latents.add(sqrt_one_minus_alphas_cumprod * delta_noise) + _target = target.add(delta_noise) + elif noise_scheduler.config.prediction_type == "v_prediction": + raise NotImplementedError("DREAM has not been implemented for v-prediction") + else: + raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}") + + return _noisy_latents, _target + + +def unet_lora_state_dict(unet: UNet2DConditionModel) -> dict[str, torch.Tensor]: + r""" + Returns: + A state dict containing just the LoRA parameters. + """ + lora_state_dict = {} + + for name, module in unet.named_modules(): + if hasattr(module, "set_lora_layer"): + lora_layer = getattr(module, "lora_layer") + if lora_layer is not None: + current_lora_layer_sd = lora_layer.state_dict() + for lora_layer_matrix_name, lora_param in current_lora_layer_sd.items(): + # The matrix name can either be "down" or "up". + lora_state_dict[f"{name}.lora.{lora_layer_matrix_name}"] = lora_param + + return lora_state_dict + + +def cast_training_params(model: torch.nn.Module | list[torch.nn.Module], dtype=torch.float32): + """ + Casts the training parameters of the model to the specified data type. + + Args: + model: The PyTorch model whose parameters will be cast. + dtype: The data type to which the model parameters will be cast. + """ + if not isinstance(model, list): + model = [model] + for m in model: + for param in m.parameters(): + # only upcast trainable parameters into fp32 + if param.requires_grad: + param.data = param.to(dtype) + + +def _set_state_dict_into_text_encoder( + lora_state_dict: dict[str, torch.Tensor], prefix: str, text_encoder: torch.nn.Module +): + """ + Sets the `lora_state_dict` into `text_encoder` coming from `transformers`. + + Args: + lora_state_dict: The state dictionary to be set. + prefix: String identifier to retrieve the portion of the state dict that belongs to `text_encoder`. + text_encoder: Where the `lora_state_dict` is to be set. + """ + + text_encoder_state_dict = { + f"{k.replace(prefix, '')}": v for k, v in lora_state_dict.items() if k.startswith(prefix) + } + text_encoder_state_dict = convert_state_dict_to_peft(convert_state_dict_to_diffusers(text_encoder_state_dict)) + set_peft_model_state_dict(text_encoder, text_encoder_state_dict, adapter_name="default") + + +def _collate_lora_metadata(modules_to_save: dict[str, torch.nn.Module]) -> dict[str, Any]: + metadatas = {} + for module_name, module in modules_to_save.items(): + if module is not None: + metadatas[f"{module_name}_lora_adapter_metadata"] = module.peft_config["default"].to_dict() + return metadatas + + +def compute_density_for_timestep_sampling( + weighting_scheme: str, + batch_size: int, + logit_mean: float = None, + logit_std: float = None, + mode_scale: float = None, + device: torch.device | str = "cpu", + generator: torch.Generator | None = None, +): + """ + Compute the density for sampling the timesteps when doing SD3 training. + + Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528. + + SD3 paper reference: https://huggingface.co/papers/2403.03206v1. + """ + if weighting_scheme == "logit_normal": + u = torch.normal(mean=logit_mean, std=logit_std, size=(batch_size,), device=device, generator=generator) + u = torch.nn.functional.sigmoid(u) + elif weighting_scheme == "mode": + u = torch.rand(size=(batch_size,), device=device, generator=generator) + u = 1 - u - mode_scale * (torch.cos(math.pi * u / 2) ** 2 - 1 + u) + else: + u = torch.rand(size=(batch_size,), device=device, generator=generator) + return u + + +def compute_loss_weighting_for_sd3(weighting_scheme: str, sigmas=None): + """ + Computes loss weighting scheme for SD3 training. + + Courtesy: This was contributed by Rafie Walker in https://github.com/huggingface/diffusers/pull/8528. + + SD3 paper reference: https://huggingface.co/papers/2403.03206v1. + """ + if weighting_scheme == "sigma_sqrt": + weighting = (sigmas**-2.0).float() + elif weighting_scheme == "cosmap": + bot = 1 - 2 * sigmas + 2 * sigmas**2 + weighting = 2 / (math.pi * bot) + else: + weighting = torch.ones_like(sigmas) + return weighting + + +def free_memory(): + """ + Runs garbage collection. Then clears the cache of the available accelerator. + """ + gc.collect() + + if torch.cuda.is_available(): + torch.cuda.empty_cache() + elif torch.backends.mps.is_available(): + torch.mps.empty_cache() + elif is_torch_npu_available(): + torch_npu.npu.empty_cache() + elif hasattr(torch, "xpu") and torch.xpu.is_available(): + torch.xpu.empty_cache() + + +@contextmanager +def offload_models(*modules: torch.nn.Module | DiffusionPipeline, device: str | torch.device, offload: bool = True): + """ + Context manager that, if offload=True, moves each module to `device` on enter, then moves it back to its original + device on exit. + + Args: + device (`str` or `torch.Device`): Device to move the `modules` to. + offload (`bool`): Flag to enable offloading. + """ + if offload: + is_model = not any(isinstance(m, DiffusionPipeline) for m in modules) + # record where each module was + if is_model: + original_devices = [next(m.parameters()).device for m in modules] + else: + assert len(modules) == 1 + # For DiffusionPipeline, wrap the device in a list to make it iterable + original_devices = [modules[0].device] + # move to target device + for m in modules: + m.to(device) + + try: + yield + finally: + if offload: + # move back to original devices + for m, orig_dev in zip(modules, original_devices): + m.to(orig_dev) + + +def parse_buckets_string(buckets_str): + """Parses a string defining buckets into a list of (height, width) tuples.""" + if not buckets_str: + raise ValueError("Bucket string cannot be empty.") + + bucket_pairs = buckets_str.strip().split(";") + parsed_buckets = [] + for pair_str in bucket_pairs: + match = re.match(r"^\s*(\d+)\s*,\s*(\d+)\s*$", pair_str) + if not match: + raise ValueError(f"Invalid bucket format: '{pair_str}'. Expected 'height,width'.") + try: + height = int(match.group(1)) + width = int(match.group(2)) + if height <= 0 or width <= 0: + raise ValueError("Bucket dimensions must be positive integers.") + if height % 8 != 0 or width % 8 != 0: + warnings.warn(f"Bucket dimension ({height},{width}) not divisible by 8. This might cause issues.") + parsed_buckets.append((height, width)) + except ValueError as e: + raise ValueError(f"Invalid integer in bucket pair '{pair_str}': {e}") from e + + if not parsed_buckets: + raise ValueError("No valid buckets found in the provided string.") + + return parsed_buckets + + +def find_nearest_bucket(h, w, bucket_options): + """Finds the closes bucket to the given height and width.""" + min_metric = float("inf") + best_bucket_idx = None + for bucket_idx, (bucket_h, bucket_w) in enumerate(bucket_options): + metric = abs(h * bucket_w - w * bucket_h) + if metric <= min_metric: + min_metric = metric + best_bucket_idx = bucket_idx + return best_bucket_idx + + +# Common aspect ratios (width, height) used to generate an aspect-ratio bucket ladder on the fly. +_DEFAULT_BUCKET_ASPECT_RATIOS = [ + (1, 1), + (4, 5), + (5, 4), + (2, 3), + (3, 2), + (3, 4), + (4, 3), + (9, 16), + (16, 9), + (1, 2), + (2, 1), +] + + +def generate_aspect_ratio_buckets(resolution, divisibility=16, base_resolutions=None): + """Generate a discrete list of ``(height, width)`` buckets for a target pixel budget. + + Buckets are sized so that ``height * width`` is close to ``resolution ** 2`` while spanning a range of aspect + ratios, with each dimension rounded to a multiple of ``divisibility``. When ``base_resolutions`` (a list of + ``(height, width)`` pairs, e.g. a model's preferred resolutions) is provided, its aspect ratios are used instead of + the default set and rescaled to the target pixel budget. + + Args: + resolution (`int`): Target resolution; the pixel budget is ``resolution ** 2``. + divisibility (`int`, *optional*, defaults to 16): Each bucket dimension is rounded to a + multiple of this value (16 satisfies the patch-size constraints of the Flux/Z-Image VAEs). + base_resolutions (`list[tuple[int, int]]`, *optional*): ``(height, width)`` pairs whose aspect + ratios seed the ladder instead of the default aspect-ratio set. + + Returns: + `list[tuple[int, int]]`: The unique ``(height, width)`` buckets. + """ + target_pixels = resolution * resolution + if base_resolutions is not None: + aspect_ratios = [(w, h) for (h, w) in base_resolutions] + else: + aspect_ratios = _DEFAULT_BUCKET_ASPECT_RATIOS + + buckets = [] + seen = set() + for rw, rh in aspect_ratios: + aspect = rw / rh + # h * w == target_pixels and w / h == aspect => h = sqrt(target_pixels / aspect) + h = (target_pixels / aspect) ** 0.5 + w = h * aspect + h = max(divisibility, round(h / divisibility) * divisibility) + w = max(divisibility, round(w / divisibility) * divisibility) + if (h, w) not in seen: + seen.add((h, w)) + buckets.append((h, w)) + return buckets + + +def _to_cpu_contiguous(state_dicts) -> dict: + return {k: v.detach().cpu().contiguous() if isinstance(v, torch.Tensor) else v for k, v in state_dicts.items()} + + +def get_fsdp_kwargs_from_accelerator(accelerator) -> dict: + """ + Extract and convert FSDP config from Accelerator into PyTorch FSDP kwargs. + """ + + kwargs = {} + fsdp_state = getattr(accelerator.state, "fsdp_plugin", None) + + if fsdp_state is None: + raise ValueError("Accelerate isn't configured to handle FSDP. Please update your installation.") + + fsdp_plugin = accelerator.state.fsdp_plugin + + if fsdp_plugin is None: + # FSDP not enabled in Accelerator + kwargs["sharding_strategy"] = ShardingStrategy.FULL_SHARD + else: + # FSDP is enabled → use plugin's strategy, or default if None + kwargs["sharding_strategy"] = fsdp_plugin.sharding_strategy or ShardingStrategy.FULL_SHARD + + return kwargs + + +def wrap_with_fsdp( + model: torch.nn.Module, + device: str | torch.device, + offload: bool = True, + use_orig_params: bool = True, + limit_all_gathers: bool = True, + fsdp_kwargs: dict[str, Any] | None = None, + transformer_layer_cls: set[type[torch.nn.Module]] | None = None, +) -> FSDP: + """ + Wrap a model with FSDP using common defaults and optional transformer auto-wrapping. + + Args: + model: Model to wrap + device: Target device (e.g., accelerator.device) + offload: Whether to enable CPU parameter offloading + use_orig_params: Whether to use original parameters + limit_all_gathers: Whether to limit all gathers + fsdp_kwargs: FSDP arguments (sharding_strategy, etc.) — usually from Accelerate config + transformer_layer_cls: Classes for auto-wrapping (if not using policy from fsdp_kwargs) + + Returns: + FSDP-wrapped model + """ + + logger = get_logger(__name__) + + if transformer_layer_cls is None: + # Set the default layers if transformer_layer_cls is not provided + transformer_layer_cls = type(model.model.language_model.layers[0]) + logger.info(f"transformer_layer_cls is not provided, auto-inferred as {transformer_layer_cls.__name__}") + + # Add auto-wrap policy if transformer layers specified + auto_wrap_policy = partial(transformer_auto_wrap_policy, transformer_layer_cls={transformer_layer_cls}) + + config = { + "device_id": device, + "cpu_offload": CPUOffload(offload_params=offload) if offload else None, + "use_orig_params": use_orig_params, + "limit_all_gathers": limit_all_gathers, + "auto_wrap_policy": auto_wrap_policy, + } + + if fsdp_kwargs: + config.update(fsdp_kwargs) + + fsdp_model = FSDP(model, **config) + return fsdp_model + + +# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14 +class EMAModel: + """ + Exponential Moving Average of models weights + """ + + def __init__( + self, + parameters: Iterable[torch.nn.Parameter], + decay: float = 0.9999, + min_decay: float = 0.0, + update_after_step: int = 0, + use_ema_warmup: bool = False, + inv_gamma: float | int = 1.0, + power: float | int = 2 / 3, + foreach: bool = False, + model_cls: Any | None = None, + model_config: dict[str, Any] | None = None, + **kwargs, + ): + """ + Args: + parameters (Iterable[torch.nn.Parameter]): The parameters to track. + decay (float): The decay factor for the exponential moving average. + min_decay (float): The minimum decay factor for the exponential moving average. + update_after_step (int): The number of steps to wait before starting to update the EMA weights. + use_ema_warmup (bool): Whether to use EMA warmup. + inv_gamma (float): + Inverse multiplicative factor of EMA warmup. Default: 1. Only used if `use_ema_warmup` is True. + power (float): Exponential factor of EMA warmup. Default: 2/3. Only used if `use_ema_warmup` is True. + foreach (bool): Use torch._foreach functions for updating shadow parameters. Should be faster. + device (str | torch.device | None): The device to store the EMA weights on. If None, the EMA + weights will be stored on CPU. + + @crowsonkb's notes on EMA Warmup: + If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan + to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps), + gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999 + at 215.4k steps). + """ + + if isinstance(parameters, torch.nn.Module): + deprecation_message = ( + "Passing a `torch.nn.Module` to `ExponentialMovingAverage` is deprecated. " + "Please pass the parameters of the module instead." + ) + deprecate( + "passing a `torch.nn.Module` to `ExponentialMovingAverage`", + "1.0.0", + deprecation_message, + standard_warn=False, + ) + parameters = parameters.parameters() + + # set use_ema_warmup to True if a torch.nn.Module is passed for backwards compatibility + use_ema_warmup = True + + if kwargs.get("max_value", None) is not None: + deprecation_message = "The `max_value` argument is deprecated. Please use `decay` instead." + deprecate("max_value", "1.0.0", deprecation_message, standard_warn=False) + decay = kwargs["max_value"] + + if kwargs.get("min_value", None) is not None: + deprecation_message = "The `min_value` argument is deprecated. Please use `min_decay` instead." + deprecate("min_value", "1.0.0", deprecation_message, standard_warn=False) + min_decay = kwargs["min_value"] + + parameters = list(parameters) + self.shadow_params = [p.clone().detach() for p in parameters] + + if kwargs.get("device", None) is not None: + deprecation_message = "The `device` argument is deprecated. Please use `to` instead." + deprecate("device", "1.0.0", deprecation_message, standard_warn=False) + self.to(device=kwargs["device"]) + + self.temp_stored_params = None + + self.decay = decay + self.min_decay = min_decay + self.update_after_step = update_after_step + self.use_ema_warmup = use_ema_warmup + self.inv_gamma = inv_gamma + self.power = power + self.optimization_step = 0 + self.cur_decay_value = None # set in `step()` + self.foreach = foreach + + self.model_cls = model_cls + self.model_config = model_config + + @classmethod + def from_pretrained(cls, path, model_cls, foreach=False) -> "EMAModel": + _, ema_kwargs = model_cls.from_config(path, return_unused_kwargs=True) + model = model_cls.from_pretrained(path) + + ema_model = cls(model.parameters(), model_cls=model_cls, model_config=model.config, foreach=foreach) + + ema_model.load_state_dict(ema_kwargs) + return ema_model + + def save_pretrained(self, path): + if self.model_cls is None: + raise ValueError("`save_pretrained` can only be used if `model_cls` was defined at __init__.") + + if self.model_config is None: + raise ValueError("`save_pretrained` can only be used if `model_config` was defined at __init__.") + + model = self.model_cls.from_config(self.model_config) + state_dict = self.state_dict() + state_dict.pop("shadow_params", None) + + model.register_to_config(**state_dict) + self.copy_to(model.parameters()) + model.save_pretrained(path) + + def get_decay(self, optimization_step: int) -> float: + """ + Compute the decay factor for the exponential moving average. + """ + step = max(0, optimization_step - self.update_after_step - 1) + + if step <= 0: + return 0.0 + + if self.use_ema_warmup: + cur_decay_value = 1 - (1 + step / self.inv_gamma) ** -self.power + else: + cur_decay_value = (1 + step) / (10 + step) + + cur_decay_value = min(cur_decay_value, self.decay) + # make sure decay is not smaller than min_decay + cur_decay_value = max(cur_decay_value, self.min_decay) + return cur_decay_value + + @torch.no_grad() + def step(self, parameters: Iterable[torch.nn.Parameter]): + if isinstance(parameters, torch.nn.Module): + deprecation_message = ( + "Passing a `torch.nn.Module` to `ExponentialMovingAverage.step` is deprecated. " + "Please pass the parameters of the module instead." + ) + deprecate( + "passing a `torch.nn.Module` to `ExponentialMovingAverage.step`", + "1.0.0", + deprecation_message, + standard_warn=False, + ) + parameters = parameters.parameters() + + parameters = list(parameters) + + self.optimization_step += 1 + + # Compute the decay factor for the exponential moving average. + decay = self.get_decay(self.optimization_step) + self.cur_decay_value = decay + one_minus_decay = 1 - decay + + context_manager = contextlib.nullcontext() + + if self.foreach: + if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): + context_manager = deepspeed.zero.GatheredParameters(parameters, modifier_rank=None) + + with context_manager: + params_grad = [param for param in parameters if param.requires_grad] + s_params_grad = [ + s_param for s_param, param in zip(self.shadow_params, parameters) if param.requires_grad + ] + + if len(params_grad) < len(parameters): + torch._foreach_copy_( + [s_param for s_param, param in zip(self.shadow_params, parameters) if not param.requires_grad], + [param for param in parameters if not param.requires_grad], + non_blocking=True, + ) + + torch._foreach_sub_( + s_params_grad, torch._foreach_sub(s_params_grad, params_grad), alpha=one_minus_decay + ) + + else: + for s_param, param in zip(self.shadow_params, parameters): + if is_transformers_available() and transformers.integrations.deepspeed.is_deepspeed_zero3_enabled(): + context_manager = deepspeed.zero.GatheredParameters(param, modifier_rank=None) + + with context_manager: + if param.requires_grad: + s_param.sub_(one_minus_decay * (s_param - param)) + else: + s_param.copy_(param) + + def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None: + """ + Copy current averaged parameters into given collection of parameters. + + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored moving averages. If `None`, the parameters with which this + `ExponentialMovingAverage` was initialized will be used. + """ + parameters = list(parameters) + if self.foreach: + torch._foreach_copy_( + [param.data for param in parameters], + [s_param.to(param.device).data for s_param, param in zip(self.shadow_params, parameters)], + ) + else: + for s_param, param in zip(self.shadow_params, parameters): + param.data.copy_(s_param.to(param.device).data) + + def pin_memory(self) -> None: + r""" + Move internal buffers of the ExponentialMovingAverage to pinned memory. Useful for non-blocking transfers for + offloading EMA params to the host. + """ + + self.shadow_params = [p.pin_memory() for p in self.shadow_params] + + def to(self, device=None, dtype=None, non_blocking=False) -> None: + r""" + Move internal buffers of the ExponentialMovingAverage to `device`. + + Args: + device: like `device` argument to `torch.Tensor.to` + """ + # .to() on the tensors handles None correctly + self.shadow_params = [ + p.to(device=device, dtype=dtype, non_blocking=non_blocking) + if p.is_floating_point() + else p.to(device=device, non_blocking=non_blocking) + for p in self.shadow_params + ] + + def state_dict(self) -> dict: + r""" + Returns the state of the ExponentialMovingAverage as a dict. This method is used by accelerate during + checkpointing to save the ema state dict. + """ + # Following PyTorch conventions, references to tensors are returned: + # "returns a reference to the state and not its copy!" - + # https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict + return { + "decay": self.decay, + "min_decay": self.min_decay, + "optimization_step": self.optimization_step, + "update_after_step": self.update_after_step, + "use_ema_warmup": self.use_ema_warmup, + "inv_gamma": self.inv_gamma, + "power": self.power, + "shadow_params": self.shadow_params, + } + + def store(self, parameters: Iterable[torch.nn.Parameter]) -> None: + r""" + Saves the current parameters for restoring later. + + Args: + parameters: Iterable of `torch.nn.Parameter`. The parameters to be temporarily stored. + """ + self.temp_stored_params = [param.detach().cpu().clone() for param in parameters] + + def restore(self, parameters: Iterable[torch.nn.Parameter]) -> None: + r""" + Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters + without: affecting the original optimization process. Store the parameters before the `copy_to()` method. After + validation (or model saving), use this to restore the former parameters. + + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. If `None`, the parameters with which this + `ExponentialMovingAverage` was initialized will be used. + """ + + if self.temp_stored_params is None: + raise RuntimeError("This ExponentialMovingAverage has no `store()`ed weights to `restore()`") + if self.foreach: + torch._foreach_copy_( + [param.data for param in parameters], [c_param.data for c_param in self.temp_stored_params] + ) + else: + for c_param, param in zip(self.temp_stored_params, parameters): + param.data.copy_(c_param.data) + + # Better memory-wise. + self.temp_stored_params = None + + def load_state_dict(self, state_dict: dict) -> None: + r""" + Loads the ExponentialMovingAverage state. This method is used by accelerate during checkpointing to save the + ema state dict. + + Args: + state_dict (dict): EMA state. Should be an object returned + from a call to :meth:`state_dict`. + """ + # deepcopy, to be consistent with module API + state_dict = copy.deepcopy(state_dict) + + self.decay = state_dict.get("decay", self.decay) + if self.decay < 0.0 or self.decay > 1.0: + raise ValueError("Decay must be between 0 and 1") + + self.min_decay = state_dict.get("min_decay", self.min_decay) + if not isinstance(self.min_decay, float): + raise ValueError("Invalid min_decay") + + self.optimization_step = state_dict.get("optimization_step", self.optimization_step) + if not isinstance(self.optimization_step, int): + raise ValueError("Invalid optimization_step") + + self.update_after_step = state_dict.get("update_after_step", self.update_after_step) + if not isinstance(self.update_after_step, int): + raise ValueError("Invalid update_after_step") + + self.use_ema_warmup = state_dict.get("use_ema_warmup", self.use_ema_warmup) + if not isinstance(self.use_ema_warmup, bool): + raise ValueError("Invalid use_ema_warmup") + + self.inv_gamma = state_dict.get("inv_gamma", self.inv_gamma) + if not isinstance(self.inv_gamma, (float, int)): + raise ValueError("Invalid inv_gamma") + + self.power = state_dict.get("power", self.power) + if not isinstance(self.power, (float, int)): + raise ValueError("Invalid power") + + shadow_params = state_dict.get("shadow_params", None) + if shadow_params is not None: + self.shadow_params = shadow_params + if not isinstance(self.shadow_params, list): + raise ValueError("shadow_params must be a list") + if not all(isinstance(p, torch.Tensor) for p in self.shadow_params): + raise ValueError("shadow_params must all be Tensors") diff --git a/diffusers/utils/__init__.py b/diffusers/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..bc193138cc84502865fba5786e192c122bc607c1 --- /dev/null +++ b/diffusers/utils/__init__.py @@ -0,0 +1,176 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import os + +from packaging import version + +from .. import __version__ +from .constants import ( + CONFIG_NAME, + DEFAULT_HF_PARALLEL_LOADING_WORKERS, + DEPRECATED_REVISION_ARGS, + DIFFUSERS_DYNAMIC_MODULE_NAME, + DIFFUSERS_LOAD_ID_FIELDS, + FLASHPACK_FILE_EXTENSION, + FLASHPACK_WEIGHTS_NAME, + GGUF_FILE_EXTENSION, + HF_ENABLE_PARALLEL_LOADING, + HF_MODULES_CACHE, + HUGGINGFACE_CO_RESOLVE_ENDPOINT, + MIN_PEFT_VERSION, + ONNX_EXTERNAL_WEIGHTS_NAME, + ONNX_WEIGHTS_NAME, + SAFE_WEIGHTS_INDEX_NAME, + SAFETENSORS_FILE_EXTENSION, + SAFETENSORS_WEIGHTS_NAME, + USE_PEFT_BACKEND, + WEIGHTS_INDEX_NAME, + WEIGHTS_NAME, +) +from .deprecation_utils import _maybe_remap_transformers_class, _resolve_dtype, deprecate +from .doc_utils import replace_example_docstring +from .dynamic_modules_utils import get_class_from_dynamic_module +from .export_utils import encode_video, export_to_gif, export_to_obj, export_to_ply, export_to_video +from .hub_utils import ( + PushToHubMixin, + _add_variant, + _get_checkpoint_shard_files, + _get_model_file, + extract_commit_hash, + http_user_agent, +) +from .import_utils import ( + BACKENDS_MAPPING, + DIFFUSERS_SLOW_IMPORT, + ENV_VARS_TRUE_AND_AUTO_VALUES, + ENV_VARS_TRUE_VALUES, + USE_TF, + USE_TORCH, + DummyObject, + OptionalDependencyNotAvailable, + _LazyModule, + get_objects_from_module, + is_accelerate_available, + is_accelerate_version, + is_aiter_available, + is_aiter_version, + is_auto_round_available, + is_av_available, + is_better_profanity_available, + is_bitsandbytes_available, + is_bitsandbytes_version, + is_bs4_available, + is_cosmos_guardrail_available, + is_flash_attn_3_available, + is_flash_attn_available, + is_flash_attn_version, + is_flashpack_available, + is_ftfy_available, + is_gguf_available, + is_gguf_version, + is_google_colab, + is_hf_hub_version, + is_hpu_available, + is_inflect_available, + is_invisible_watermark_available, + is_kernels_available, + is_kernels_version, + is_kornia_available, + is_librosa_available, + is_matplotlib_available, + is_nltk_available, + is_note_seq_available, + is_nvidia_modelopt_available, + is_nvidia_modelopt_version, + is_onnx_available, + is_opencv_available, + is_optimum_quanto_available, + is_optimum_quanto_version, + is_outlines_available, + is_peft_available, + is_peft_version, + is_pytorch_retinaface_available, + is_safetensors_available, + is_sageattention_available, + is_sageattention_version, + is_scipy_available, + is_sdnq_available, + is_sdnq_version, + is_sentencepiece_available, + is_tensorboard_available, + is_timm_available, + is_torch_available, + is_torch_mlu_available, + is_torch_neuronx_available, + is_torch_npu_available, + is_torch_version, + is_torch_xla_available, + is_torch_xla_version, + is_torchao_available, + is_torchao_version, + is_torchsde_available, + is_torchvision_available, + is_transformers_available, + is_transformers_version, + is_unidecode_available, + is_wandb_available, + is_xformers_available, + is_xformers_version, + requires_backends, +) +from .loading_utils import get_module_from_name, get_submodule_by_name, load_image, load_video +from .logging import get_logger +from .outputs import BaseOutput +from .peft_utils import ( + apply_lora_scale, + check_peft_version, + delete_adapter_layers, + get_adapter_name, + get_peft_kwargs, + recurse_remove_peft_layers, + scale_lora_layers, + set_adapter_layers, + set_weights_and_activate_adapters, + unscale_lora_layers, +) +from .pil_utils import PIL_INTERPOLATION, make_image_grid, numpy_to_pil, pt_to_pil +from .remote_utils import remote_decode +from .state_dict_utils import ( + convert_all_state_dict_to_peft, + convert_sai_sd_control_lora_state_dict_to_peft, + convert_state_dict_to_diffusers, + convert_state_dict_to_kohya, + convert_state_dict_to_peft, + convert_unet_state_dict_to_peft, + state_dict_all_zero, +) +from .typing_utils import _get_detailed_type, _is_valid_type + + +logger = get_logger(__name__) + + +def check_min_version(min_version): + if version.parse(__version__) < version.parse(min_version): + if "dev" in min_version: + error_message = ( + "This example requires a source install from HuggingFace diffusers (see " + "`https://huggingface.co/docs/diffusers/installation#install-from-source`)," + ) + else: + error_message = f"This example requires a minimum version of {min_version}," + error_message += f" but the version found is {__version__}.\n" + raise ImportError(error_message) diff --git a/diffusers/utils/accelerate_utils.py b/diffusers/utils/accelerate_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..ae6a8ca747ac09a316a2622cb9bbf455ddbe3a51 --- /dev/null +++ b/diffusers/utils/accelerate_utils.py @@ -0,0 +1,48 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Accelerate utilities: Utilities related to accelerate +""" + +from packaging import version + +from .import_utils import is_accelerate_available + + +if is_accelerate_available(): + import accelerate + + +def apply_forward_hook(method): + """ + Decorator that applies a registered CpuOffload hook to an arbitrary function rather than `forward`. This is useful + for cases where a PyTorch module provides functions other than `forward` that should trigger a move to the + appropriate acceleration device. This is the case for `encode` and `decode` in [`AutoencoderKL`]. + + This decorator looks inside the internal `_hf_hook` property to find a registered offload hook. + + :param method: The method to decorate. This method should be a method of a PyTorch module. + """ + if not is_accelerate_available(): + return method + accelerate_version = version.parse(accelerate.__version__).base_version + if version.parse(accelerate_version) < version.parse("0.17.0"): + return method + + def wrapper(self, *args, **kwargs): + if hasattr(self, "_hf_hook") and hasattr(self._hf_hook, "pre_forward"): + self._hf_hook.pre_forward(self) + return method(self, *args, **kwargs) + + return wrapper diff --git a/diffusers/utils/constants.py b/diffusers/utils/constants.py new file mode 100644 index 0000000000000000000000000000000000000000..6deb523061d3eba48a4b0e835322c7671efc2ebc --- /dev/null +++ b/diffusers/utils/constants.py @@ -0,0 +1,85 @@ +# Copyright 2025 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import importlib +import os + +from huggingface_hub.constants import HF_HOME +from packaging import version + +from ..dependency_versions_check import dep_version_check +from .import_utils import ENV_VARS_TRUE_VALUES, is_peft_available, is_transformers_available + + +MIN_PEFT_VERSION = "0.6.0" +MIN_TRANSFORMERS_VERSION = "4.34.0" +_CHECK_PEFT = os.environ.get("_CHECK_PEFT", "1") in ENV_VARS_TRUE_VALUES + + +CONFIG_NAME = "config.json" +WEIGHTS_NAME = "diffusion_pytorch_model.bin" +WEIGHTS_INDEX_NAME = "diffusion_pytorch_model.bin.index.json" +ONNX_WEIGHTS_NAME = "model.onnx" +SAFETENSORS_WEIGHTS_NAME = "diffusion_pytorch_model.safetensors" +SAFE_WEIGHTS_INDEX_NAME = "diffusion_pytorch_model.safetensors.index.json" +SAFETENSORS_FILE_EXTENSION = "safetensors" +FLASHPACK_WEIGHTS_NAME = "model.flashpack" +FLASHPACK_FILE_EXTENSION = "flashpack" +GGUF_FILE_EXTENSION = "gguf" +ONNX_EXTERNAL_WEIGHTS_NAME = "weights.pb" +HUGGINGFACE_CO_RESOLVE_ENDPOINT = os.environ.get("HF_ENDPOINT", "https://huggingface.co") +DIFFUSERS_DYNAMIC_MODULE_NAME = "diffusers_modules" +HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(HF_HOME, "modules")) +DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"] +DIFFUSERS_REQUEST_TIMEOUT = 60 +DIFFUSERS_ATTN_BACKEND = os.getenv("DIFFUSERS_ATTN_BACKEND", "native") +DIFFUSERS_ATTN_CHECKS = os.getenv("DIFFUSERS_ATTN_CHECKS", "0").upper() in ENV_VARS_TRUE_VALUES +DEFAULT_HF_PARALLEL_LOADING_WORKERS = 8 +HF_ENABLE_PARALLEL_LOADING = os.environ.get("HF_ENABLE_PARALLEL_LOADING", "").upper() in ENV_VARS_TRUE_VALUES +DIFFUSERS_DISABLE_REMOTE_CODE = os.getenv("DIFFUSERS_DISABLE_REMOTE_CODE", "false").upper() in ENV_VARS_TRUE_VALUES +DIFFUSERS_SDNQ_TRANSFORMERS = os.getenv("DIFFUSERS_SDNQ_TRANSFORMERS", "false").upper() in ENV_VARS_TRUE_VALUES + +# Below should be `True` if the current version of `peft` and `transformers` are compatible with +# PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are +# available. +# For PEFT it is has to be greater than or equal to 0.6.0 and for transformers it has to be greater than or equal to 4.34.0. +_required_peft_version = is_peft_available() and version.parse( + version.parse(importlib.metadata.version("peft")).base_version +) >= version.parse(MIN_PEFT_VERSION) +_required_transformers_version = is_transformers_available() and version.parse( + version.parse(importlib.metadata.version("transformers")).base_version +) >= version.parse(MIN_TRANSFORMERS_VERSION) + +USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version + +if USE_PEFT_BACKEND and _CHECK_PEFT: + dep_version_check("peft") + + +DECODE_ENDPOINT_SD_V1 = "https://q1bj3bpq6kzilnsu.us-east-1.aws.endpoints.huggingface.cloud/" +DECODE_ENDPOINT_SD_XL = "https://x2dmsqunjd6k9prw.us-east-1.aws.endpoints.huggingface.cloud/" +DECODE_ENDPOINT_FLUX = "https://whhx50ex1aryqvw6.us-east-1.aws.endpoints.huggingface.cloud/" +DECODE_ENDPOINT_HUNYUAN_VIDEO = "https://o7ywnmrahorts457.us-east-1.aws.endpoints.huggingface.cloud/" + + +ENCODE_ENDPOINT_SD_V1 = "https://qc6479g0aac6qwy9.us-east-1.aws.endpoints.huggingface.cloud/" +ENCODE_ENDPOINT_SD_XL = "https://xjqqhmyn62rog84g.us-east-1.aws.endpoints.huggingface.cloud/" +ENCODE_ENDPOINT_FLUX = "https://ptccx55jz97f9zgo.us-east-1.aws.endpoints.huggingface.cloud/" + + +DIFFUSERS_LOAD_ID_FIELDS = [ + "pretrained_model_name_or_path", + "subfolder", + "variant", + "revision", +] diff --git a/diffusers/utils/deprecation_utils.py b/diffusers/utils/deprecation_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..27cee39c73a43b60b4d70cb552116ef6594271e7 --- /dev/null +++ b/diffusers/utils/deprecation_utils.py @@ -0,0 +1,112 @@ +import inspect +import warnings +from typing import Any + +from packaging import version + +from ..utils import logging + + +logger = logging.get_logger(__name__) + +_TORCH_DTYPE_DEPRECATION_MESSAGE = "Please use `dtype` instead." + + +def _resolve_dtype(dtype, torch_dtype): + """Resolve the deprecated `torch_dtype` argument into `dtype`.""" + if torch_dtype is None: + return dtype + if dtype is not None: + raise ValueError( + "You have passed both `dtype` and `torch_dtype`. Please only pass `dtype`, `torch_dtype` is deprecated." + ) + deprecate("torch_dtype", "1.0.0", _TORCH_DTYPE_DEPRECATION_MESSAGE) + return torch_dtype + + +# Mapping for deprecated Transformers classes to their replacements +# This is used to handle models that reference deprecated class names in their configs +# Reference: https://github.com/huggingface/transformers/issues/40822 +# Format: { +# "DeprecatedClassName": { +# "new_class": "NewClassName", +# "transformers_version": (">=", "5.0.0"), # (operation, version) tuple +# } +# } +_TRANSFORMERS_CLASS_REMAPPING = { + "CLIPFeatureExtractor": { + "new_class": "CLIPImageProcessor", + "transformers_version": (">", "4.57.0"), + }, +} + + +def _maybe_remap_transformers_class(class_name: str) -> str | None: + """ + Check if a Transformers class should be remapped to a newer version. + + Args: + class_name: The name of the class to check + + Returns: + The new class name if remapping should occur, None otherwise + """ + if class_name not in _TRANSFORMERS_CLASS_REMAPPING: + return None + + from .import_utils import is_transformers_version + + mapping = _TRANSFORMERS_CLASS_REMAPPING[class_name] + operation, required_version = mapping["transformers_version"] + + # Only remap if the transformers version meets the requirement + if is_transformers_version(operation, required_version): + new_class = mapping["new_class"] + logger.warning(f"{class_name} appears to have been deprecated in transformers. Using {new_class} instead.") + return mapping["new_class"] + + return None + + +def deprecate(*args, take_from: dict | Any | None = None, standard_warn=True, stacklevel=2): + from .. import __version__ + + deprecated_kwargs = take_from + values = () + if not isinstance(args[0], tuple): + args = (args,) + + for attribute, version_name, message in args: + if version.parse(version.parse(__version__).base_version) >= version.parse(version_name): + raise ValueError( + f"The deprecation tuple {(attribute, version_name, message)} should be removed since diffusers'" + f" version {__version__} is >= {version_name}" + ) + + warning = None + if isinstance(deprecated_kwargs, dict) and attribute in deprecated_kwargs: + values += (deprecated_kwargs.pop(attribute),) + warning = f"The `{attribute}` argument is deprecated and will be removed in version {version_name}." + elif hasattr(deprecated_kwargs, attribute): + values += (getattr(deprecated_kwargs, attribute),) + warning = f"The `{attribute}` attribute is deprecated and will be removed in version {version_name}." + elif deprecated_kwargs is None: + warning = f"`{attribute}` is deprecated and will be removed in version {version_name}." + + if warning is not None: + warning = warning + " " if standard_warn else "" + warnings.warn(warning + message, FutureWarning, stacklevel=stacklevel) + + if isinstance(deprecated_kwargs, dict) and len(deprecated_kwargs) > 0: + call_frame = inspect.getouterframes(inspect.currentframe())[1] + filename = call_frame.filename + line_number = call_frame.lineno + function = call_frame.function + key, value = next(iter(deprecated_kwargs.items())) + raise TypeError(f"{function} in {filename} line {line_number - 1} got an unexpected keyword argument `{key}`") + + if len(values) == 0: + return + elif len(values) == 1: + return values[0] + return values diff --git a/diffusers/utils/distributed_utils.py b/diffusers/utils/distributed_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..239b7b26200d6b10a371eeb6e67e11f1e9f70882 --- /dev/null +++ b/diffusers/utils/distributed_utils.py @@ -0,0 +1,36 @@ +# Copyright 2025 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +try: + import torch +except ImportError: + torch = None + + +def is_torch_dist_rank_zero() -> bool: + if torch is None: + return True + + dist_module = getattr(torch, "distributed", None) + if dist_module is None or not dist_module.is_available(): + return True + + if not dist_module.is_initialized(): + return True + + try: + return dist_module.get_rank() == 0 + except (RuntimeError, ValueError): + return True diff --git a/diffusers/utils/doc_utils.py b/diffusers/utils/doc_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..3dcd39dbf09b816ac82421d3cb8d74ff963c3b1d --- /dev/null +++ b/diffusers/utils/doc_utils.py @@ -0,0 +1,39 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Doc utilities: Utilities related to documentation +""" + +import re + + +def replace_example_docstring(example_docstring): + def docstring_decorator(fn): + func_doc = fn.__doc__ + lines = func_doc.split("\n") + i = 0 + while i < len(lines) and re.search(r"^\s*Examples?:\s*$", lines[i]) is None: + i += 1 + if i < len(lines): + lines[i] = example_docstring + func_doc = "\n".join(lines) + else: + raise ValueError( + f"The function {fn} should have an empty 'Examples:' in its docstring as placeholder, " + f"current docstring is:\n{func_doc}" + ) + fn.__doc__ = func_doc + return fn + + return docstring_decorator diff --git a/diffusers/utils/dummy_auto_round_objects.py b/diffusers/utils/dummy_auto_round_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..be7a6b8403cb4a3ac167aeb341279bc2a57346c8 --- /dev/null +++ b/diffusers/utils/dummy_auto_round_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class AutoRoundConfig(metaclass=DummyObject): + _backends = ["auto_round"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["auto_round"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["auto_round"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["auto_round"]) diff --git a/diffusers/utils/dummy_bitsandbytes_objects.py b/diffusers/utils/dummy_bitsandbytes_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..2dc589428de960f9c46f5a22874e2ef6475c418c --- /dev/null +++ b/diffusers/utils/dummy_bitsandbytes_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class BitsAndBytesConfig(metaclass=DummyObject): + _backends = ["bitsandbytes"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["bitsandbytes"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["bitsandbytes"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["bitsandbytes"]) diff --git a/diffusers/utils/dummy_gguf_objects.py b/diffusers/utils/dummy_gguf_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..4a6d9a060a1397c5ef0e402e67170960bb3de516 --- /dev/null +++ b/diffusers/utils/dummy_gguf_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class GGUFQuantizationConfig(metaclass=DummyObject): + _backends = ["gguf"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["gguf"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["gguf"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["gguf"]) diff --git a/diffusers/utils/dummy_note_seq_objects.py b/diffusers/utils/dummy_note_seq_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..c02d0b015aedc37c01fb3b843bc79547aae5da68 --- /dev/null +++ b/diffusers/utils/dummy_note_seq_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class MidiProcessor(metaclass=DummyObject): + _backends = ["note_seq"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["note_seq"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["note_seq"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["note_seq"]) diff --git a/diffusers/utils/dummy_nunchaku_lite_objects.py b/diffusers/utils/dummy_nunchaku_lite_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..4ebe7f0ec8630e1ae661d4c207578a0d64b41b18 --- /dev/null +++ b/diffusers/utils/dummy_nunchaku_lite_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class NunchakuLiteQuantizationConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) diff --git a/diffusers/utils/dummy_nvidia_modelopt_objects.py b/diffusers/utils/dummy_nvidia_modelopt_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..046b28223b3d4efd541cbe35d6596b746aecca5f --- /dev/null +++ b/diffusers/utils/dummy_nvidia_modelopt_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class NVIDIAModelOptConfig(metaclass=DummyObject): + _backends = ["nvidia_modelopt"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["nvidia_modelopt"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["nvidia_modelopt"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["nvidia_modelopt"]) diff --git a/diffusers/utils/dummy_onnx_objects.py b/diffusers/utils/dummy_onnx_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..bde5f6ad0793e2d81bc638600b46ff81748d09ee --- /dev/null +++ b/diffusers/utils/dummy_onnx_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class OnnxRuntimeModel(metaclass=DummyObject): + _backends = ["onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["onnx"]) diff --git a/diffusers/utils/dummy_optimum_quanto_objects.py b/diffusers/utils/dummy_optimum_quanto_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..44f8eaffc2460360d17720578b8224496f91cf94 --- /dev/null +++ b/diffusers/utils/dummy_optimum_quanto_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class QuantoConfig(metaclass=DummyObject): + _backends = ["optimum_quanto"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["optimum_quanto"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["optimum_quanto"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["optimum_quanto"]) diff --git a/diffusers/utils/dummy_pt_objects.py b/diffusers/utils/dummy_pt_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..f7a8b6ec9de777828cf4762a9c92d8e294282fb0 --- /dev/null +++ b/diffusers/utils/dummy_pt_objects.py @@ -0,0 +1,3662 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class AdaptiveProjectedGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AdaptiveProjectedMixGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BaseGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ClassifierFreeGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ClassifierFreeZeroStarGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FrequencyDecoupledGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PerturbedAttentionGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SkipLayerGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SmoothedEnergyGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TangentialClassifierFreeGuidance(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FasterCacheConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FirstBlockCacheConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HookRegistry(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LayerSkipConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MagCacheConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PyramidAttentionBroadcastConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SmoothedEnergyGuidanceConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TaylorSeerCacheConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TextKVCacheConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +def apply_faster_cache(*args, **kwargs): + requires_backends(apply_faster_cache, ["torch"]) + + +def apply_first_block_cache(*args, **kwargs): + requires_backends(apply_first_block_cache, ["torch"]) + + +def apply_layer_skip(*args, **kwargs): + requires_backends(apply_layer_skip, ["torch"]) + + +def apply_mag_cache(*args, **kwargs): + requires_backends(apply_mag_cache, ["torch"]) + + +def apply_pyramid_attention_broadcast(*args, **kwargs): + requires_backends(apply_pyramid_attention_broadcast, ["torch"]) + + +def apply_taylorseer_cache(*args, **kwargs): + requires_backends(apply_taylorseer_cache, ["torch"]) + + +def apply_text_kv_cache(*args, **kwargs): + requires_backends(apply_text_kv_cache, ["torch"]) + + +class InpaintProcessor(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class IPAdapterMaskProcessor(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PixArtImageProcessor(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VaeImageProcessor(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VaeImageProcessorLDM3D(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AceStepTransformer1DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AllegroTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AnimaTextConditioner(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AnyFlowFARTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AnyFlowTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AsymmetricAutoencoderKL(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AttentionBackendName(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AuraFlowTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderDC(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKL(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLAllegro(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLCogVideoX(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLCosmos(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLFlux2(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLHunyuanImage(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLHunyuanImageRefiner(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLHunyuanVideo(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLHunyuanVideo15(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLKVAE(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLKVAEVideo(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLLTX2Audio(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLLTX2Video(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLLTXVideo(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLMagvit(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLMiniMaxH3(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLMiniMaxH3Audio(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLMochi(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLQwenImage(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLTemporalDecoder(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderKLWan(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderOobleck(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderRAE(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderTiny(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoencoderVidTok(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BriaFiboTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BriaTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CacheMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ChromaTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ChronoEditTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogVideoXTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogView3PlusTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogView4Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConsisIDTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConsistencyDecoderVAE(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ContextParallelConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ControlNetUnionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ControlNetXSAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Cosmos3AVAEAudioTokenizer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Cosmos3OmniTransformer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CosmosControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CosmosTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiTTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DreamLiteTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DreamLiteUNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EasyAnimateTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ErnieImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Flux2Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FluxControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FluxMultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FluxTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class GlmImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HeliosTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HiDreamImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanDiT2DControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanDiT2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanDiT2DMultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanVideo15Transformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanVideoFramepackTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HunyuanVideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class I2VGenXLUNet(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Ideogram4Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class JoyImageEditPlusTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class JoyImageEditTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Kandinsky3UNet(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Kandinsky5Transformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Krea2Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LatteTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LongCatAudioDiTTransformer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LongCatAudioDiTVae(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LongCatImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LTX2VideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LTXVideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Lumina2Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LuminaNextDiT2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MiniMaxH3Transformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MochiTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ModelMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MotifVideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MotionAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MultiAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class NucleusMoEImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class OmniGenTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class OvisImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ParallelConfig(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PixArtTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PriorTransformer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PRXTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class QwenImageControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class QwenImageMultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class QwenImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SanaControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SanaTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SanaVideoTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SD3ControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SD3MultiControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SD3Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SkyReelsV2Transformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SparseControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class StableAudioDiTModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class T2IAdapter(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class T5FilmDecoder(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class Transformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TransformerTemporalModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet1DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet2DConditionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNet3DConditionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNetControlNetXSModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNetMotionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UNetSpatioTemporalConditionModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UVit2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VQModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class WanAnimateTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class WanTransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class WanVACETransformer3DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ZImageControlNetModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ZImageTransformer2DModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +def attention_backend(*args, **kwargs): + requires_backends(attention_backend, ["torch"]) + + +class AutoPipelineBlocks(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ComponentsManager(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ComponentSpec(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConditionalPipelineBlocks(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConfigSpec(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class InputParam(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LoopSequentialPipelineBlocks(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ModularPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ModularPipelineBlocks(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class OutputParam(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SequentialPipelineBlocks(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +def get_constant_schedule(*args, **kwargs): + requires_backends(get_constant_schedule, ["torch"]) + + +def get_constant_schedule_with_warmup(*args, **kwargs): + requires_backends(get_constant_schedule_with_warmup, ["torch"]) + + +def get_cosine_schedule_with_warmup(*args, **kwargs): + requires_backends(get_cosine_schedule_with_warmup, ["torch"]) + + +def get_cosine_with_hard_restarts_schedule_with_warmup(*args, **kwargs): + requires_backends(get_cosine_with_hard_restarts_schedule_with_warmup, ["torch"]) + + +def get_linear_schedule_with_warmup(*args, **kwargs): + requires_backends(get_linear_schedule_with_warmup, ["torch"]) + + +def get_polynomial_decay_schedule_with_warmup(*args, **kwargs): + requires_backends(get_polynomial_decay_schedule_with_warmup, ["torch"]) + + +def get_scheduler(*args, **kwargs): + requires_backends(get_scheduler, ["torch"]) + + +class AudioPipelineOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForImage2Image(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForInpainting(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForText2Audio(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AutoPipelineForText2Image(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BlipDiffusionControlNetPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BlipDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CLIPImageProjection(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ConsistencyModelPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DanceDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiffusionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiTPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ImagePipelineOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KarrasVePipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LDMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LDMSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PNDMPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class RePaintPipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ScoreSdeVePipeline(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class StableDiffusionMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiffusersQuantizer(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class AmusedScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BlockRefinementScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class BlockRefinementSchedulerOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CMStochasticIterativeScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogVideoXDDIMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class CogVideoXDPMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMInverseScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMParallelScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDIMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMParallelScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DDPMWuerstchenScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DEISMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiscreteDDIMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DiscreteDDIMSchedulerOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DPMSolverMultistepInverseScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class DPMSolverSinglestepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EDMDPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EDMEulerScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EntropyBoundScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EntropyBoundSchedulerOutput(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EulerAncestralDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EulerDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMapEulerDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMatchEulerDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMatchHeunDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class FlowMatchLCMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HeliosDMDScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HeliosScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class HeunDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class IPNDMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KarrasVeScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KDPM2AncestralDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class KDPM2DiscreteScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LCMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class LTXEulerAncestralRFScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class MiniMaxH3Scheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class PNDMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class RePaintScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SASolverScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SchedulerMixin(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class SCMScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class ScoreSdeVeScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class TCDScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UnCLIPScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class UniPCMultistepScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VQDiffusionScheduler(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class EMAModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + +class VideoProcessor(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch"]) diff --git a/diffusers/utils/dummy_sdnq_objects.py b/diffusers/utils/dummy_sdnq_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..688e9632f9093bc5cbee1873efddea48014286f9 --- /dev/null +++ b/diffusers/utils/dummy_sdnq_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class SDNQConfig(metaclass=DummyObject): + _backends = ["sdnq"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["sdnq"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["sdnq"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["sdnq"]) diff --git a/diffusers/utils/dummy_torch_and_librosa_objects.py b/diffusers/utils/dummy_torch_and_librosa_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..2088bc4a744198284f22fe54e6f1055cf3568566 --- /dev/null +++ b/diffusers/utils/dummy_torch_and_librosa_objects.py @@ -0,0 +1,32 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class AudioDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "librosa"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "librosa"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) + + +class Mel(metaclass=DummyObject): + _backends = ["torch", "librosa"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "librosa"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "librosa"]) diff --git a/diffusers/utils/dummy_torch_and_scipy_objects.py b/diffusers/utils/dummy_torch_and_scipy_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..a1ff25863822b04971d2c6dfdc17f5b28774cf05 --- /dev/null +++ b/diffusers/utils/dummy_torch_and_scipy_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class LMSDiscreteScheduler(metaclass=DummyObject): + _backends = ["torch", "scipy"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "scipy"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "scipy"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "scipy"]) diff --git a/diffusers/utils/dummy_torch_and_torchsde_objects.py b/diffusers/utils/dummy_torch_and_torchsde_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..6ff14231b9cc8c88effc4d11f7b439fa16bb7f20 --- /dev/null +++ b/diffusers/utils/dummy_torch_and_torchsde_objects.py @@ -0,0 +1,32 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class CosineDPMSolverMultistepScheduler(metaclass=DummyObject): + _backends = ["torch", "torchsde"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "torchsde"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) + + +class DPMSolverSDEScheduler(metaclass=DummyObject): + _backends = ["torch", "torchsde"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "torchsde"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "torchsde"]) diff --git a/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py b/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..b7afad8226b87292100270e3e7daad6885be0e7f --- /dev/null +++ b/diffusers/utils/dummy_torch_and_transformers_and_onnx_objects.py @@ -0,0 +1,92 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class OnnxStableDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionInpaintPipelineLegacy(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class OnnxStableDiffusionUpscalePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + +class StableDiffusionOnnxPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "onnx"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "onnx"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "onnx"]) diff --git a/diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py b/diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..c0c4d9df5eeeca4cca0a5147058eba70f89bcbc2 --- /dev/null +++ b/diffusers/utils/dummy_torch_and_transformers_and_opencv_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class ConsisIDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "opencv"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "opencv"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "opencv"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "opencv"]) diff --git a/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py b/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..349db10d1fa5b8505173e1578d3c83ea17e95fbc --- /dev/null +++ b/diffusers/utils/dummy_torch_and_transformers_and_sentencepiece_objects.py @@ -0,0 +1,47 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class KolorsImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + +class KolorsPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + +class KolorsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers", "sentencepiece"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers", "sentencepiece"]) diff --git a/diffusers/utils/dummy_torch_and_transformers_objects.py b/diffusers/utils/dummy_torch_and_transformers_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..d5abf9efa891b524ba67cdedf8109cdf9781a2bc --- /dev/null +++ b/diffusers/utils/dummy_torch_and_transformers_objects.py @@ -0,0 +1,5252 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class AnimaAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimaModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos3DistilledBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos3DistilledModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos3OmniBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos3OmniModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ErnieImageAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ErnieImageModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2AutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinBaseAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinBaseModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxKontextAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxKontextModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosPyramidAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosPyramidDistilledAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosPyramidDistilledModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosPyramidModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideo15AutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideo15ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Ideogram4AutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Ideogram4ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Krea2AutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Krea2ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Krea2TurboAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Krea2TurboModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MiniMaxH3Blocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MiniMaxH3ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MiniMaxH3Ref2VABlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MiniMaxH3Ref2VAModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditPlusAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditPlusModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageLayeredAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageLayeredModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3AutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Wan22Blocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Wan22Image2VideoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Wan22Image2VideoModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Wan22ModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanImage2VideoAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanImage2VideoModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageAutoBlocks(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageModularPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AceStepAudioTokenDetokenizer(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AceStepAudioTokenizer(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AceStepConditionEncoder(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AceStepPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AllegroPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AltDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AltDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AmusedImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AmusedInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AmusedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffSDXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffSparseControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffVideoToVideoControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnimateDiffVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnyFlowFARPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AnyFlowPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDM2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDM2ProjectionModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDM2UNet2DConditionModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AudioLDMPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class AuraFlowPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class BriaFiboEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class BriaFiboPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class BriaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ChromaImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ChromaInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ChromaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ChronoEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CLIPImageProjection(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXFunControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogVideoXVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogView3PlusPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogView4ControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CogView4Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ConsisIDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos2_5_PredictBasePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos2_5_TransferPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos2TextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos2VideoToWorldPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Cosmos3OmniPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CosmosActionCondition(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CosmosTextToWorldPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CosmosVideoToWorldPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class CycleDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class DiffusionGemmaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class DiffusionGemmaPipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class DreamLiteMobilePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class DreamLitePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class DreamLitePipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class EasyAnimateControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class EasyAnimateInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class EasyAnimatePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ErnieImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinKVPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2KleinPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Flux2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlNetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxControlPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxFillPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxKontextInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxKontextPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class FluxPriorReduxPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class GlmImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HeliosPyramidPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HiDreamImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanDiTControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanDiTPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanDiTPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanImageRefinerPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanSkyreelsImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideo15ImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideo15Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideoFramepackPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideoImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class HunyuanVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class I2VGenXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Ideogram4Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Ideogram4PromptEnhancerHead(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFImg2ImgSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFInpaintingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFInpaintingSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class IFSuperResolutionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ImageTextPipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class JoyImageEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class JoyImageEditPipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class JoyImageEditPlusPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class JoyImageEditPlusPipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky3Img2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky3Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky5I2IPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky5I2VPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky5T2IPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Kandinsky5T2VPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyImg2ImgCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyInpaintCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyPriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22CombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22ControlnetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22ControlnetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22Img2ImgCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22Img2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22InpaintCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22InpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22PriorEmb2EmbPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class KandinskyV22PriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Krea2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LatentConsistencyModelImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LatentConsistencyModelPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LattePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LDMTextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LEditsPPPipelineStableDiffusion(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LEditsPPPipelineStableDiffusionXL(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LLaDA2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LLaDA2PipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LongCatAudioDiTPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LongCatImageEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LongCatImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTX2ConditionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTX2HDRPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTX2ImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTX2InContextPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTX2LatentUpsamplePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTX2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXConditionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXI2VLongMultiPromptPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXLatentUpsamplePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LTXPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LucyEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Lumina2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class Lumina2Text2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LuminaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class LuminaText2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MarigoldDepthPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MarigoldIntrinsicsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MarigoldNormalsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MochiPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MotifVideoImage2VideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MotifVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MotifVideoPipelineOutput(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class MusicLDMPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class NucleusMoEImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class OmniGenPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class OvisImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PaintByExamplePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PIAPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PixArtAlphaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PixArtSigmaPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PixArtSigmaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PRXPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class PRXPixelPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageEditPlusPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImageLayeredPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class QwenImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ReduxImageEncoder(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaSprintImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaSprintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SanaVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SemanticStableDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ShapEImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ShapEPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SkyReelsV2DiffusionForcingImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SkyReelsV2DiffusionForcingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SkyReelsV2DiffusionForcingVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SkyReelsV2ImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class SkyReelsV2Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableAudioPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableAudioProjectionModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableCascadeCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableCascadeDecoderPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableCascadePriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3ControlNetInpaintingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3ControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3Img2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3InpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3PAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3PAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusion3Pipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionAdapterPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionAttendAndExcitePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetPAGInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionControlNetXSPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionDepth2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionDiffEditPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionGLIGENPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionGLIGENTextImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionInpaintPipelineLegacy(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionInstructPix2PixPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionLatentUpscalePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionLDM3DPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionModelEditingPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPAGInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPanoramaPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionParadigmsPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPipelineSafe(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionPix2PixZeroPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionSAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionUpscalePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLAdapterPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetPAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetUnionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLControlNetXSPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLInstructPix2PixPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPAGImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPAGInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPAGPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableDiffusionXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableUnCLIPImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableUnCLIPPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class StableVideoDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class TextToVideoSDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class TextToVideoZeroPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class TextToVideoZeroSDXLPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UnCLIPImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UnCLIPPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UniDiffuserModel(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UniDiffuserPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class UniDiffuserTextDecoder(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionDualGuidedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionImageVariationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VersatileDiffusionTextToImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VideoToVideoSDPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VisualClozeGenerationPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VisualClozePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class VQDiffusionPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanAnimatePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanImageToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanVACEPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WanVideoToVideoPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WuerstchenCombinedPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WuerstchenDecoderPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class WuerstchenPriorPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageControlNetInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageControlNetPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageImg2ImgPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageInpaintPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImageOmniPipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + +class ZImagePipeline(metaclass=DummyObject): + _backends = ["torch", "transformers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch", "transformers"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torch", "transformers"]) diff --git a/diffusers/utils/dummy_torchao_objects.py b/diffusers/utils/dummy_torchao_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..16f0f6a55f64099c73ce4d81af8c87053b833e44 --- /dev/null +++ b/diffusers/utils/dummy_torchao_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class TorchAoConfig(metaclass=DummyObject): + _backends = ["torchao"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torchao"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["torchao"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["torchao"]) diff --git a/diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py b/diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py new file mode 100644 index 0000000000000000000000000000000000000000..fbde04e33f0abd86d12f3dee048a4f0585c9f19d --- /dev/null +++ b/diffusers/utils/dummy_transformers_and_torch_and_note_seq_objects.py @@ -0,0 +1,17 @@ +# This file is autogenerated by the command `make fix-copies`, do not edit. +from ..utils import DummyObject, requires_backends + + +class SpectrogramDiffusionPipeline(metaclass=DummyObject): + _backends = ["transformers", "torch", "note_seq"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["transformers", "torch", "note_seq"]) + + @classmethod + def from_config(cls, *args, **kwargs): + requires_backends(cls, ["transformers", "torch", "note_seq"]) + + @classmethod + def from_pretrained(cls, *args, **kwargs): + requires_backends(cls, ["transformers", "torch", "note_seq"]) diff --git a/diffusers/utils/dynamic_modules_utils.py b/diffusers/utils/dynamic_modules_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..46aeb514484d3fac0f8954a9ede794c46cbbd6cb --- /dev/null +++ b/diffusers/utils/dynamic_modules_utils.py @@ -0,0 +1,544 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Utilities to dynamically load objects from the Hub.""" + +import importlib +import inspect +import json +import os +import re +import shutil +import sys +import threading +from pathlib import Path +from types import ModuleType +from urllib import request + +from huggingface_hub import hf_hub_download, model_info +from huggingface_hub.utils import RevisionNotFoundError, validate_hf_hub_args +from packaging import version + +from .. import __version__ +from . import DIFFUSERS_DYNAMIC_MODULE_NAME, HF_MODULES_CACHE, logging +from .constants import DIFFUSERS_DISABLE_REMOTE_CODE + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +# See https://huggingface.co/datasets/diffusers/community-pipelines-mirror +COMMUNITY_PIPELINES_MIRROR_ID = "diffusers/community-pipelines-mirror" +TIME_OUT_REMOTE_CODE = int(os.getenv("DIFFUSERS_TIMEOUT_REMOTE_CODE", 15)) +_HF_REMOTE_CODE_LOCK = threading.Lock() + + +def get_diffusers_versions(): + url = "https://pypi.org/pypi/diffusers/json" + releases = json.loads(request.urlopen(url).read())["releases"].keys() + return sorted(releases, key=lambda x: version.Version(x)) + + +def init_hf_modules(): + """ + Creates the cache directory for modules with an init, and adds it to the Python path. + """ + # This function has already been executed if HF_MODULES_CACHE already is in the Python path. + if HF_MODULES_CACHE in sys.path: + return + + sys.path.append(HF_MODULES_CACHE) + os.makedirs(HF_MODULES_CACHE, exist_ok=True) + init_path = Path(HF_MODULES_CACHE) / "__init__.py" + if not init_path.exists(): + init_path.touch() + + +def create_dynamic_module(name: str | os.PathLike): + """ + Creates a dynamic module in the cache directory for modules. + """ + init_hf_modules() + dynamic_module_path = Path(HF_MODULES_CACHE) / name + # If the parent module does not exist yet, recursively create it. + if not dynamic_module_path.parent.exists(): + create_dynamic_module(dynamic_module_path.parent) + os.makedirs(dynamic_module_path, exist_ok=True) + init_path = dynamic_module_path / "__init__.py" + if not init_path.exists(): + init_path.touch() + + +def get_relative_imports(module_file): + """ + Get the list of modules that are relatively imported in a module file. + + Args: + module_file (`str` or `os.PathLike`): The module file to inspect. + """ + with open(module_file, "r", encoding="utf-8") as f: + content = f.read() + + # Imports of the form `import .xxx` + relative_imports = re.findall(r"^\s*import\s+\.(\S+)\s*$", content, flags=re.MULTILINE) + # Imports of the form `from .xxx import yyy` + relative_imports += re.findall(r"^\s*from\s+\.(\S+)\s+import", content, flags=re.MULTILINE) + # Unique-ify + return list(set(relative_imports)) + + +def get_relative_import_files(module_file): + """ + Get the list of all files that are needed for a given module. Note that this function recurses through the relative + imports (if a imports b and b imports c, it will return module files for b and c). + + Args: + module_file (`str` or `os.PathLike`): The module file to inspect. + """ + no_change = False + files_to_check = [module_file] + all_relative_imports = [] + + # Let's recurse through all relative imports + while not no_change: + new_imports = [] + for f in files_to_check: + new_imports.extend(get_relative_imports(f)) + + module_path = Path(module_file).parent + new_import_files = [str(module_path / m) for m in new_imports] + new_import_files = [f for f in new_import_files if f not in all_relative_imports] + files_to_check = [f"{f}.py" for f in new_import_files] + + no_change = len(new_import_files) == 0 + all_relative_imports.extend(files_to_check) + + return all_relative_imports + + +def check_imports(filename): + """ + Check if the current Python environment contains all the libraries that are imported in a file. + """ + with open(filename, "r", encoding="utf-8") as f: + content = f.read() + + # Imports of the form `import xxx` + imports = re.findall(r"^\s*import\s+(\S+)\s*$", content, flags=re.MULTILINE) + # Imports of the form `from xxx import yyy` + imports += re.findall(r"^\s*from\s+(\S+)\s+import", content, flags=re.MULTILINE) + # Only keep the top-level module + imports = [imp.split(".")[0] for imp in imports if not imp.startswith(".")] + + # Unique-ify and test we got them all + imports = list(set(imports)) + missing_packages = [] + for imp in imports: + try: + importlib.import_module(imp) + except ImportError: + missing_packages.append(imp) + + if len(missing_packages) > 0: + logger.warning( + "This modeling file might require the following packages that were not found in your environment: " + f"{', '.join(missing_packages)}. Run `pip install {' '.join(missing_packages)}`" + ) + + return get_relative_imports(filename) + + +def resolve_trust_remote_code(trust_remote_code, model_name, has_remote_code): + trust_remote_code = trust_remote_code and not DIFFUSERS_DISABLE_REMOTE_CODE + if DIFFUSERS_DISABLE_REMOTE_CODE: + logger.warning( + "Downloading remote code is disabled globally via the DIFFUSERS_DISABLE_REMOTE_CODE environment variable. Ignoring `trust_remote_code`." + ) + + if has_remote_code and not trust_remote_code: + error_msg = f"The repository for {model_name} contains custom code. " + error_msg += ( + "Downloading remote code is disabled globally via the DIFFUSERS_DISABLE_REMOTE_CODE environment variable." + if DIFFUSERS_DISABLE_REMOTE_CODE + else "Pass `trust_remote_code=True` to allow loading remote code modules." + ) + raise ValueError(error_msg) + + elif has_remote_code and trust_remote_code: + logger.warning( + f"`trust_remote_code` is enabled. Downloading code from {model_name}. Please ensure you trust the contents of this repository" + ) + + return trust_remote_code + + +def get_class_in_module(class_name, module_path, force_reload=False): + """ + Import a module on the cache directory for modules and extract a class from it. + """ + name = os.path.normpath(module_path) + if name.endswith(".py"): + name = name[:-3] + name = name.replace(os.path.sep, ".") + module_file: Path = Path(HF_MODULES_CACHE) / module_path + + with _HF_REMOTE_CODE_LOCK: + if force_reload: + sys.modules.pop(name, None) + importlib.invalidate_caches() + cached_module: ModuleType | None = sys.modules.get(name) + module_spec = importlib.util.spec_from_file_location(name, location=module_file) + + module: ModuleType + if cached_module is None: + module = importlib.util.module_from_spec(module_spec) + # insert it into sys.modules before any loading begins + sys.modules[name] = module + else: + module = cached_module + + module_spec.loader.exec_module(module) + + if class_name is None: + return find_pipeline_class(module) + + return getattr(module, class_name) + + +def find_pipeline_class(loaded_module): + """ + Retrieve pipeline class that inherits from `DiffusionPipeline`. Note that there has to be exactly one class + inheriting from `DiffusionPipeline`. + """ + from ..pipelines import DiffusionPipeline + + cls_members = dict(inspect.getmembers(loaded_module, inspect.isclass)) + + pipeline_class = None + for cls_name, cls in cls_members.items(): + if ( + cls_name != DiffusionPipeline.__name__ + and issubclass(cls, DiffusionPipeline) + and cls.__module__.split(".")[0] != "diffusers" + ): + if pipeline_class is not None: + raise ValueError( + f"Multiple classes that inherit from {DiffusionPipeline.__name__} have been found:" + f" {pipeline_class.__name__}, and {cls_name}. Please make sure to define only one in" + f" {loaded_module}." + ) + pipeline_class = cls + + return pipeline_class + + +@validate_hf_hub_args +def get_cached_module_file( + pretrained_model_name_or_path: str | os.PathLike, + module_file: str, + subfolder: str | None = None, + cache_dir: str | os.PathLike | None = None, + force_download: bool = False, + proxies: dict[str, str] | None = None, + token: bool | str | None = None, + revision: str | None = None, + local_files_only: bool = False, + local_dir: str | None = None, + trust_remote_code: bool = False, +): + """ + Prepares Downloads a module from a local folder or a distant repo and returns its path inside the cached + Transformers module. + + Args: + pretrained_model_name_or_path (`str` or `os.PathLike`): + This can be either: + + - a string, the *model id* of a pretrained model configuration hosted inside a model repo on + huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced + under a user or organization name, like `dbmdz/bert-base-german-cased`. + - a path to a *directory* containing a configuration file saved using the + [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`. + + module_file (`str`): + The name of the module file containing the class to look for. + cache_dir (`str` or `os.PathLike`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the standard + cache should not be used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force to (re-)download the configuration files and override the cached versions if they + exist. + proxies (`dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request. + token (`str` or *bool*, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated + when running `transformers-cli login` (stored in `~/.huggingface`). + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a + git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any + identifier allowed by git. + local_files_only (`bool`, *optional*, defaults to `False`): + If `True`, will only try to load the tokenizer configuration from local files. + trust_remote_code (`bool`, *optional*, defaults to `False`): + Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This + option should only be set to `True` for repositories you trust and in which you have read the code, as it + will execute code present on the Hub on your local machine. + + > [!TIP] > You may pass a token in `token` if you are not logged in (`hf auth login`) and want to use private or + [gated > models](https://huggingface.co/docs/hub/models-gated#gated-models). + + Returns: + `str`: The path to the module inside the cache. + """ + # Download and cache module_file from the repo `pretrained_model_name_or_path` of grab it if it's a local file. + pretrained_model_name_or_path = str(pretrained_model_name_or_path) + + if DIFFUSERS_DISABLE_REMOTE_CODE: + raise ValueError( + "Downloading remote code is disabled globally via the DIFFUSERS_DISABLE_REMOTE_CODE environment variable." + ) + + if subfolder is not None: + module_file_or_url = os.path.join(pretrained_model_name_or_path, subfolder, module_file) + else: + module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file) + + is_local_file = os.path.isfile(module_file_or_url) + is_community_pipeline = not is_local_file and pretrained_model_name_or_path.count("/") == 0 + + if is_local_file: + resolved_module_file = module_file_or_url + submodule = "local" + if not trust_remote_code: + raise ValueError( + f"The directory {pretrained_model_name_or_path} contains custom code in {module_file} which must be executed to correctly " + f"load the model. You can inspect the file content at {module_file_or_url}.\n" + f"Pass `trust_remote_code=True` to allow loading remote code modules." + ) + elif is_community_pipeline: + available_versions = get_diffusers_versions() + # cut ".dev0" + latest_version = "v" + ".".join(__version__.split(".")[:3]) + + # retrieve github version that matches + if revision is None: + revision = latest_version if latest_version[1:] in available_versions else "main" + logger.info(f"Defaulting to latest_version: {revision}.") + elif revision in available_versions: + revision = f"v{revision}" + elif revision == "main": + revision = revision + else: + raise ValueError( + f"`custom_revision`: {revision} does not exist. Please make sure to choose one of" + f" {', '.join(available_versions + ['main'])}." + ) + + try: + resolved_module_file = hf_hub_download( + repo_id=COMMUNITY_PIPELINES_MIRROR_ID, + repo_type="dataset", + filename=f"{revision}/{pretrained_model_name_or_path}.py", + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + local_dir=local_dir, + ) + submodule = "git" + module_file = pretrained_model_name_or_path + ".py" + except RevisionNotFoundError as e: + raise EnvironmentError( + f"Revision '{revision}' not found in the community pipelines mirror. Check available revisions on" + " https://huggingface.co/datasets/diffusers/community-pipelines-mirror/tree/main." + " If you don't find the revision you are looking for, please open an issue on https://github.com/huggingface/diffusers/issues." + ) from e + except EnvironmentError: + logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.") + raise + else: + if not trust_remote_code: + raise ValueError( + f"The repository for {pretrained_model_name_or_path} contains custom code in {module_file} which must be executed to correctly " + f"load the model. You can inspect the repository content at https://hf.co/{pretrained_model_name_or_path}/blob/main/{module_file}.\n" + f"Pass `trust_remote_code=True` to allow loading remote code modules." + ) + try: + # Load from URL or cache if already cached + resolved_module_file = hf_hub_download( + pretrained_model_name_or_path, + module_file, + subfolder=subfolder, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + local_dir=local_dir, + revision=revision, + token=token, + ) + submodule = os.path.join("local", "--".join(pretrained_model_name_or_path.split("/"))) + except EnvironmentError: + logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.") + raise + + # Check we have all the requirements in our environment + modules_needed = check_imports(resolved_module_file) + + # Now we move the module inside our cached dynamic modules. + full_submodule = DIFFUSERS_DYNAMIC_MODULE_NAME + os.path.sep + submodule + create_dynamic_module(full_submodule) + submodule_path = Path(HF_MODULES_CACHE) / full_submodule + if submodule == "local" or submodule == "git": + # We always copy local files (we could hash the file to see if there was a change, and give them the name of + # that hash, to only copy when there is a modification but it seems overkill for now). + # The only reason we do the copy is to avoid putting too many folders in sys.path. + shutil.copyfile(resolved_module_file, submodule_path / module_file) + for module_needed in modules_needed: + if len(module_needed.split(".")) == 2: + module_needed = "/".join(module_needed.split(".")) + module_folder = module_needed.split("/")[0] + if not os.path.exists(submodule_path / module_folder): + os.makedirs(submodule_path / module_folder) + module_needed = f"{module_needed}.py" + if subfolder is not None: + source_path = os.path.join(pretrained_model_name_or_path, subfolder, module_needed) + else: + source_path = os.path.join(pretrained_model_name_or_path, module_needed) + shutil.copyfile(source_path, submodule_path / module_needed) + else: + # Get the commit hash + # TODO: we will get this info in the etag soon, so retrieve it from there and not here. + commit_hash = model_info(pretrained_model_name_or_path, revision=revision, token=token).sha + + # The module file will end up being placed in a subfolder with the git hash of the repo. This way we get the + # benefit of versioning. + submodule_path = submodule_path / commit_hash + full_submodule = full_submodule + os.path.sep + commit_hash + create_dynamic_module(full_submodule) + + if not (submodule_path / module_file).exists(): + if len(module_file.split("/")) == 2: + module_folder = module_file.split("/")[0] + if not os.path.exists(submodule_path / module_folder): + os.makedirs(submodule_path / module_folder) + shutil.copyfile(resolved_module_file, submodule_path / module_file) + + # Make sure we also have every file with relative + for module_needed in modules_needed: + if len(module_needed.split(".")) == 2: + module_needed = "/".join(module_needed.split(".")) + if not (submodule_path / module_needed).exists(): + get_cached_module_file( + pretrained_model_name_or_path, + f"{module_needed}.py", + subfolder=subfolder, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + token=token, + revision=revision, + local_files_only=local_files_only, + local_dir=local_dir, + trust_remote_code=trust_remote_code, + ) + return os.path.join(full_submodule, module_file) + + +@validate_hf_hub_args +def get_class_from_dynamic_module( + pretrained_model_name_or_path: str | os.PathLike, + module_file: str, + subfolder: str | None = None, + class_name: str | None = None, + cache_dir: str | os.PathLike | None = None, + force_download: bool = False, + proxies: dict[str, str] | None = None, + token: bool | str | None = None, + revision: str | None = None, + local_files_only: bool = False, + local_dir: str | None = None, + trust_remote_code: bool = False, +): + """ + Extracts a class from a module file, present in the local folder or repository of a model. + + > [!WARNING] > Calling this function will execute the code in the module file found locally or downloaded from the + Hub. It should > therefore only be called on trusted repos. + + Args: + pretrained_model_name_or_path (`str` or `os.PathLike`): + This can be either: + + - a string, the *model id* of a pretrained model configuration hosted inside a model repo on + huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced + under a user or organization name, like `dbmdz/bert-base-german-cased`. + - a path to a *directory* containing a configuration file saved using the + [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`. + + module_file (`str`): + The name of the module file containing the class to look for. + class_name (`str`): + The name of the class to import in the module. + cache_dir (`str` or `os.PathLike`, *optional*): + Path to a directory in which a downloaded pretrained model configuration should be cached if the standard + cache should not be used. + force_download (`bool`, *optional*, defaults to `False`): + Whether or not to force to (re-)download the configuration files and override the cached versions if they + exist. + proxies (`dict[str, str]`, *optional*): + A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128', + 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request. + token (`str` or `bool`, *optional*): + The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated + when running `transformers-cli login` (stored in `~/.huggingface`). + revision (`str`, *optional*, defaults to `"main"`): + The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a + git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any + identifier allowed by git. + local_files_only (`bool`, *optional*, defaults to `False`): + If `True`, will only try to load the tokenizer configuration from local files. + trust_remote_code (`bool`, *optional*, defaults to `False`): + Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This + option should only be set to `True` for repositories you trust and in which you have read the code, as it + will execute code present on the Hub on your local machine. + + > [!TIP] > You may pass a token in `token` if you are not logged in (`hf auth login`) and want to use private or + [gated > models](https://huggingface.co/docs/hub/models-gated#gated-models). + + Returns: + `type`: The class, dynamically imported from the module. + + Examples: + + ```python + # Download module `modeling.py` from huggingface.co and cache then extract the class `MyBertModel` from this + # module. + cls = get_class_from_dynamic_module("sgugger/my-bert-model", "modeling.py", "MyBertModel") + ```""" + # And lastly we get the class inside our newly created module + final_module = get_cached_module_file( + pretrained_model_name_or_path, + module_file, + subfolder=subfolder, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + token=token, + revision=revision, + local_files_only=local_files_only, + local_dir=local_dir, + trust_remote_code=trust_remote_code, + ) + return get_class_in_module(class_name, final_module) diff --git a/diffusers/utils/export_utils.py b/diffusers/utils/export_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..f06f172e79b2c7564bf649dbed2dfcbbe646a916 --- /dev/null +++ b/diffusers/utils/export_utils.py @@ -0,0 +1,388 @@ +from __future__ import annotations + +import io +import random +import struct +import tempfile +from collections.abc import Iterator +from contextlib import contextmanager +from fractions import Fraction +from itertools import chain +from typing import TYPE_CHECKING + +import numpy as np +import PIL.Image +import PIL.ImageOps + +from .import_utils import BACKENDS_MAPPING, is_imageio_available, is_opencv_available +from .logging import get_logger + + +if TYPE_CHECKING: + import torch + + +global_rng = random.Random() + +logger = get_logger(__name__) + + +@contextmanager +def buffered_writer(raw_f): + f = io.BufferedWriter(raw_f) + yield f + f.flush() + + +def export_to_gif(image: list[PIL.Image.Image], output_gif_path: str = None, fps: int = 10) -> str: + if output_gif_path is None: + output_gif_path = tempfile.NamedTemporaryFile(suffix=".gif").name + + image[0].save( + output_gif_path, + save_all=True, + append_images=image[1:], + optimize=False, + duration=1000 // fps, + loop=0, + ) + return output_gif_path + + +def export_to_ply(mesh, output_ply_path: str = None): + """ + Write a PLY file for a mesh. + """ + if output_ply_path is None: + output_ply_path = tempfile.NamedTemporaryFile(suffix=".ply").name + + coords = mesh.verts.detach().cpu().numpy() + faces = mesh.faces.cpu().numpy() + rgb = np.stack([mesh.vertex_channels[x].detach().cpu().numpy() for x in "RGB"], axis=1) + + with buffered_writer(open(output_ply_path, "wb")) as f: + f.write(b"ply\n") + f.write(b"format binary_little_endian 1.0\n") + f.write(bytes(f"element vertex {len(coords)}\n", "ascii")) + f.write(b"property float x\n") + f.write(b"property float y\n") + f.write(b"property float z\n") + if rgb is not None: + f.write(b"property uchar red\n") + f.write(b"property uchar green\n") + f.write(b"property uchar blue\n") + if faces is not None: + f.write(bytes(f"element face {len(faces)}\n", "ascii")) + f.write(b"property list uchar int vertex_index\n") + f.write(b"end_header\n") + + if rgb is not None: + rgb = (rgb * 255.499).round().astype(int) + vertices = [ + (*coord, *rgb) + for coord, rgb in zip( + coords.tolist(), + rgb.tolist(), + ) + ] + format = struct.Struct("<3f3B") + for item in vertices: + f.write(format.pack(*item)) + else: + format = struct.Struct("<3f") + for vertex in coords.tolist(): + f.write(format.pack(*vertex)) + + if faces is not None: + format = struct.Struct(" str: + """ + quality: + Video output quality. Default is 5. Uses variable bit rate. Highest quality is 10, lowest is 0. Set to None to + prevent variable bitrate flags to FFMPEG so you can manually specify them using output_params instead. + Specifying a fixed bitrate using `bitrate` disables this parameter. + + bitrate: + Set a constant bitrate for the video encoding. Default is None causing `quality` parameter to be used instead. + Better quality videos with smaller file sizes will result from using the `quality` variable bitrate parameter + rather than specifying a fixed bitrate with this parameter. + + macro_block_size: + Size constraint for video. Width and height, must be divisible by this number. If not divisible by this number + imageio will tell ffmpeg to scale the image up to the next closest size divisible by this number. Most codecs + are compatible with a macroblock size of 16 (default), some can go smaller (4, 8). To disable this automatic + feature set it to None or 1, however be warned many players can't decode videos that are odd in size and some + codecs will produce poor results or fail. See https://en.wikipedia.org/wiki/Macroblock. + """ + # TODO: Dhruv. Remove by Diffusers release 0.33.0 + # Added to prevent breaking existing code + if not is_imageio_available(): + logger.warning( + ( + "It is recommended to use `export_to_video` with `imageio` and `imageio-ffmpeg` as a backend. \n" + "These libraries are not present in your environment. Attempting to use legacy OpenCV backend to export video. \n" + "Support for the OpenCV backend will be deprecated in a future Diffusers version" + ) + ) + return _legacy_export_to_video(video_frames, output_video_path, fps) + + if is_imageio_available(): + import imageio + else: + raise ImportError(BACKENDS_MAPPING["imageio"][1].format("export_to_video")) + + try: + imageio.plugins.ffmpeg.get_exe() + except AttributeError: + raise AttributeError( + ( + "Found an existing imageio backend in your environment. Attempting to export video with imageio. \n" + "Unable to find a compatible ffmpeg installation in your environment to use with imageio. Please install via `pip install imageio-ffmpeg" + ) + ) + + if output_video_path is None: + output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name + + if isinstance(video_frames[0], np.ndarray): + video_frames = [(frame * 255).astype(np.uint8) for frame in video_frames] + + elif isinstance(video_frames[0], PIL.Image.Image): + video_frames = [np.array(frame) for frame in video_frames] + + with imageio.get_writer( + output_video_path, fps=fps, quality=quality, bitrate=bitrate, macro_block_size=macro_block_size + ) as writer: + for frame in video_frames: + writer.append_data(frame) + + return output_video_path + + +def _import_av(): + try: + import av + except ImportError as error: + raise ImportError( + "PyAV is required to use `encode_video`. You can install it with `pip install av`." + ) from error + + return av + + +def _prepare_audio_stream(container, audio_sample_rate: int): + """ + Prepare the audio stream for writing. + """ + audio_stream = container.add_stream("aac", rate=audio_sample_rate) + audio_stream.codec_context.sample_rate = audio_sample_rate + audio_stream.codec_context.layout = "stereo" + audio_stream.codec_context.time_base = Fraction(1, audio_sample_rate) + return audio_stream + + +def _resample_audio(container, audio_stream, frame_in, av_module) -> None: + cc = audio_stream.codec_context + + # Use the encoder's format/layout/rate as the target. + target_format = cc.format or "fltp" # AAC is usually fltp. + target_layout = cc.layout or "stereo" + target_rate = cc.sample_rate or frame_in.sample_rate + + audio_resampler = av_module.audio.resampler.AudioResampler( + format=target_format, + layout=target_layout, + rate=target_rate, + ) + + audio_next_pts = 0 + for rframe in audio_resampler.resample(frame_in): + if rframe.pts is None: + rframe.pts = audio_next_pts + audio_next_pts += rframe.samples + rframe.sample_rate = frame_in.sample_rate + container.mux(audio_stream.encode(rframe)) + + # Flush audio encoder. + for packet in audio_stream.encode(): + container.mux(packet) + + +def _write_audio( + container, + audio_stream, + samples: "torch.Tensor", + audio_sample_rate: int, + av_module, +) -> None: + import torch + + if samples.ndim == 1: + samples = samples[:, None] + + if samples.shape[1] != 2 and samples.shape[0] == 2: + samples = samples.T + + if samples.shape[1] != 2: + raise ValueError(f"Expected samples with 2 channels; got shape {samples.shape}.") + + # Convert to int16 packed for ingestion; resampler converts to the encoder format. + if samples.dtype != torch.int16: + samples = torch.clip(samples, -1.0, 1.0) + samples = (samples * 32767.0).to(torch.int16) + + frame_in = av_module.AudioFrame.from_ndarray( + samples.contiguous().reshape(1, -1).cpu().numpy(), + format="s16", + layout="stereo", + ) + frame_in.sample_rate = audio_sample_rate + + _resample_audio(container, audio_stream, frame_in, av_module) + + +def encode_video( + video: list[PIL.Image.Image] | np.ndarray | "torch.Tensor" | Iterator["torch.Tensor"], + fps: int, + output_path: str, + audio: "torch.Tensor" | None = None, + audio_sample_rate: int | None = None, + video_chunks_number: int = 1, +) -> None: + """ + Encodes a video with optional audio using the PyAV library. Based on code from the original LTX-2 repo: + https://github.com/Lightricks/LTX-2/blob/4f410820b198e05074a1e92de793e3b59e9ab5a0/packages/ltx-pipelines/src/ltx_pipelines/utils/media_io.py#L182 + + Args: + video (`List[PIL.Image.Image]` or `np.ndarray` or `torch.Tensor`): + A video tensor of shape [frames, height, width, channels] with integer pixel values in [0, 255]. If the + input is a `np.ndarray`, it is expected to be a float array with values in [0, 1] (which is what pipelines + usually return with `output_type="np"`). + fps (`int`) + The frames per second (FPS) of the encoded video. + output_path (`str`): + The path to save the encoded video to. + audio (`torch.Tensor`, *optional*): + An audio waveform of shape [audio_channels, samples]. + audio_sample_rate: (`int`, *optional*): + The sampling rate of the audio waveform. + video_chunks_number (`int`, *optional*, defaults to `1`): + The number of chunks to split the video into for encoding. Each chunk will be encoded separately. The + number of chunks to use often depends on the tiling config for the video VAE. + """ + av = _import_av() + import torch + from tqdm import tqdm + + if isinstance(video, list) and isinstance(video[0], PIL.Image.Image): + # Pipeline output_type="pil"; assumes each image is in "RGB" mode. + video_frames = [np.array(frame) for frame in video] + video = np.stack(video_frames, axis=0) + video = torch.from_numpy(video) + elif isinstance(video, np.ndarray): + # Pipeline output_type="np". + is_denormalized = np.logical_and(np.zeros_like(video) <= video, video <= np.ones_like(video)) + if np.all(is_denormalized): + video = (video * 255).round().astype("uint8") + else: + logger.warning( + "Supplied `numpy.ndarray` does not have values in [0, 1]. The values will be assumed to be pixel " + "values in [0, ..., 255] and will be used as is." + ) + video = torch.from_numpy(video) + + if isinstance(video, torch.Tensor): + # Split into video_chunks_number along the frame dimension. + video = torch.tensor_split(video, video_chunks_number, dim=0) + video = iter(video) + + first_chunk = next(video) + + _, height, width, _ = first_chunk.shape + + container = av.open(output_path, mode="w") + stream = container.add_stream("libx264", rate=int(fps)) + stream.width = width + stream.height = height + stream.pix_fmt = "yuv420p" + + if audio is not None: + if audio_sample_rate is None: + raise ValueError("audio_sample_rate is required when audio is provided") + + audio_stream = _prepare_audio_stream(container, audio_sample_rate) + + for video_chunk in tqdm(chain([first_chunk], video), total=video_chunks_number, desc="Encoding video chunks"): + video_chunk_cpu = video_chunk.to("cpu").numpy() + for frame_array in video_chunk_cpu: + frame = av.VideoFrame.from_ndarray(frame_array, format="rgb24") + for packet in stream.encode(frame): + container.mux(packet) + + # Flush video encoder. + for packet in stream.encode(): + container.mux(packet) + + if audio is not None: + _write_audio(container, audio_stream, audio, audio_sample_rate, av) + + container.close() diff --git a/diffusers/utils/hub_utils.py b/diffusers/utils/hub_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..97f27976a7b7fabf54057c85031d7d99954c3b8c --- /dev/null +++ b/diffusers/utils/hub_utils.py @@ -0,0 +1,593 @@ +# coding=utf-8 +# Copyright 2025 The HuggingFace Inc. team. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import json +import os +import re +import sys +import tempfile +import warnings +from pathlib import Path +from uuid import uuid4 + +from huggingface_hub import ( + DDUFEntry, + ModelCard, + ModelCardData, + create_repo, + hf_hub_download, + model_info, + snapshot_download, + upload_folder, +) +from huggingface_hub.constants import HF_HUB_DISABLE_TELEMETRY, HF_HUB_OFFLINE +from huggingface_hub.file_download import REGEX_COMMIT_HASH +from huggingface_hub.utils import ( + EntryNotFoundError, + HfHubHTTPError, + RepositoryNotFoundError, + RevisionNotFoundError, + is_jinja_available, + validate_hf_hub_args, +) +from packaging import version + +from .. import __version__ +from .constants import ( + DEPRECATED_REVISION_ARGS, + HUGGINGFACE_CO_RESOLVE_ENDPOINT, + SAFETENSORS_WEIGHTS_NAME, + WEIGHTS_NAME, +) +from .import_utils import ( + ENV_VARS_TRUE_VALUES, + _onnxruntime_version, + _torch_version, + is_onnx_available, + is_torch_available, +) +from .logging import get_logger + + +logger = get_logger(__name__) + +MODEL_CARD_TEMPLATE_PATH = Path(__file__).parent / "model_card_template.md" +SESSION_ID = uuid4().hex + + +def http_user_agent(user_agent: dict | str | None = None) -> str: + """ + Formats a user-agent string with basic info about a request. + """ + ua = f"diffusers/{__version__}; python/{sys.version.split()[0]}; session_id/{SESSION_ID}" + if HF_HUB_DISABLE_TELEMETRY or HF_HUB_OFFLINE: + return ua + "; telemetry/off" + if is_torch_available(): + ua += f"; torch/{_torch_version}" + if is_onnx_available(): + ua += f"; onnxruntime/{_onnxruntime_version}" + # CI will set this value to True + if os.environ.get("DIFFUSERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES: + ua += "; is_ci/true" + if isinstance(user_agent, dict): + ua += "; " + "; ".join(f"{k}/{v}" for k, v in user_agent.items()) + elif isinstance(user_agent, str): + ua += "; " + user_agent + return ua + + +def load_or_create_model_card( + repo_id_or_path: str = None, + token: str | None = None, + is_pipeline: bool = False, + from_training: bool = False, + model_description: str | None = None, + base_model: str = None, + prompt: str | None = None, + license: str | None = None, + widget: list[dict] | None = None, + inference: bool | None = None, + is_modular: bool = False, + update_model_card: bool = False, +) -> ModelCard: + """ + Loads or creates a model card. + + Args: + repo_id_or_path (`str`): + The repo id (e.g., "stable-diffusion-v1-5/stable-diffusion-v1-5") or local path where to look for the model + card. + token (`str`, *optional*): + Authentication token. Will default to the stored token. See https://huggingface.co/settings/token for more + details. + is_pipeline (`bool`): + Boolean to indicate if we're adding tag to a [`DiffusionPipeline`]. + from_training: (`bool`): Boolean flag to denote if the model card is being created from a training script. + model_description (`str`, *optional*): Model description to add to the model card. Helpful when using + `load_or_create_model_card` from a training script. + base_model (`str`): Base model identifier (e.g., "stabilityai/stable-diffusion-xl-base-1.0"). Useful + for DreamBooth-like training. + prompt (`str`, *optional*): Prompt used for training. Useful for DreamBooth-like training. + license: (`str`, *optional*): License of the output artifact. Helpful when using + `load_or_create_model_card` from a training script. + widget (`list[dict]`, *optional*): Widget to accompany a gallery template. + inference: (`bool`, optional): Whether to turn on inference widget. Helpful when using + `load_or_create_model_card` from a training script. + is_modular: (`bool`, optional): Boolean flag to denote if the model card is for a modular pipeline. + When True, uses model_description as-is without additional template formatting. + update_model_card: (`bool`, optional): When True, regenerates the model card content even if one + already exists on the remote repo. Existing card metadata (tags, license, etc.) is preserved. Only + supported for modular pipelines (i.e., `is_modular=True`). + """ + if not is_jinja_available(): + raise ValueError( + "Modelcard rendering is based on Jinja templates." + " Please make sure to have `jinja` installed before using `load_or_create_model_card`." + " To install it, please run `pip install Jinja2`." + ) + + if update_model_card and not is_modular: + raise ValueError("`update_model_card=True` is only supported for modular pipelines (`is_modular=True`).") + + try: + # Check if the model card is present on the remote repo + model_card = ModelCard.load(repo_id_or_path, token=token) + # For modular pipelines, regenerate card content when requested (preserve existing metadata) + if update_model_card and is_modular and model_description is not None: + existing_data = model_card.data + model_card = ModelCard(model_description) + model_card.data = existing_data + except (EntryNotFoundError, RepositoryNotFoundError): + # Otherwise create a model card from template + if from_training: + model_card = ModelCard.from_template( + card_data=ModelCardData( # Card metadata object that will be converted to YAML block + license=license, + library_name="diffusers", + inference=inference, + base_model=base_model, + instance_prompt=prompt, + widget=widget, + ), + template_path=MODEL_CARD_TEMPLATE_PATH, + model_description=model_description, + ) + else: + card_data = ModelCardData() + if is_modular and model_description is not None: + model_card = ModelCard(model_description) + model_card.data = card_data + else: + component = "pipeline" if is_pipeline else "model" + if model_description is None: + model_description = f"This is the model card of a 🧨 diffusers {component} that has been pushed on the Hub. This model card has been automatically generated." + model_card = ModelCard.from_template(card_data, model_description=model_description) + + return model_card + + +def populate_model_card(model_card: ModelCard, tags: str | list[str] | None = None) -> ModelCard: + """Populates the `model_card` with library name and optional tags.""" + if model_card.data.library_name is None: + model_card.data.library_name = "diffusers" + + if tags is not None: + if isinstance(tags, str): + tags = [tags] + if model_card.data.tags is None: + model_card.data.tags = [] + for tag in tags: + model_card.data.tags.append(tag) + + return model_card + + +def extract_commit_hash(resolved_file: str | None, commit_hash: str | None = None): + """ + Extracts the commit hash from a resolved filename toward a cache file. + """ + if resolved_file is None or commit_hash is not None: + return commit_hash + resolved_file = str(Path(resolved_file).as_posix()) + search = re.search(r"snapshots/([^/]+)/", resolved_file) + if search is None: + return None + commit_hash = search.groups()[0] + return commit_hash if REGEX_COMMIT_HASH.match(commit_hash) else None + + +def _add_variant(weights_name: str, variant: str | None = None) -> str: + if variant is not None: + splits = weights_name.split(".") + splits = splits[:-1] + [variant] + splits[-1:] + weights_name = ".".join(splits) + + return weights_name + + +@validate_hf_hub_args +def _get_model_file( + pretrained_model_name_or_path: str | Path, + *, + weights_name: str, + subfolder: str | None = None, + cache_dir: str | None = None, + force_download: bool = False, + proxies: dict | None = None, + local_files_only: bool = False, + token: str | None = None, + user_agent: dict | str | None = None, + revision: str | None = None, + commit_hash: str | None = None, + dduf_entries: dict[str, DDUFEntry] | None = None, +): + pretrained_model_name_or_path = str(pretrained_model_name_or_path) + + if dduf_entries: + if subfolder is not None: + raise ValueError( + "DDUF file only allow for 1 level of directory (e.g transformer/model1/model.safetentors is not allowed). " + "Please check the DDUF structure" + ) + model_file = ( + weights_name + if pretrained_model_name_or_path == "" + else "/".join([pretrained_model_name_or_path, weights_name]) + ) + if model_file in dduf_entries: + return model_file + else: + raise EnvironmentError(f"Error no file named {weights_name} found in archive {dduf_entries.keys()}.") + elif os.path.isfile(pretrained_model_name_or_path): + return pretrained_model_name_or_path + elif os.path.isdir(pretrained_model_name_or_path): + if os.path.isfile(os.path.join(pretrained_model_name_or_path, weights_name)): + # Load from a PyTorch checkpoint + model_file = os.path.join(pretrained_model_name_or_path, weights_name) + return model_file + elif subfolder is not None and os.path.isfile( + os.path.join(pretrained_model_name_or_path, subfolder, weights_name) + ): + model_file = os.path.join(pretrained_model_name_or_path, subfolder, weights_name) + return model_file + else: + raise EnvironmentError( + f"Error no file named {weights_name} found in directory {pretrained_model_name_or_path}." + ) + else: + # 1. First check if deprecated way of loading from branches is used + if ( + revision in DEPRECATED_REVISION_ARGS + and (weights_name == WEIGHTS_NAME or weights_name == SAFETENSORS_WEIGHTS_NAME) + and version.parse(version.parse(__version__).base_version) >= version.parse("0.22.0") + ): + try: + model_file = hf_hub_download( + pretrained_model_name_or_path, + filename=_add_variant(weights_name, revision), + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + user_agent=user_agent, + subfolder=subfolder, + revision=revision or commit_hash, + ) + warnings.warn( + f"Loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` is deprecated. Loading instead from `revision='main'` with `variant={revision}`. Loading model variants via `revision='{revision}'` will be removed in diffusers v1. Please use `variant='{revision}'` instead.", + FutureWarning, + ) + return model_file + except: # noqa: E722 + warnings.warn( + f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have a {_add_variant(weights_name, revision)} file in the 'main' branch of {pretrained_model_name_or_path}. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {_add_variant(weights_name, revision)}' so that the correct variant file can be added.", + FutureWarning, + ) + try: + # 2. Load model file as usual + model_file = hf_hub_download( + pretrained_model_name_or_path, + filename=weights_name, + cache_dir=cache_dir, + force_download=force_download, + proxies=proxies, + local_files_only=local_files_only, + token=token, + user_agent=user_agent, + subfolder=subfolder, + revision=revision or commit_hash, + ) + return model_file + + except RepositoryNotFoundError as e: + raise EnvironmentError( + f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier " + "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a " + "token having permission to this repo with `token` or log in with `hf auth login`." + ) from e + except RevisionNotFoundError as e: + raise EnvironmentError( + f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for " + "this model name. Check the model page at " + f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions." + ) from e + except EntryNotFoundError as e: + raise EnvironmentError( + f"{pretrained_model_name_or_path} does not appear to have a file named {weights_name}." + ) from e + except HfHubHTTPError as e: + raise EnvironmentError( + f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n{e}" + ) from e + except ValueError as e: + raise EnvironmentError( + f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it" + f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a" + f" directory containing a file named {weights_name} or" + " \nCheckout your internet connection or see how to run the library in" + " offline mode at 'https://huggingface.co/docs/diffusers/installation#offline-mode'." + ) from e + except EnvironmentError as e: + raise EnvironmentError( + f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from " + "'https://huggingface.co/models', make sure you don't have a local directory with the same name. " + f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory " + f"containing a file named {weights_name}" + ) from e + + +def _get_checkpoint_shard_files( + pretrained_model_name_or_path, + index_filename, + cache_dir=None, + proxies=None, + local_files_only=False, + token=None, + user_agent=None, + revision=None, + subfolder="", + dduf_entries: dict[str, DDUFEntry] | None = None, +): + """ + For a given model: + + - download and cache all the shards of a sharded checkpoint if `pretrained_model_name_or_path` is a model ID on the + Hub + - returns the list of paths to all the shards, as well as some metadata. + + For the description of each arg, see [`PreTrainedModel.from_pretrained`]. `index_filename` is the full path to the + index (downloaded and cached if `pretrained_model_name_or_path` is a model ID on the Hub). + """ + if dduf_entries: + if index_filename not in dduf_entries: + raise ValueError(f"Can't find a checkpoint index ({index_filename}) in {pretrained_model_name_or_path}.") + else: + if not os.path.isfile(index_filename): + raise ValueError(f"Can't find a checkpoint index ({index_filename}) in {pretrained_model_name_or_path}.") + + if dduf_entries: + index = json.loads(dduf_entries[index_filename].read_text()) + else: + with open(index_filename, "r") as f: + index = json.loads(f.read()) + + original_shard_filenames = sorted(set(index["weight_map"].values())) + for shard_filename in original_shard_filenames: + if os.path.basename(shard_filename) != shard_filename: + raise ValueError( + f"The shard filename {shard_filename!r} in the checkpoint index contains a path separator or a " + "parent-directory reference, which is not allowed. Shard filenames must be plain filenames located " + "in the model directory." + ) + sharded_metadata = index["metadata"] + sharded_metadata["all_checkpoint_keys"] = list(index["weight_map"].keys()) + sharded_metadata["weight_map"] = index["weight_map"].copy() + shards_path = os.path.join(pretrained_model_name_or_path, subfolder) + + # First, let's deal with local folder. + if os.path.isdir(pretrained_model_name_or_path) or dduf_entries: + shard_filenames = [os.path.join(shards_path, f) for f in original_shard_filenames] + for shard_file in shard_filenames: + if dduf_entries: + if shard_file not in dduf_entries: + raise FileNotFoundError( + f"{shards_path} does not appear to have a file named {shard_file} which is " + "required according to the checkpoint index." + ) + else: + if not os.path.exists(shard_file): + raise FileNotFoundError( + f"{shards_path} does not appear to have a file named {shard_file} which is " + "required according to the checkpoint index." + ) + return shard_filenames, sharded_metadata + + # At this stage pretrained_model_name_or_path is a model identifier on the Hub + allow_patterns = original_shard_filenames + if subfolder is not None: + allow_patterns = [os.path.join(subfolder, p) for p in allow_patterns] + + ignore_patterns = ["*.json", "*.md"] + + # If the repo doesn't have the required shards, error out early even before downloading anything. + if not local_files_only: + model_files_info = model_info(pretrained_model_name_or_path, revision=revision, token=token) + for shard_file in original_shard_filenames: + shard_file_present = any(shard_file in k.rfilename for k in model_files_info.siblings) + if not shard_file_present: + raise EnvironmentError( + f"{shards_path} does not appear to have a file named {shard_file} which is " + "required according to the checkpoint index." + ) + + try: + # Load from URL + cached_folder = snapshot_download( + pretrained_model_name_or_path, + cache_dir=cache_dir, + proxies=proxies, + local_files_only=local_files_only, + token=token, + revision=revision, + allow_patterns=allow_patterns, + ignore_patterns=ignore_patterns, + user_agent=user_agent, + ) + if subfolder is not None: + cached_folder = os.path.join(cached_folder, subfolder) + + # We have already dealt with RepositoryNotFoundError and RevisionNotFoundError when getting the index, so + # we don't have to catch them here. We have also dealt with EntryNotFoundError. + except HfHubHTTPError as e: + raise EnvironmentError( + f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load {pretrained_model_name_or_path}. You should try" + " again after checking your internet connection." + ) from e + + cached_filenames = [os.path.join(cached_folder, f) for f in original_shard_filenames] + for cached_file in cached_filenames: + if not os.path.isfile(cached_file): + raise EnvironmentError( + f"{cached_folder} does not have a file named {cached_file} which is required according to the checkpoint index." + ) + + return cached_filenames, sharded_metadata + + +def _check_legacy_sharding_variant_format(folder: str = None, filenames: list[str] = None, variant: str = None): + if filenames and folder: + raise ValueError("Both `filenames` and `folder` cannot be provided.") + if not filenames: + filenames = [] + for _, _, files in os.walk(folder): + for file in files: + filenames.append(os.path.basename(file)) + transformers_index_format = r"\d{5}-of-\d{5}" + variant_file_re = re.compile(rf".*-{transformers_index_format}\.{variant}\.[a-z]+$") + return any(variant_file_re.match(f) is not None for f in filenames) + + +class PushToHubMixin: + """ + A Mixin to push a model, scheduler, or pipeline to the Hugging Face Hub. + """ + + def _upload_folder( + self, + working_dir: str | os.PathLike, + repo_id: str, + token: str | None = None, + commit_message: str | None = None, + create_pr: bool = False, + subfolder: str | None = None, + ): + """ + Uploads all files in `working_dir` to `repo_id`. + """ + if commit_message is None: + if "Model" in self.__class__.__name__: + commit_message = "Upload model" + elif "Scheduler" in self.__class__.__name__: + commit_message = "Upload scheduler" + else: + commit_message = f"Upload {self.__class__.__name__}" + + logger.info(f"Uploading the files of {working_dir} to {repo_id}.") + return upload_folder( + repo_id=repo_id, + folder_path=working_dir, + token=token, + commit_message=commit_message, + create_pr=create_pr, + path_in_repo=subfolder, + ) + + def push_to_hub( + self, + repo_id: str, + commit_message: str | None = None, + private: bool | None = None, + token: str | None = None, + create_pr: bool = False, + safe_serialization: bool = True, + variant: str | None = None, + subfolder: str | None = None, + ) -> str: + """ + Upload model, scheduler, or pipeline files to the 🤗 Hugging Face Hub. + + Parameters: + repo_id (`str`): + The name of the repository you want to push your model, scheduler, or pipeline files to. It should + contain your organization name when pushing to an organization. `repo_id` can also be a path to a local + directory. + commit_message (`str`, *optional*): + Message to commit while pushing. Default to `"Upload {object}"`. + private (`bool`, *optional*): + Whether to make the repo private. If `None` (default), the repo will be public unless the + organization's default is private. This value is ignored if the repo already exists. + token (`str`, *optional*): + The token to use as HTTP bearer authorization for remote files. The token generated when running `hf + auth login` (stored in `~/.huggingface`). + create_pr (`bool`, *optional*, defaults to `False`): + Whether or not to create a PR with the uploaded files or directly commit. + safe_serialization (`bool`, *optional*, defaults to `True`): + Whether or not to convert the model weights to the `safetensors` format. + variant (`str`, *optional*): + If specified, weights are saved in the format `pytorch_model..bin`. + + Examples: + + ```python + from diffusers import UNet2DConditionModel + + unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet") + + # Push the `unet` to your namespace with the name "my-finetuned-unet". + unet.push_to_hub("my-finetuned-unet") + + # Push the `unet` to an organization with the name "my-finetuned-unet". + unet.push_to_hub("your-org/my-finetuned-unet") + ``` + """ + repo_id = create_repo(repo_id, private=private, token=token, exist_ok=True).repo_id + + # Create a new empty model card and eventually tag it + if not subfolder: + model_card = load_or_create_model_card(repo_id, token=token) + model_card = populate_model_card(model_card) + + # Save all files. + save_kwargs = {"safe_serialization": safe_serialization} + if "Scheduler" not in self.__class__.__name__: + save_kwargs.update({"variant": variant}) + + with tempfile.TemporaryDirectory() as tmpdir: + self.save_pretrained(tmpdir, **save_kwargs) + + # Update model card if needed: + if not subfolder: + model_card.save(os.path.join(tmpdir, "README.md")) + + return self._upload_folder( + tmpdir, + repo_id, + token=token, + commit_message=commit_message, + create_pr=create_pr, + subfolder=subfolder, + ) diff --git a/diffusers/utils/import_utils.py b/diffusers/utils/import_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..639ed59d63931b7545605e9b9f891327c80e64c3 --- /dev/null +++ b/diffusers/utils/import_utils.py @@ -0,0 +1,1044 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Import utilities: Utilities related to imports and our lazy inits. +""" + +import importlib.util +import inspect +import operator as op +import os +import sys +from collections import OrderedDict, defaultdict +from functools import lru_cache as cache +from itertools import chain +from types import ModuleType +from typing import Any + +from huggingface_hub.utils import is_jinja_available # noqa: F401 +from packaging.version import Version, parse + +from . import logging + + +# The package importlib_metadata is in a different place, depending on the python version. +if sys.version_info < (3, 8): + import importlib_metadata +else: + import importlib.metadata as importlib_metadata +try: + _package_map = importlib_metadata.packages_distributions() # load-once to avoid expensive calls +except Exception: + _package_map = None + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +ENV_VARS_TRUE_VALUES = {"1", "ON", "YES", "TRUE"} +ENV_VARS_TRUE_AND_AUTO_VALUES = ENV_VARS_TRUE_VALUES.union({"AUTO"}) + +USE_TF = os.environ.get("USE_TF", "AUTO").upper() +USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper() +USE_SAFETENSORS = os.environ.get("USE_SAFETENSORS", "AUTO").upper() +DIFFUSERS_SLOW_IMPORT = os.environ.get("DIFFUSERS_SLOW_IMPORT", "FALSE").upper() +DIFFUSERS_SLOW_IMPORT = DIFFUSERS_SLOW_IMPORT in ENV_VARS_TRUE_VALUES + +STR_OPERATION_TO_FUNC = {">": op.gt, ">=": op.ge, "==": op.eq, "!=": op.ne, "<=": op.le, "<": op.lt} + +_is_google_colab = "google.colab" in sys.modules or any(k.startswith("COLAB_") for k in os.environ) + + +def _is_package_available(pkg_name: str, get_dist_name: bool = False) -> tuple[bool, str]: + global _package_map + pkg_exists = importlib.util.find_spec(pkg_name) is not None + pkg_version = "N/A" + + if pkg_exists: + if _package_map is None: + _package_map = defaultdict(list) + try: + # Fallback for Python < 3.10 + for dist in importlib_metadata.distributions(): + _top_level_declared = (dist.read_text("top_level.txt") or "").split() + # Infer top-level package names from file structure + _inferred_opt_names = { + f.parts[0] if len(f.parts) > 1 else inspect.getmodulename(f) for f in (dist.files or []) + } - {None} + _top_level_inferred = filter(lambda name: "." not in name, _inferred_opt_names) + for pkg in _top_level_declared or _top_level_inferred: + _package_map[pkg].append(dist.metadata["Name"]) + except Exception as _: + pass + try: + if get_dist_name and pkg_name in _package_map and _package_map[pkg_name]: + if len(_package_map[pkg_name]) > 1: + logger.warning( + f"Multiple distributions found for package {pkg_name}. Picked distribution: {_package_map[pkg_name][0]}" + ) + pkg_name = _package_map[pkg_name][0] + pkg_version = importlib_metadata.version(pkg_name) + logger.debug(f"Successfully imported {pkg_name} version {pkg_version}") + except (ImportError, importlib_metadata.PackageNotFoundError): + pkg_exists = False + + return pkg_exists, pkg_version + + +if USE_TORCH in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TF not in ENV_VARS_TRUE_VALUES: + _torch_available, _torch_version = _is_package_available("torch") + +else: + logger.info("Disabling PyTorch because USE_TORCH is set") + _torch_available = False + _torch_version = "N/A" + +if USE_SAFETENSORS in ENV_VARS_TRUE_AND_AUTO_VALUES: + _safetensors_available, _safetensors_version = _is_package_available("safetensors") + +else: + logger.info("Disabling Safetensors because USE_SAFETENSORS is set") + _safetensors_available = False + +_onnxruntime_version = "N/A" +_onnx_available = importlib.util.find_spec("onnxruntime") is not None +if _onnx_available: + candidates = ( + "onnxruntime", + "onnxruntime-cann", + "onnxruntime-directml", + "ort_nightly_directml", + "onnxruntime-gpu", + "ort_nightly_gpu", + "onnxruntime-migraphx", + "onnxruntime-openvino", + "onnxruntime-qnn", + "onnxruntime-rocm", + "onnxruntime-training", + "onnxruntime-vitisai", + ) + _onnxruntime_version = None + # For the metadata, we have to look for both onnxruntime and onnxruntime-x + for pkg in candidates: + try: + _onnxruntime_version = importlib_metadata.version(pkg) + break + except importlib_metadata.PackageNotFoundError: + pass + _onnx_available = _onnxruntime_version is not None + if _onnx_available: + logger.debug(f"Successfully imported onnxruntime version {_onnxruntime_version}") + +# (sayakpaul): importlib.util.find_spec("opencv-python") returns None even when it's installed. +# _opencv_available = importlib.util.find_spec("opencv-python") is not None +try: + candidates = ( + "opencv-python", + "opencv-contrib-python", + "opencv-python-headless", + "opencv-contrib-python-headless", + ) + _opencv_version = None + for pkg in candidates: + try: + _opencv_version = importlib_metadata.version(pkg) + break + except importlib_metadata.PackageNotFoundError: + pass + _opencv_available = _opencv_version is not None + if _opencv_available: + logger.debug(f"Successfully imported cv2 version {_opencv_version}") +except importlib_metadata.PackageNotFoundError: + _opencv_available = False + +_bs4_available = importlib.util.find_spec("bs4") is not None +try: + # importlib metadata under different name + _bs4_version = importlib_metadata.version("beautifulsoup4") + logger.debug(f"Successfully imported ftfy version {_bs4_version}") +except importlib_metadata.PackageNotFoundError: + _bs4_available = False + +_invisible_watermark_available = importlib.util.find_spec("imwatermark") is not None +try: + _invisible_watermark_version = importlib_metadata.version("invisible-watermark") + logger.debug(f"Successfully imported invisible-watermark version {_invisible_watermark_version}") +except importlib_metadata.PackageNotFoundError: + _invisible_watermark_available = False + +_torch_xla_available, _torch_xla_version = _is_package_available("torch_xla") +_torch_npu_available, _torch_npu_version = _is_package_available("torch_npu") +_torch_mlu_available, _torch_mlu_version = _is_package_available("torch_mlu") +_torch_neuronx_available, _torch_neuronx_version = _is_package_available("torch_neuronx") +_transformers_available, _transformers_version = _is_package_available("transformers") +_hf_hub_available, _hf_hub_version = _is_package_available("huggingface_hub") +_kernels_available, _kernels_version = _is_package_available("kernels") +_inflect_available, _inflect_version = _is_package_available("inflect") +_unidecode_available, _unidecode_version = _is_package_available("unidecode") + +_note_seq_available, _note_seq_version = _is_package_available("note_seq") +_wandb_available, _wandb_version = _is_package_available("wandb") +_tensorboard_available, _tensorboard_version = _is_package_available("tensorboard") +_compel_available, _compel_version = _is_package_available("compel") +_sentencepiece_available, _sentencepiece_version = _is_package_available("sentencepiece") +_outlines_available, _outlines_version = _is_package_available("outlines") +_torchsde_available, _torchsde_version = _is_package_available("torchsde") +_peft_available, _peft_version = _is_package_available("peft") +_torchvision_available, _torchvision_version = _is_package_available("torchvision") +_matplotlib_available, _matplotlib_version = _is_package_available("matplotlib") +_timm_available, _timm_version = _is_package_available("timm") +_bitsandbytes_available, _bitsandbytes_version = _is_package_available("bitsandbytes") +_imageio_available, _imageio_version = _is_package_available("imageio") +_ftfy_available, _ftfy_version = _is_package_available("ftfy") +_scipy_available, _scipy_version = _is_package_available("scipy") +_librosa_available, _librosa_version = _is_package_available("librosa") +_accelerate_available, _accelerate_version = _is_package_available("accelerate") +_xformers_available, _xformers_version = _is_package_available("xformers") +_gguf_available, _gguf_version = _is_package_available("gguf") +_torchao_available, _torchao_version = _is_package_available("torchao") +_bitsandbytes_available, _bitsandbytes_version = _is_package_available("bitsandbytes") +_optimum_quanto_available, _optimum_quanto_version = _is_package_available("optimum", get_dist_name=True) +_pytorch_retinaface_available, _pytorch_retinaface_version = _is_package_available("pytorch_retinaface") +_better_profanity_available, _better_profanity_version = _is_package_available("better_profanity") +_nltk_available, _nltk_version = _is_package_available("nltk") +_cosmos_guardrail_available, _cosmos_guardrail_version = _is_package_available("cosmos_guardrail") +_sageattention_available, _sageattention_version = _is_package_available("sageattention") +_flash_attn_available, _flash_attn_version = _is_package_available("flash_attn") +_flash_attn_3_available, _flash_attn_3_version = _is_package_available("flash_attn_3") +_aiter_available, _aiter_version = _is_package_available("aiter", get_dist_name=True) +_kornia_available, _kornia_version = _is_package_available("kornia") +_nvidia_modelopt_available, _nvidia_modelopt_version = _is_package_available("modelopt", get_dist_name=True) +_auto_round_available, _auto_round_version = _is_package_available("auto_round") +_sdnq_available, _sdnq_version = _is_package_available("sdnq") +_flashpack_available, _flashpack_version = _is_package_available("flashpack") +_av_available, _av_version = _is_package_available("av") + + +def is_torch_available(): + return _torch_available + + +def is_torch_xla_available(): + return _torch_xla_available + + +def is_torch_npu_available(): + return _torch_npu_available + + +def is_torch_mlu_available(): + return _torch_mlu_available + + +def is_torch_neuronx_available(): + return _torch_neuronx_available + + +def is_transformers_available(): + return _transformers_available + + +def is_inflect_available(): + return _inflect_available + + +def is_unidecode_available(): + return _unidecode_available + + +def is_onnx_available(): + return _onnx_available + + +def is_opencv_available(): + return _opencv_available + + +def is_scipy_available(): + return _scipy_available + + +def is_librosa_available(): + return _librosa_available + + +def is_xformers_available(): + return _xformers_available + + +def is_accelerate_available(): + return _accelerate_available + + +def is_kernels_available(): + return _kernels_available + + +def is_note_seq_available(): + return _note_seq_available + + +def is_wandb_available(): + return _wandb_available + + +def is_tensorboard_available(): + return _tensorboard_available + + +def is_compel_available(): + return _compel_available + + +def is_ftfy_available(): + return _ftfy_available + + +def is_bs4_available(): + return _bs4_available + + +def is_torchsde_available(): + return _torchsde_available + + +def is_invisible_watermark_available(): + return _invisible_watermark_available + + +def is_peft_available(): + return _peft_available + + +def is_torchvision_available(): + return _torchvision_available + + +def is_matplotlib_available(): + return _matplotlib_available + + +def is_safetensors_available(): + return _safetensors_available + + +def is_bitsandbytes_available(): + return _bitsandbytes_available + + +def is_google_colab(): + return _is_google_colab + + +def is_sentencepiece_available(): + return _sentencepiece_available + + +def is_outlines_available(): + return _outlines_available + + +def is_imageio_available(): + return _imageio_available + + +def is_gguf_available(): + return _gguf_available + + +def is_flashpack_available(): + return _flashpack_available + + +def is_torchao_available(): + return _torchao_available + + +def is_optimum_quanto_available(): + return _optimum_quanto_available + + +def is_nvidia_modelopt_available(): + return _nvidia_modelopt_available + + +def is_auto_round_available(): + return _auto_round_available + + +def is_sdnq_available(): + return _sdnq_available + + +def is_timm_available(): + return _timm_available + + +def is_pytorch_retinaface_available(): + return _pytorch_retinaface_available + + +def is_better_profanity_available(): + return _better_profanity_available + + +def is_nltk_available(): + return _nltk_available + + +def is_cosmos_guardrail_available(): + return _cosmos_guardrail_available + + +def is_hpu_available(): + return all(importlib.util.find_spec(lib) for lib in ("habana_frameworks", "habana_frameworks.torch")) + + +def is_sageattention_available(): + return _sageattention_available + + +def is_flash_attn_available(): + return _flash_attn_available + + +def is_flash_attn_3_available(): + return _flash_attn_3_available + + +def is_aiter_available(): + return _aiter_available + + +def is_kornia_available(): + return _kornia_available + + +def is_av_available(): + return _av_available + + +# docstyle-ignore +INFLECT_IMPORT_ERROR = """ +{0} requires the inflect library but it was not found in your environment. You can install it with pip: `pip install +inflect` +""" + +# docstyle-ignore +PYTORCH_IMPORT_ERROR = """ +{0} requires the PyTorch library but it was not found in your environment. Checkout the instructions on the +installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your environment. +""" + +# docstyle-ignore +ONNX_IMPORT_ERROR = """ +{0} requires the onnxruntime library but it was not found in your environment. You can install it with pip: `pip +install onnxruntime` +""" + +# docstyle-ignore +OPENCV_IMPORT_ERROR = """ +{0} requires the OpenCV library but it was not found in your environment. You can install it with pip: `pip +install opencv-python` +""" + +# docstyle-ignore +SCIPY_IMPORT_ERROR = """ +{0} requires the scipy library but it was not found in your environment. You can install it with pip: `pip install +scipy` +""" + +# docstyle-ignore +LIBROSA_IMPORT_ERROR = """ +{0} requires the librosa library but it was not found in your environment. Checkout the instructions on the +installation page: https://librosa.org/doc/latest/install.html and follow the ones that match your environment. +""" + +# docstyle-ignore +TRANSFORMERS_IMPORT_ERROR = """ +{0} requires the transformers library but it was not found in your environment. You can install it with pip: `pip +install transformers` +""" + +# docstyle-ignore +UNIDECODE_IMPORT_ERROR = """ +{0} requires the unidecode library but it was not found in your environment. You can install it with pip: `pip install +Unidecode` +""" + +# docstyle-ignore +NOTE_SEQ_IMPORT_ERROR = """ +{0} requires the note-seq library but it was not found in your environment. You can install it with pip: `pip +install note-seq` +""" + +# docstyle-ignore +WANDB_IMPORT_ERROR = """ +{0} requires the wandb library but it was not found in your environment. You can install it with pip: `pip +install wandb` +""" + +# docstyle-ignore +TENSORBOARD_IMPORT_ERROR = """ +{0} requires the tensorboard library but it was not found in your environment. You can install it with pip: `pip +install tensorboard` +""" + + +# docstyle-ignore +COMPEL_IMPORT_ERROR = """ +{0} requires the compel library but it was not found in your environment. You can install it with pip: `pip install compel` +""" + +# docstyle-ignore +BS4_IMPORT_ERROR = """ +{0} requires the Beautiful Soup library but it was not found in your environment. You can install it with pip: +`pip install beautifulsoup4`. Please note that you may need to restart your runtime after installation. +""" + +# docstyle-ignore +FTFY_IMPORT_ERROR = """ +{0} requires the ftfy library but it was not found in your environment. Checkout the instructions on the +installation section: https://github.com/rspeer/python-ftfy/tree/master#installing and follow the ones +that match your environment. Please note that you may need to restart your runtime after installation. +""" + +# docstyle-ignore +TORCHSDE_IMPORT_ERROR = """ +{0} requires the torchsde library but it was not found in your environment. You can install it with pip: `pip install torchsde` +""" + +# docstyle-ignore +INVISIBLE_WATERMARK_IMPORT_ERROR = """ +{0} requires the invisible-watermark library but it was not found in your environment. You can install it with pip: `pip install invisible-watermark>=0.2.0` +""" + +# docstyle-ignore +PEFT_IMPORT_ERROR = """ +{0} requires the peft library but it was not found in your environment. You can install it with pip: `pip install peft` +""" + +# docstyle-ignore +SAFETENSORS_IMPORT_ERROR = """ +{0} requires the safetensors library but it was not found in your environment. You can install it with pip: `pip install safetensors` +""" + +# docstyle-ignore +SENTENCEPIECE_IMPORT_ERROR = """ +{0} requires the sentencepiece library but it was not found in your environment. You can install it with pip: `pip install sentencepiece` +""" + + +# docstyle-ignore +BITSANDBYTES_IMPORT_ERROR = """ +{0} requires the bitsandbytes library but it was not found in your environment. You can install it with pip: `pip install bitsandbytes` +""" + +# docstyle-ignore +IMAGEIO_IMPORT_ERROR = """ +{0} requires the imageio library and ffmpeg but it was not found in your environment. You can install it with pip: `pip install imageio imageio-ffmpeg` +""" + +# docstyle-ignore +GGUF_IMPORT_ERROR = """ +{0} requires the gguf library but it was not found in your environment. You can install it with pip: `pip install gguf` +""" + +TORCHAO_IMPORT_ERROR = """ +{0} requires the torchao library but it was not found in your environment. You can install it with pip: `pip install +torchao` +""" + +QUANTO_IMPORT_ERROR = """ +{0} requires the optimum-quanto library but it was not found in your environment. You can install it with pip: `pip +install optimum-quanto` +""" + +SDNQ_IMPORT_ERROR = """ +{0} requires the sdnq library but it was not found in your environment. You can install it with pip: `pip install sdnq` +""" + +# docstyle-ignore +PYTORCH_RETINAFACE_IMPORT_ERROR = """ +{0} requires the pytorch_retinaface library but it was not found in your environment. You can install it with pip: `pip install pytorch_retinaface` +""" + +# docstyle-ignore +BETTER_PROFANITY_IMPORT_ERROR = """ +{0} requires the better_profanity library but it was not found in your environment. You can install it with pip: `pip install better_profanity` +""" + +# docstyle-ignore +NLTK_IMPORT_ERROR = """ +{0} requires the nltk library but it was not found in your environment. You can install it with pip: `pip install nltk` +""" + + +TORCH_NEURONX_IMPORT_ERROR = """ +{0} requires the torch_neuronx library (AWS Neuron SDK) but it was not found in your environment. Please install it +following the AWS Neuron documentation: https://awsdocs-neuron.readthedocs-hosted.com/en/latest/ +""" + +BACKENDS_MAPPING = OrderedDict( + [ + ("bs4", (is_bs4_available, BS4_IMPORT_ERROR)), + ("inflect", (is_inflect_available, INFLECT_IMPORT_ERROR)), + ("onnx", (is_onnx_available, ONNX_IMPORT_ERROR)), + ("opencv", (is_opencv_available, OPENCV_IMPORT_ERROR)), + ("scipy", (is_scipy_available, SCIPY_IMPORT_ERROR)), + ("torch", (is_torch_available, PYTORCH_IMPORT_ERROR)), + ("transformers", (is_transformers_available, TRANSFORMERS_IMPORT_ERROR)), + ("unidecode", (is_unidecode_available, UNIDECODE_IMPORT_ERROR)), + ("librosa", (is_librosa_available, LIBROSA_IMPORT_ERROR)), + ("note_seq", (is_note_seq_available, NOTE_SEQ_IMPORT_ERROR)), + ("wandb", (is_wandb_available, WANDB_IMPORT_ERROR)), + ("tensorboard", (is_tensorboard_available, TENSORBOARD_IMPORT_ERROR)), + ("compel", (is_compel_available, COMPEL_IMPORT_ERROR)), + ("ftfy", (is_ftfy_available, FTFY_IMPORT_ERROR)), + ("torchsde", (is_torchsde_available, TORCHSDE_IMPORT_ERROR)), + ("invisible_watermark", (is_invisible_watermark_available, INVISIBLE_WATERMARK_IMPORT_ERROR)), + ("peft", (is_peft_available, PEFT_IMPORT_ERROR)), + ("safetensors", (is_safetensors_available, SAFETENSORS_IMPORT_ERROR)), + ("bitsandbytes", (is_bitsandbytes_available, BITSANDBYTES_IMPORT_ERROR)), + ("sentencepiece", (is_sentencepiece_available, SENTENCEPIECE_IMPORT_ERROR)), + ("imageio", (is_imageio_available, IMAGEIO_IMPORT_ERROR)), + ("gguf", (is_gguf_available, GGUF_IMPORT_ERROR)), + ("torchao", (is_torchao_available, TORCHAO_IMPORT_ERROR)), + ("quanto", (is_optimum_quanto_available, QUANTO_IMPORT_ERROR)), + ("sdnq", (is_sdnq_available, SDNQ_IMPORT_ERROR)), + ("pytorch_retinaface", (is_pytorch_retinaface_available, PYTORCH_RETINAFACE_IMPORT_ERROR)), + ("better_profanity", (is_better_profanity_available, BETTER_PROFANITY_IMPORT_ERROR)), + ("nltk", (is_nltk_available, NLTK_IMPORT_ERROR)), + ("torch_neuronx", (is_torch_neuronx_available, TORCH_NEURONX_IMPORT_ERROR)), + ] +) + + +def requires_backends(obj, backends): + if not isinstance(backends, (list, tuple)): + backends = [backends] + + name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__ + checks = (BACKENDS_MAPPING[backend] for backend in backends) + failed = [msg.format(name) for available, msg in checks if not available()] + if failed: + raise ImportError("".join(failed)) + + if name in [ + "VersatileDiffusionTextToImagePipeline", + "VersatileDiffusionPipeline", + "VersatileDiffusionDualGuidedPipeline", + "StableDiffusionImageVariationPipeline", + "UnCLIPPipeline", + ] and is_transformers_version("<", "4.25.0"): + raise ImportError( + f"You need to install `transformers>=4.25` in order to use {name}: \n```\n pip install" + " --upgrade transformers \n```" + ) + + if name in ["StableDiffusionDepth2ImgPipeline", "StableDiffusionPix2PixZeroPipeline"] and is_transformers_version( + "<", "4.26.0" + ): + raise ImportError( + f"You need to install `transformers>=4.26` in order to use {name}: \n```\n pip install" + " --upgrade transformers \n```" + ) + + +class DummyObject(type): + """ + Metaclass for the dummy objects. Any class inheriting from it will return the ImportError generated by + `requires_backend` each time a user tries to access any method of that class. + """ + + def __getattr__(cls, key): + if key.startswith("_") and key not in ["_load_connected_pipes", "_is_onnx"]: + return super().__getattr__(cls, key) + requires_backends(cls, cls._backends) + + +# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L319 +def compare_versions(library_or_version: str | Version, operation: str, requirement_version: str): + """ + Compares a library version to some requirement using a given operation. + + Args: + library_or_version (`str` or `packaging.version.Version`): + A library name or a version to check. + operation (`str`): + A string representation of an operator, such as `">"` or `"<="`. + requirement_version (`str`): + The version to compare the library version against + """ + if operation not in STR_OPERATION_TO_FUNC.keys(): + raise ValueError(f"`operation` must be one of {list(STR_OPERATION_TO_FUNC.keys())}, received {operation}") + operation = STR_OPERATION_TO_FUNC[operation] + if isinstance(library_or_version, str): + library_or_version = parse(importlib_metadata.version(library_or_version)) + return operation(library_or_version, parse(requirement_version)) + + +# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L338 +@cache +def is_torch_version(operation: str, version: str): + """ + Compares the current PyTorch version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A string version of PyTorch + """ + return compare_versions(parse(_torch_version), operation, version) + + +@cache +def is_torch_xla_version(operation: str, version: str): + """ + Compares the current torch_xla version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A string version of torch_xla + """ + if not is_torch_xla_available: + return False + return compare_versions(parse(_torch_xla_version), operation, version) + + +@cache +def is_transformers_version(operation: str, version: str): + """ + Compares the current Transformers version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _transformers_available: + return False + return compare_versions(parse(_transformers_version), operation, version) + + +@cache +def is_kernels_version(operation: str, version: str): + """ + Compares the current Kernels version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _kernels_available: + return False + return compare_versions(parse(_kernels_version), operation, version) + + +@cache +def is_hf_hub_version(operation: str, version: str): + """ + Compares the current Hugging Face Hub version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _hf_hub_available: + return False + return compare_versions(parse(_hf_hub_version), operation, version) + + +@cache +def is_accelerate_version(operation: str, version: str): + """ + Compares the current Accelerate version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _accelerate_available: + return False + return compare_versions(parse(_accelerate_version), operation, version) + + +@cache +def is_peft_version(operation: str, version: str): + """ + Compares the current PEFT version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _peft_available: + return False + return compare_versions(parse(_peft_version), operation, version) + + +@cache +def is_bitsandbytes_version(operation: str, version: str): + """ + Args: + Compares the current bitsandbytes version to a given reference with an operation. + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _bitsandbytes_available: + return False + return compare_versions(parse(_bitsandbytes_version), operation, version) + + +@cache +def is_gguf_version(operation: str, version: str): + """ + Compares the current Accelerate version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _gguf_available: + return False + return compare_versions(parse(_gguf_version), operation, version) + + +@cache +def is_sdnq_version(operation: str, version: str): + """ + Compares the current sdnq version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _sdnq_available: + return False + return compare_versions(parse(_sdnq_version), operation, version) + + +@cache +def is_torchao_version(operation: str, version: str): + """ + Compares the current torchao version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _torchao_available: + return False + return compare_versions(parse(_torchao_version), operation, version) + + +@cache +def is_optimum_quanto_version(operation: str, version: str): + """ + Compares the current Accelerate version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _optimum_quanto_available: + return False + return compare_versions(parse(_optimum_quanto_version), operation, version) + + +@cache +def is_nvidia_modelopt_version(operation: str, version: str): + """ + Compares the current Nvidia ModelOpt version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _nvidia_modelopt_available: + return False + return compare_versions(parse(_nvidia_modelopt_version), operation, version) + + +@cache +def is_xformers_version(operation: str, version: str): + """ + Compares the current xformers version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _xformers_available: + return False + return compare_versions(parse(_xformers_version), operation, version) + + +@cache +def is_sageattention_version(operation: str, version: str): + """ + Compares the current sageattention version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _sageattention_available: + return False + return compare_versions(parse(_sageattention_version), operation, version) + + +@cache +def is_flash_attn_version(operation: str, version: str): + """ + Compares the current flash-attention version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _flash_attn_available: + return False + return compare_versions(parse(_flash_attn_version), operation, version) + + +@cache +def is_aiter_version(operation: str, version: str): + """ + Compares the current aiter version to a given reference with an operation. + + Args: + operation (`str`): + A string representation of an operator, such as `">"` or `"<="` + version (`str`): + A version string + """ + if not _aiter_available: + return False + return compare_versions(parse(_aiter_version), operation, version) + + +def get_objects_from_module(module): + """ + Returns a dict of object names and values in a module, while skipping private/internal objects + + Args: + module (ModuleType): + Module to extract the objects from. + + Returns: + dict: Dictionary of object names and corresponding values + """ + + objects = {} + for name in dir(module): + if name.startswith("_"): + continue + objects[name] = getattr(module, name) + + return objects + + +class OptionalDependencyNotAvailable(BaseException): + """ + An error indicating that an optional dependency of Diffusers was not found in the environment. + """ + + +class _LazyModule(ModuleType): + """ + Module class that surfaces all objects but only performs associated imports when the objects are requested. + """ + + # Very heavily inspired by optuna.integration._IntegrationModule + # https://github.com/optuna/optuna/blob/master/optuna/integration/__init__.py + def __init__(self, name, module_file, import_structure, module_spec=None, extra_objects=None): + super().__init__(name) + self._modules = set(import_structure.keys()) + self._class_to_module = {} + for key, values in import_structure.items(): + for value in values: + self._class_to_module[value] = key + # Needed for autocompletion in an IDE + self.__all__ = list(import_structure.keys()) + list(chain(*import_structure.values())) + self.__file__ = module_file + self.__spec__ = module_spec + self.__path__ = [os.path.dirname(module_file)] + self._objects = {} if extra_objects is None else extra_objects + self._name = name + self._import_structure = import_structure + + # Needed for autocompletion in an IDE + def __dir__(self): + result = super().__dir__() + # The elements of self.__all__ that are submodules may or may not be in the dir already, depending on whether + # they have been accessed or not. So we only add the elements of self.__all__ that are not already in the dir. + for attr in self.__all__: + if attr not in result: + result.append(attr) + return result + + def __getattr__(self, name: str) -> Any: + if name in self._objects: + return self._objects[name] + if name in self._modules: + value = self._get_module(name) + elif name in self._class_to_module.keys(): + module = self._get_module(self._class_to_module[name]) + value = getattr(module, name) + else: + raise AttributeError(f"module {self.__name__} has no attribute {name}") + + setattr(self, name, value) + return value + + def _get_module(self, module_name: str): + try: + return importlib.import_module("." + module_name, self.__name__) + except Exception as e: + raise RuntimeError( + f"Failed to import {self.__name__}.{module_name} because of the following error (look up to see its" + f" traceback):\n{e}" + ) from e + + def __reduce__(self): + return (self.__class__, (self._name, self.__file__, self._import_structure)) diff --git a/diffusers/utils/loading_utils.py b/diffusers/utils/loading_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..c4fee0cfdd833a2bd013e982f8039ca54490b1ad --- /dev/null +++ b/diffusers/utils/loading_utils.py @@ -0,0 +1,163 @@ +import os +import tempfile +from typing import Any, Callable +from urllib.parse import unquote, urlparse + +import PIL.Image +import PIL.ImageOps +import requests + +from .constants import DIFFUSERS_REQUEST_TIMEOUT +from .import_utils import BACKENDS_MAPPING, is_imageio_available + + +def load_image( + image: str | PIL.Image.Image, convert_method: Callable[[PIL.Image.Image], PIL.Image.Image] | None = None +) -> PIL.Image.Image: + """ + Loads `image` to a PIL Image. + + Args: + image (`str` or `PIL.Image.Image`): + The image to convert to the PIL Image format. + convert_method (Callable[[PIL.Image.Image], PIL.Image.Image], *optional*): + A conversion method to apply to the image after loading it. When set to `None` the image will be converted + "RGB". + + Returns: + `PIL.Image.Image`: + A PIL Image. + """ + if isinstance(image, str): + if image.startswith("http://") or image.startswith("https://"): + image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw) + elif os.path.isfile(image): + image = PIL.Image.open(image) + else: + raise ValueError( + f"Incorrect path or URL. URLs must start with `http://` or `https://`, and {image} is not a valid path." + ) + elif isinstance(image, PIL.Image.Image): + image = image + else: + raise ValueError( + "Incorrect format used for the image. Should be a URL linking to an image, a local path, or a PIL image." + ) + + image = PIL.ImageOps.exif_transpose(image) + + if convert_method is not None: + image = convert_method(image) + else: + image = image.convert("RGB") + + return image + + +def load_video( + video: str, + convert_method: Callable[[list[PIL.Image.Image]], list[PIL.Image.Image]] | None = None, +) -> list[PIL.Image.Image]: + """ + Loads `video` to a list of PIL Image. + + Args: + video (`str`): + A URL or Path to a video to convert to a list of PIL Image format. + convert_method (Callable[[list[PIL.Image.Image]], list[PIL.Image.Image]], *optional*): + A conversion method to apply to the video after loading it. When set to `None` the images will be converted + to "RGB". + + Returns: + `list[PIL.Image.Image]`: + The video as a list of PIL images. + """ + is_url = video.startswith("http://") or video.startswith("https://") + is_file = os.path.isfile(video) + was_tempfile_created = False + + if not (is_url or is_file): + raise ValueError( + f"Incorrect path or URL. URLs must start with `http://` or `https://`, and {video} is not a valid path." + ) + + if is_url: + response = requests.get(video, stream=True) + if response.status_code != 200: + raise ValueError(f"Failed to download video. Status code: {response.status_code}") + + parsed_url = urlparse(video) + file_name = os.path.basename(unquote(parsed_url.path)) + + suffix = os.path.splitext(file_name)[1] or ".mp4" + video_path = tempfile.NamedTemporaryFile(suffix=suffix, delete=False).name + + was_tempfile_created = True + + video_data = response.iter_content(chunk_size=8192) + with open(video_path, "wb") as f: + for chunk in video_data: + f.write(chunk) + + video = video_path + + pil_images = [] + if video.endswith(".gif"): + gif = PIL.Image.open(video) + try: + while True: + pil_images.append(gif.copy()) + gif.seek(gif.tell() + 1) + except EOFError: + pass + + else: + if is_imageio_available(): + import imageio + else: + raise ImportError(BACKENDS_MAPPING["imageio"][1].format("load_video")) + + try: + imageio.plugins.ffmpeg.get_exe() + except AttributeError: + raise AttributeError( + "`Unable to find an ffmpeg installation on your machine. Please install via `pip install imageio-ffmpeg" + ) + + with imageio.get_reader(video) as reader: + # Read all frames + for frame in reader: + pil_images.append(PIL.Image.fromarray(frame)) + + if was_tempfile_created: + os.remove(video_path) + + if convert_method is not None: + pil_images = convert_method(pil_images) + + return pil_images + + +# Taken from `transformers`. +def get_module_from_name(module, tensor_name: str) -> tuple[Any, str]: + if "." in tensor_name: + splits = tensor_name.split(".") + for split in splits[:-1]: + new_module = getattr(module, split) + if new_module is None: + raise ValueError(f"{module} has no attribute {split}.") + module = new_module + tensor_name = splits[-1] + return module, tensor_name + + +def get_submodule_by_name(root_module, module_path: str): + current = root_module + parts = module_path.split(".") + for part in parts: + if part.isdigit(): + idx = int(part) + current = current[idx] # e.g., for nn.ModuleList or nn.Sequential + else: + current = getattr(current, part) + return current diff --git a/diffusers/utils/logging.py b/diffusers/utils/logging.py new file mode 100644 index 0000000000000000000000000000000000000000..21c4bba8a830e6f687dd35bdaf61743c151dacff --- /dev/null +++ b/diffusers/utils/logging.py @@ -0,0 +1,361 @@ +# coding=utf-8 +# Copyright 2025 Optuna, Hugging Face +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Logging utilities.""" + +import logging +import os +import sys +import threading +from logging import ( + CRITICAL, # NOQA + DEBUG, # NOQA + ERROR, # NOQA + FATAL, # NOQA + INFO, # NOQA + NOTSET, # NOQA + WARN, # NOQA + WARNING, # NOQA +) + +from tqdm import auto as tqdm_lib + +from .distributed_utils import is_torch_dist_rank_zero + + +_lock = threading.Lock() +_default_handler: logging.Handler | None = None + +log_levels = { + "debug": logging.DEBUG, + "info": logging.INFO, + "warning": logging.WARNING, + "error": logging.ERROR, + "critical": logging.CRITICAL, +} + +_default_log_level = logging.WARNING + +_tqdm_active = True +_rank_zero_filter = None + + +class _RankZeroFilter(logging.Filter): + def filter(self, record): + # Always allow rank-zero logs, but keep debug-level messages from all ranks for troubleshooting. + return is_torch_dist_rank_zero() or record.levelno <= logging.DEBUG + + +def _ensure_rank_zero_filter(logger: logging.Logger) -> None: + global _rank_zero_filter + + if _rank_zero_filter is None: + _rank_zero_filter = _RankZeroFilter() + + if not any(isinstance(f, _RankZeroFilter) for f in logger.filters): + logger.addFilter(_rank_zero_filter) + + +def _get_default_logging_level() -> int: + """ + If DIFFUSERS_VERBOSITY env var is set to one of the valid choices return that as the new default level. If it is + not - fall back to `_default_log_level` + """ + env_level_str = os.getenv("DIFFUSERS_VERBOSITY", None) + if env_level_str: + if env_level_str in log_levels: + return log_levels[env_level_str] + else: + logging.getLogger().warning( + f"Unknown option DIFFUSERS_VERBOSITY={env_level_str}, has to be one of: {', '.join(log_levels.keys())}" + ) + return _default_log_level + + +def _get_library_name() -> str: + return __name__.split(".")[0] + + +def _get_library_root_logger() -> logging.Logger: + return logging.getLogger(_get_library_name()) + + +def _configure_library_root_logger() -> None: + global _default_handler + + with _lock: + if _default_handler: + # This library has already configured the library root logger. + return + _default_handler = logging.StreamHandler() # Set sys.stderr as stream. + + if sys.stderr: # only if sys.stderr exists, e.g. when not using pythonw in windows + _default_handler.flush = sys.stderr.flush + + # Apply our default configuration to the library root logger. + library_root_logger = _get_library_root_logger() + library_root_logger.addHandler(_default_handler) + library_root_logger.setLevel(_get_default_logging_level()) + library_root_logger.propagate = False + _ensure_rank_zero_filter(library_root_logger) + + +def _reset_library_root_logger() -> None: + global _default_handler + + with _lock: + if not _default_handler: + return + + library_root_logger = _get_library_root_logger() + library_root_logger.removeHandler(_default_handler) + library_root_logger.setLevel(logging.NOTSET) + _default_handler = None + + +def get_log_levels_dict() -> dict[str, int]: + return log_levels + + +def get_logger(name: str | None = None) -> logging.Logger: + """ + Return a logger with the specified name. + + This function is not supposed to be directly accessed unless you are writing a custom diffusers module. + """ + + if name is None: + name = _get_library_name() + + _configure_library_root_logger() + logger = logging.getLogger(name) + _ensure_rank_zero_filter(logger) + return logger + + +def get_verbosity() -> int: + """ + Return the current level for the 🤗 Diffusers' root logger as an `int`. + + Returns: + `int`: + Logging level integers which can be one of: + + - `50`: `diffusers.logging.CRITICAL` or `diffusers.logging.FATAL` + - `40`: `diffusers.logging.ERROR` + - `30`: `diffusers.logging.WARNING` or `diffusers.logging.WARN` + - `20`: `diffusers.logging.INFO` + - `10`: `diffusers.logging.DEBUG` + + """ + + _configure_library_root_logger() + return _get_library_root_logger().getEffectiveLevel() + + +def set_verbosity(verbosity: int) -> None: + """ + Set the verbosity level for the 🤗 Diffusers' root logger. + + Args: + verbosity (`int`): + Logging level which can be one of: + + - `diffusers.logging.CRITICAL` or `diffusers.logging.FATAL` + - `diffusers.logging.ERROR` + - `diffusers.logging.WARNING` or `diffusers.logging.WARN` + - `diffusers.logging.INFO` + - `diffusers.logging.DEBUG` + """ + + _configure_library_root_logger() + _get_library_root_logger().setLevel(verbosity) + + +def set_verbosity_info() -> None: + """Set the verbosity to the `INFO` level.""" + return set_verbosity(INFO) + + +def set_verbosity_warning() -> None: + """Set the verbosity to the `WARNING` level.""" + return set_verbosity(WARNING) + + +def set_verbosity_debug() -> None: + """Set the verbosity to the `DEBUG` level.""" + return set_verbosity(DEBUG) + + +def set_verbosity_error() -> None: + """Set the verbosity to the `ERROR` level.""" + return set_verbosity(ERROR) + + +def disable_default_handler() -> None: + """Disable the default handler of the 🤗 Diffusers' root logger.""" + + _configure_library_root_logger() + + assert _default_handler is not None + _get_library_root_logger().removeHandler(_default_handler) + + +def enable_default_handler() -> None: + """Enable the default handler of the 🤗 Diffusers' root logger.""" + + _configure_library_root_logger() + + assert _default_handler is not None + _get_library_root_logger().addHandler(_default_handler) + + +def add_handler(handler: logging.Handler) -> None: + """adds a handler to the HuggingFace Diffusers' root logger.""" + + _configure_library_root_logger() + + assert handler is not None + _get_library_root_logger().addHandler(handler) + + +def remove_handler(handler: logging.Handler) -> None: + """removes given handler from the HuggingFace Diffusers' root logger.""" + + _configure_library_root_logger() + + assert handler is not None and handler in _get_library_root_logger().handlers + _get_library_root_logger().removeHandler(handler) + + +def disable_propagation() -> None: + """ + Disable propagation of the library log outputs. Note that log propagation is disabled by default. + """ + + _configure_library_root_logger() + _get_library_root_logger().propagate = False + + +def enable_propagation() -> None: + """ + Enable propagation of the library log outputs. Please disable the HuggingFace Diffusers' default handler to prevent + double logging if the root logger has been configured. + """ + + _configure_library_root_logger() + _get_library_root_logger().propagate = True + + +def enable_explicit_format() -> None: + """ + Enable explicit formatting for every 🤗 Diffusers' logger. The explicit formatter is as follows: + ``` + [LEVELNAME|FILENAME|LINE NUMBER] TIME >> MESSAGE + ``` + All handlers currently bound to the root logger are affected by this method. + """ + handlers = _get_library_root_logger().handlers + + for handler in handlers: + formatter = logging.Formatter("[%(levelname)s|%(filename)s:%(lineno)s] %(asctime)s >> %(message)s") + handler.setFormatter(formatter) + + +def reset_format() -> None: + """ + Resets the formatting for 🤗 Diffusers' loggers. + + All handlers currently bound to the root logger are affected by this method. + """ + handlers = _get_library_root_logger().handlers + + for handler in handlers: + handler.setFormatter(None) + + +def warning_advice(self, *args, **kwargs) -> None: + """ + This method is identical to `logger.warning()`, but if env var DIFFUSERS_NO_ADVISORY_WARNINGS=1 is set, this + warning will not be printed + """ + no_advisory_warnings = os.getenv("DIFFUSERS_NO_ADVISORY_WARNINGS", False) + if no_advisory_warnings: + return + self.warning(*args, **kwargs) + + +logging.Logger.warning_advice = warning_advice + + +class EmptyTqdm: + """Dummy tqdm which doesn't do anything.""" + + def __init__(self, *args, **kwargs): # pylint: disable=unused-argument + self._iterator = args[0] if args else None + + def __iter__(self): + return iter(self._iterator) + + def __getattr__(self, _): + """Return empty function.""" + + def empty_fn(*args, **kwargs): # pylint: disable=unused-argument + return + + return empty_fn + + def __enter__(self): + return self + + def __exit__(self, type_, value, traceback): + return + + +class _tqdm_cls: + def __call__(self, *args, **kwargs): + if _tqdm_active: + return tqdm_lib.tqdm(*args, **kwargs) + else: + return EmptyTqdm(*args, **kwargs) + + def set_lock(self, *args, **kwargs): + self._lock = None + if _tqdm_active: + return tqdm_lib.tqdm.set_lock(*args, **kwargs) + + def get_lock(self): + if _tqdm_active: + return tqdm_lib.tqdm.get_lock() + + +tqdm = _tqdm_cls() + + +def is_progress_bar_enabled() -> bool: + """Return a boolean indicating whether tqdm progress bars are enabled.""" + global _tqdm_active + return bool(_tqdm_active) + + +def enable_progress_bar() -> None: + """Enable tqdm progress bar.""" + global _tqdm_active + _tqdm_active = True + + +def disable_progress_bar() -> None: + """Disable tqdm progress bar.""" + global _tqdm_active + _tqdm_active = False diff --git a/diffusers/utils/model_card_template.md b/diffusers/utils/model_card_template.md new file mode 100644 index 0000000000000000000000000000000000000000..f41b71e24e2081425d3049ae51b50036d5d28b6a --- /dev/null +++ b/diffusers/utils/model_card_template.md @@ -0,0 +1,24 @@ +--- +{{ card_data }} +--- + + + +{{ model_description }} + +## Intended uses & limitations + +#### How to use + +```python +# TODO: add an example code snippet for running this diffusion pipeline +``` + +#### Limitations and bias + +[TODO: provide examples of latent issues and potential remediations] + +## Training details + +[TODO: describe the data used to train the model] diff --git a/diffusers/utils/outputs.py b/diffusers/utils/outputs.py new file mode 100644 index 0000000000000000000000000000000000000000..84c025d797739ad1e7a3ef01184ca9304b3f18e0 --- /dev/null +++ b/diffusers/utils/outputs.py @@ -0,0 +1,134 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Generic utilities +""" + +from collections import OrderedDict +from dataclasses import fields, is_dataclass +from typing import Any + +import numpy as np + +from .import_utils import is_torch_available, is_torch_version + + +def is_tensor(x) -> bool: + """ + Tests if `x` is a `torch.Tensor` or `np.ndarray`. + """ + if is_torch_available(): + import torch + + if isinstance(x, torch.Tensor): + return True + + return isinstance(x, np.ndarray) + + +class BaseOutput(OrderedDict): + """ + Base class for all model outputs as dataclass. Has a `__getitem__` that allows indexing by integer or slice (like a + tuple) or strings (like a dictionary) that will ignore the `None` attributes. Otherwise behaves like a regular + Python dictionary. + + > [!WARNING] > You can't unpack a [`BaseOutput`] directly. Use the [`~utils.BaseOutput.to_tuple`] method to convert + it to a tuple > first. + """ + + def __init_subclass__(cls) -> None: + """Register subclasses as pytree nodes. + + This is necessary to synchronize gradients when using `torch.nn.parallel.DistributedDataParallel` with + `static_graph=True` with modules that output `ModelOutput` subclasses. + """ + if is_torch_available(): + import torch.utils._pytree + + if is_torch_version("<", "2.2"): + torch.utils._pytree._register_pytree_node( + cls, + torch.utils._pytree._dict_flatten, + lambda values, context: cls(**torch.utils._pytree._dict_unflatten(values, context)), + ) + else: + torch.utils._pytree.register_pytree_node( + cls, + torch.utils._pytree._dict_flatten, + lambda values, context: cls(**torch.utils._pytree._dict_unflatten(values, context)), + serialized_type_name=f"{cls.__module__}.{cls.__name__}", + ) + + def __post_init__(self) -> None: + class_fields = fields(self) + + # Safety and consistency checks + if not len(class_fields): + raise ValueError(f"{self.__class__.__name__} has no fields.") + + first_field = getattr(self, class_fields[0].name) + other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:]) + + if other_fields_are_none and isinstance(first_field, dict): + for key, value in first_field.items(): + self[key] = value + else: + for field in class_fields: + v = getattr(self, field.name) + if v is not None: + self[field.name] = v + + def __delitem__(self, *args, **kwargs): + raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.") + + def setdefault(self, *args, **kwargs): + raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.") + + def pop(self, *args, **kwargs): + raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.") + + def update(self, *args, **kwargs): + raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.") + + def __getitem__(self, k: Any) -> Any: + if isinstance(k, str): + inner_dict = dict(self.items()) + return inner_dict[k] + else: + return self.to_tuple()[k] + + def __setattr__(self, name: Any, value: Any) -> None: + if name in self.keys() and value is not None: + # Don't call self.__setitem__ to avoid recursion errors + super().__setitem__(name, value) + super().__setattr__(name, value) + + def __setitem__(self, key, value): + # Will raise a KeyException if needed + super().__setitem__(key, value) + # Don't call self.__setattr__ to avoid recursion errors + super().__setattr__(key, value) + + def __reduce__(self): + if not is_dataclass(self): + return super().__reduce__() + callable, _args, *remaining = super().__reduce__() + args = tuple(getattr(self, field.name) for field in fields(self)) + return callable, args, *remaining + + def to_tuple(self) -> tuple[Any, ...]: + """ + Convert self to a tuple containing all the attributes/keys that are not `None`. + """ + return tuple(self[k] for k in self.keys()) diff --git a/diffusers/utils/peft_utils.py b/diffusers/utils/peft_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..748ecfe58e6ca0d52247d89b4c71524137aca9dc --- /dev/null +++ b/diffusers/utils/peft_utils.py @@ -0,0 +1,425 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +PEFT utilities: Utilities related to peft library +""" + +import collections +import functools +import importlib + +from packaging import version + +from . import logging +from .import_utils import is_peft_available, is_peft_version, is_torch_available +from .torch_utils import empty_device_cache + + +logger = logging.get_logger(__name__) + +if is_torch_available(): + import torch + + +def recurse_remove_peft_layers(model): + r""" + Recursively replace all instances of `LoraLayer` with corresponding new layers in `model`. + """ + from peft.tuners.tuners_utils import BaseTunerLayer + + has_base_layer_pattern = False + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + has_base_layer_pattern = hasattr(module, "base_layer") + break + + if has_base_layer_pattern: + from peft.utils import _get_submodules + + key_list = [key for key, _ in model.named_modules() if "lora" not in key] + for key in key_list: + try: + parent, target, target_name = _get_submodules(model, key) + except AttributeError: + continue + if hasattr(target, "base_layer"): + setattr(parent, target_name, target.get_base_layer()) + else: + # This is for backwards compatibility with PEFT <= 0.6.2. + # TODO can be removed once that PEFT version is no longer supported. + from peft.tuners.lora import LoraLayer + + for name, module in model.named_children(): + if len(list(module.children())) > 0: + ## compound module, go inside it + recurse_remove_peft_layers(module) + + module_replaced = False + + if isinstance(module, LoraLayer) and isinstance(module, torch.nn.Linear): + new_module = torch.nn.Linear( + module.in_features, + module.out_features, + bias=module.bias is not None, + ).to(module.weight.device) + new_module.weight = module.weight + if module.bias is not None: + new_module.bias = module.bias + + module_replaced = True + elif isinstance(module, LoraLayer) and isinstance(module, torch.nn.Conv2d): + new_module = torch.nn.Conv2d( + module.in_channels, + module.out_channels, + module.kernel_size, + module.stride, + module.padding, + module.dilation, + module.groups, + ).to(module.weight.device) + + new_module.weight = module.weight + if module.bias is not None: + new_module.bias = module.bias + + module_replaced = True + + if module_replaced: + setattr(model, name, new_module) + del module + + empty_device_cache() + return model + + +def scale_lora_layers(model, weight): + """ + Adjust the weightage given to the LoRA layers of the model. + + Args: + model (`torch.nn.Module`): + The model to scale. + weight (`float`): + The weight to be given to the LoRA layers. + """ + from peft.tuners.tuners_utils import BaseTunerLayer + + if weight == 1.0: + return + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + module.scale_layer(weight) + + +def unscale_lora_layers(model, weight: float | None = None): + """ + Removes the previously passed weight given to the LoRA layers of the model. + + Args: + model (`torch.nn.Module`): + The model to scale. + weight (`float`, *optional*): + The weight to be given to the LoRA layers. If no scale is passed the scale of the lora layer will be + re-initialized to the correct value. If 0.0 is passed, we will re-initialize the scale with the correct + value. + """ + from peft.tuners.tuners_utils import BaseTunerLayer + + if weight is None or weight == 1.0: + return + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + if weight != 0: + module.unscale_layer(weight) + else: + for adapter_name in module.active_adapters: + # if weight == 0 unscale should re-set the scale to the original value. + module.set_scale(adapter_name, 1.0) + + +def get_peft_kwargs( + rank_dict, network_alpha_dict, peft_state_dict, is_unet=True, model_state_dict=None, adapter_name=None +): + rank_pattern = {} + alpha_pattern = {} + r = lora_alpha = list(rank_dict.values())[0] + + if len(set(rank_dict.values())) > 1: + # get the rank occurring the most number of times + r = collections.Counter(rank_dict.values()).most_common()[0][0] + + # for modules with rank different from the most occurring rank, add it to the `rank_pattern` + rank_pattern = dict(filter(lambda x: x[1] != r, rank_dict.items())) + rank_pattern = {k.split(".lora_B.")[0]: v for k, v in rank_pattern.items()} + + if network_alpha_dict is not None and len(network_alpha_dict) > 0: + if len(set(network_alpha_dict.values())) > 1: + # get the alpha occurring the most number of times + lora_alpha = collections.Counter(network_alpha_dict.values()).most_common()[0][0] + + # for modules with alpha different from the most occurring alpha, add it to the `alpha_pattern` + alpha_pattern = dict(filter(lambda x: x[1] != lora_alpha, network_alpha_dict.items())) + if is_unet: + alpha_pattern = { + ".".join(k.split(".lora_A.")[0].split(".")).replace(".alpha", ""): v + for k, v in alpha_pattern.items() + } + else: + alpha_pattern = {".".join(k.split(".down.")[0].split(".")[:-1]): v for k, v in alpha_pattern.items()} + else: + lora_alpha = set(network_alpha_dict.values()).pop() + + target_modules = list({name.split(".lora")[0] for name in peft_state_dict.keys()}) + use_dora = any("lora_magnitude_vector" in k for k in peft_state_dict) + # for now we know that the "bias" keys are only associated with `lora_B`. + lora_bias = any("lora_B" in k and k.endswith(".bias") for k in peft_state_dict) + + lora_config_kwargs = { + "r": r, + "lora_alpha": lora_alpha, + "rank_pattern": rank_pattern, + "alpha_pattern": alpha_pattern, + "target_modules": target_modules, + "use_dora": use_dora, + "lora_bias": lora_bias, + } + + return lora_config_kwargs + + +def get_adapter_name(model): + from peft.tuners.tuners_utils import BaseTunerLayer + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + return f"default_{len(module.r)}" + return "default_0" + + +def set_adapter_layers(model, enabled=True): + from peft.tuners.tuners_utils import BaseTunerLayer + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + # The recent version of PEFT needs to call `enable_adapters` instead + if hasattr(module, "enable_adapters"): + module.enable_adapters(enabled=enabled) + else: + module.disable_adapters = not enabled + + +def delete_adapter_layers(model, adapter_name): + from peft.tuners.tuners_utils import BaseTunerLayer + + for module in model.modules(): + if isinstance(module, BaseTunerLayer): + if hasattr(module, "delete_adapter"): + module.delete_adapter(adapter_name) + else: + raise ValueError( + "The version of PEFT you are using is not compatible, please use a version that is greater than 0.6.1" + ) + + # For transformers integration - we need to pop the adapter from the config + if getattr(model, "_hf_peft_config_loaded", False) and hasattr(model, "peft_config"): + model.peft_config.pop(adapter_name, None) + # In case all adapters are deleted, we need to delete the config + # and make sure to set the flag to False + if len(model.peft_config) == 0: + del model.peft_config + model._hf_peft_config_loaded = None + + +def set_weights_and_activate_adapters(model, adapter_names, weights): + from peft.tuners.tuners_utils import BaseTunerLayer + + def get_module_weight(weight_for_adapter, module_name): + if not isinstance(weight_for_adapter, dict): + # If weight_for_adapter is a single number, always return it. + return weight_for_adapter + + for layer_name, weight_ in weight_for_adapter.items(): + if layer_name in module_name: + return weight_ + + parts = module_name.split(".") + # e.g. key = "down_blocks.1.attentions.0" + key = f"{parts[0]}.{parts[1]}.attentions.{parts[3]}" + block_weight = weight_for_adapter.get(key, 1.0) + + return block_weight + + for module_name, module in model.named_modules(): + if isinstance(module, BaseTunerLayer): + # For backward compatibility with previous PEFT versions, set multiple active adapters + if hasattr(module, "set_adapter"): + module.set_adapter(adapter_names) + else: + module.active_adapter = adapter_names + + # Set the scaling weight for each adapter for this module + for adapter_name, weight in zip(adapter_names, weights): + module.set_scale(adapter_name, get_module_weight(weight, module_name)) + + +def apply_lora_scale(kwargs_name: str = "joint_attention_kwargs"): + """ + Decorator to automatically handle LoRA layer scaling/unscaling in forward methods. + + This decorator extracts the `lora_scale` from the specified kwargs parameter, applies scaling before the forward + pass, and ensures unscaling happens after, even if an exception occurs. + + Args: + kwargs_name (`str`, defaults to `"joint_attention_kwargs"`): + The name of the keyword argument that contains the LoRA scale. Common values include + "joint_attention_kwargs", "attention_kwargs", "cross_attention_kwargs", etc. + """ + + def decorator(forward_fn): + @functools.wraps(forward_fn) + def wrapper(self, *args, **kwargs): + from . import USE_PEFT_BACKEND + + lora_scale = 1.0 + attention_kwargs = kwargs.get(kwargs_name) + + if attention_kwargs is not None: + attention_kwargs = attention_kwargs.copy() + kwargs[kwargs_name] = attention_kwargs + lora_scale = attention_kwargs.pop("scale", 1.0) + + if not USE_PEFT_BACKEND and lora_scale != 1.0: + logger.warning( + f"Passing `scale` via `{kwargs_name}` when not using the PEFT backend is ineffective." + ) + + # Apply LoRA scaling if using PEFT backend + if USE_PEFT_BACKEND: + scale_lora_layers(self, lora_scale) + + try: + # Execute the forward pass + result = forward_fn(self, *args, **kwargs) + return result + finally: + # Always unscale, even if forward pass raises an exception + if USE_PEFT_BACKEND: + unscale_lora_layers(self, lora_scale) + + return wrapper + + return decorator + + +def check_peft_version(min_version: str) -> None: + r""" + Checks if the version of PEFT is compatible. + + Args: + version (`str`): + The version of PEFT to check against. + """ + if not is_peft_available(): + raise ValueError("PEFT is not installed. Please install it with `pip install peft`") + + is_peft_version_compatible = version.parse(importlib.metadata.version("peft")) > version.parse(min_version) + + if not is_peft_version_compatible: + raise ValueError( + f"The version of PEFT you are using is not compatible, please use a version that is greater" + f" than {min_version}" + ) + + +def _create_lora_config( + state_dict, network_alphas, metadata, rank_pattern_dict, is_unet=True, model_state_dict=None, adapter_name=None +): + from peft import LoraConfig + + if metadata is not None: + lora_config_kwargs = metadata + else: + lora_config_kwargs = get_peft_kwargs( + rank_pattern_dict, + network_alpha_dict=network_alphas, + peft_state_dict=state_dict, + is_unet=is_unet, + model_state_dict=model_state_dict, + adapter_name=adapter_name, + ) + + _maybe_raise_error_for_ambiguous_keys(lora_config_kwargs) + + # Version checks for DoRA and lora_bias + if "use_dora" in lora_config_kwargs and lora_config_kwargs["use_dora"]: + if is_peft_version("<", "0.9.0"): + raise ValueError("DoRA requires PEFT >= 0.9.0. Please upgrade.") + + if "lora_bias" in lora_config_kwargs and lora_config_kwargs["lora_bias"]: + if is_peft_version("<=", "0.13.2"): + raise ValueError("lora_bias requires PEFT >= 0.14.0. Please upgrade.") + + try: + return LoraConfig(**lora_config_kwargs) + except TypeError as e: + raise TypeError("`LoraConfig` class could not be instantiated.") from e + + +def _maybe_raise_error_for_ambiguous_keys(config): + rank_pattern = config["rank_pattern"].copy() + target_modules = config["target_modules"] + + for key in list(rank_pattern.keys()): + # try to detect ambiguity + # `target_modules` can also be a str, in which case this loop would loop + # over the chars of the str. The technically correct way to match LoRA keys + # in PEFT is to use LoraModel._check_target_module_exists (lora_config, key). + # But this cuts it for now. + exact_matches = [mod for mod in target_modules if mod == key] + substring_matches = [mod for mod in target_modules if key in mod and mod != key] + + if exact_matches and substring_matches: + if is_peft_version("<", "0.14.1"): + raise ValueError( + "There are ambiguous keys present in this LoRA. To load it, please update your `peft` installation - `pip install -U peft`." + ) + + +def _maybe_warn_for_unhandled_keys(incompatible_keys, adapter_name): + warn_msg = "" + if incompatible_keys is not None: + # Check only for unexpected keys. + unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None) + if unexpected_keys: + lora_unexpected_keys = [k for k in unexpected_keys if "lora_" in k and adapter_name in k] + if lora_unexpected_keys: + warn_msg = ( + f"Loading adapter weights from state_dict led to unexpected keys found in the model:" + f" {', '.join(lora_unexpected_keys)}. " + ) + + # Filter missing keys specific to the current adapter. + missing_keys = getattr(incompatible_keys, "missing_keys", None) + if missing_keys: + lora_missing_keys = [k for k in missing_keys if "lora_" in k and adapter_name in k] + if lora_missing_keys: + warn_msg += ( + f"Loading adapter weights from state_dict led to missing keys in the model:" + f" {', '.join(lora_missing_keys)}." + ) + + if warn_msg: + logger.warning(warn_msg) diff --git a/diffusers/utils/pil_utils.py b/diffusers/utils/pil_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..72d4704fa945f0cc4edfa8e6a15955cd75fd18f1 --- /dev/null +++ b/diffusers/utils/pil_utils.py @@ -0,0 +1,65 @@ +import PIL.Image +import PIL.ImageOps +from packaging import version +from PIL import Image + + +if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"): + PIL_INTERPOLATION = { + "linear": PIL.Image.Resampling.BILINEAR, + "bilinear": PIL.Image.Resampling.BILINEAR, + "bicubic": PIL.Image.Resampling.BICUBIC, + "lanczos": PIL.Image.Resampling.LANCZOS, + "nearest": PIL.Image.Resampling.NEAREST, + } +else: + PIL_INTERPOLATION = { + "linear": PIL.Image.LINEAR, + "bilinear": PIL.Image.BILINEAR, + "bicubic": PIL.Image.BICUBIC, + "lanczos": PIL.Image.LANCZOS, + "nearest": PIL.Image.NEAREST, + } + + +def pt_to_pil(images): + """ + Convert a torch image to a PIL image. + """ + images = (images / 2 + 0.5).clamp(0, 1) + images = images.cpu().permute(0, 2, 3, 1).float().numpy() + images = numpy_to_pil(images) + return images + + +def numpy_to_pil(images): + """ + Convert a numpy image or a batch of images to a PIL image. + """ + if images.ndim == 3: + images = images[None, ...] + images = (images * 255).round().astype("uint8") + if images.shape[-1] == 1: + # special case for grayscale (single channel) images + pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images] + else: + pil_images = [Image.fromarray(image) for image in images] + + return pil_images + + +def make_image_grid(images: list[PIL.Image.Image], rows: int, cols: int, resize: int = None) -> PIL.Image.Image: + """ + Prepares a single grid of images. Useful for visualization purposes. + """ + assert len(images) == rows * cols + + if resize is not None: + images = [img.resize((resize, resize)) for img in images] + + w, h = images[0].size + grid = Image.new("RGB", size=(cols * w, rows * h)) + + for i, img in enumerate(images): + grid.paste(img, box=(i % cols * w, i // cols * h)) + return grid diff --git a/diffusers/utils/remote_utils.py b/diffusers/utils/remote_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..3a478ecf4085c62f62fec126cb92c4a4583dfabf --- /dev/null +++ b/diffusers/utils/remote_utils.py @@ -0,0 +1,427 @@ +# coding=utf-8 +# Copyright 2026 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from __future__ import annotations + +import io +import json +from typing import Literal, cast + +import requests + +from .deprecation_utils import deprecate +from .import_utils import is_safetensors_available, is_torch_available + + +if is_torch_available(): + import torch + + from ..image_processor import VaeImageProcessor + from ..video_processor import VideoProcessor + + if is_safetensors_available(): + import safetensors.torch + + DTYPE_MAP = { + "float16": torch.float16, + "float32": torch.float32, + "bfloat16": torch.bfloat16, + "uint8": torch.uint8, + } + + +from PIL import Image + + +def detect_image_type(data: bytes) -> str: + if data.startswith(b"\xff\xd8"): + return "jpeg" + elif data.startswith(b"\x89PNG\r\n\x1a\n"): + return "png" + elif data.startswith(b"GIF87a") or data.startswith(b"GIF89a"): + return "gif" + elif data.startswith(b"BM"): + return "bmp" + return "unknown" + + +def check_inputs_decode( + endpoint: str, + tensor: "torch.Tensor", + processor: "VaeImageProcessor" | "VideoProcessor" | None = None, + do_scaling: bool = True, + scaling_factor: float | None = None, + shift_factor: float | None = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + return_type: Literal["mp4", "pil", "pt"] = "pil", + image_format: Literal["png", "jpg"] = "jpg", + partial_postprocess: bool = False, + input_tensor_type: Literal["binary"] = "binary", + output_tensor_type: Literal["binary"] = "binary", + height: int | None = None, + width: int | None = None, +): + if tensor.ndim == 3 and height is None and width is None: + raise ValueError("`height` and `width` required for packed latents.") + if ( + output_type == "pt" + and return_type == "pil" + and not partial_postprocess + and not isinstance(processor, (VaeImageProcessor, VideoProcessor)) + ): + raise ValueError("`processor` is required.") + if do_scaling and scaling_factor is None: + deprecate( + "do_scaling", + "1.0.0", + "`do_scaling` is deprecated, pass `scaling_factor` and `shift_factor` if required.", + standard_warn=False, + ) + + +def postprocess_decode( + response: requests.Response, + processor: "VaeImageProcessor" | "VideoProcessor" | None = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + return_type: Literal["mp4", "pil", "pt"] = "pil", + partial_postprocess: bool = False, +): + if output_type == "pt" or (output_type == "pil" and processor is not None): + output_tensor = response.content + parameters = response.headers + shape = json.loads(parameters["shape"]) + dtype = parameters["dtype"] + torch_dtype = DTYPE_MAP[dtype] + output_tensor = torch.frombuffer(bytearray(output_tensor), dtype=torch_dtype).reshape(shape) + if output_type == "pt": + if partial_postprocess: + if return_type == "pil": + output = [Image.fromarray(image.numpy()) for image in output_tensor] + if len(output) == 1: + output = output[0] + elif return_type == "pt": + output = output_tensor + else: + if processor is None or return_type == "pt": + output = output_tensor + else: + if isinstance(processor, VideoProcessor): + output = cast( + list[Image.Image], + processor.postprocess_video(output_tensor, output_type="pil")[0], + ) + else: + output = cast( + Image.Image, + processor.postprocess(output_tensor, output_type="pil")[0], + ) + elif output_type == "pil" and return_type == "pil" and processor is None: + output = Image.open(io.BytesIO(response.content)).convert("RGB") + detected_format = detect_image_type(response.content) + output.format = detected_format + elif output_type == "pil" and processor is not None: + if return_type == "pil": + output = [ + Image.fromarray(image) + for image in (output_tensor.permute(0, 2, 3, 1).float().numpy() * 255).round().astype("uint8") + ] + elif return_type == "pt": + output = output_tensor + elif output_type == "mp4" and return_type == "mp4": + output = response.content + return output + + +def prepare_decode( + tensor: "torch.Tensor", + processor: "VaeImageProcessor" | "VideoProcessor" | None = None, + do_scaling: bool = True, + scaling_factor: float | None = None, + shift_factor: float | None = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + image_format: Literal["png", "jpg"] = "jpg", + partial_postprocess: bool = False, + height: int | None = None, + width: int | None = None, +): + headers = {} + parameters = { + "image_format": image_format, + "output_type": output_type, + "partial_postprocess": partial_postprocess, + "shape": list(tensor.shape), + "dtype": str(tensor.dtype).split(".")[-1], + } + if do_scaling and scaling_factor is not None: + parameters["scaling_factor"] = scaling_factor + if do_scaling and shift_factor is not None: + parameters["shift_factor"] = shift_factor + if do_scaling and scaling_factor is None: + parameters["do_scaling"] = do_scaling + elif do_scaling and scaling_factor is None and shift_factor is None: + parameters["do_scaling"] = do_scaling + if height is not None and width is not None: + parameters["height"] = height + parameters["width"] = width + headers["Content-Type"] = "tensor/binary" + headers["Accept"] = "tensor/binary" + if output_type == "pil" and image_format == "jpg" and processor is None: + headers["Accept"] = "image/jpeg" + elif output_type == "pil" and image_format == "png" and processor is None: + headers["Accept"] = "image/png" + elif output_type == "mp4": + headers["Accept"] = "text/plain" + tensor_data = safetensors.torch._to_ndarray(tensor)[0].tobytes() + return {"data": tensor_data, "params": parameters, "headers": headers} + + +def remote_decode( + endpoint: str, + tensor: "torch.Tensor", + processor: "VaeImageProcessor" | "VideoProcessor" | None = None, + do_scaling: bool = True, + scaling_factor: float | None = None, + shift_factor: float | None = None, + output_type: Literal["mp4", "pil", "pt"] = "pil", + return_type: Literal["mp4", "pil", "pt"] = "pil", + image_format: Literal["png", "jpg"] = "jpg", + partial_postprocess: bool = False, + input_tensor_type: Literal["binary"] = "binary", + output_tensor_type: Literal["binary"] = "binary", + height: int | None = None, + width: int | None = None, +) -> Image.Image | list[Image.Image] | bytes | "torch.Tensor": + """ + Hugging Face Hybrid Inference that allow running VAE decode remotely. + + Args: + endpoint (`str`): + Endpoint for Remote Decode. + tensor (`torch.Tensor`): + Tensor to be decoded. + processor (`VaeImageProcessor` or `VideoProcessor`, *optional*): + Used with `return_type="pt"`, and `return_type="pil"` for Video models. + do_scaling (`bool`, default `True`, *optional*): + **DEPRECATED**. **pass `scaling_factor`/`shift_factor` instead.** **still set + do_scaling=None/do_scaling=False for no scaling until option is removed** When `True` scaling e.g. `latents + / self.vae.config.scaling_factor` is applied remotely. If `False`, input must be passed with scaling + applied. + scaling_factor (`float`, *optional*): + Scaling is applied when passed e.g. [`latents / + self.vae.config.scaling_factor`](https://github.com/huggingface/diffusers/blob/7007febae5cff000d4df9059d9cf35133e8b2ca9/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py#L1083C37-L1083C77). + - SD v1: 0.18215 + - SD XL: 0.13025 + - Flux: 0.3611 + If `None`, input must be passed with scaling applied. + shift_factor (`float`, *optional*): + Shift is applied when passed e.g. `latents + self.vae.config.shift_factor`. + - Flux: 0.1159 + If `None`, input must be passed with scaling applied. + output_type (`"mp4"` or `"pil"` or `"pt", default `"pil"): + **Endpoint** output type. Subject to change. Report feedback on preferred type. + + `"mp4": Supported by video models. Endpoint returns `bytes` of video. `"pil"`: Supported by image and video + models. + Image models: Endpoint returns `bytes` of an image in `image_format`. Video models: Endpoint returns + `torch.Tensor` with partial `postprocessing` applied. + Requires `processor` as a flag (any `None` value will work). + `"pt"`: Support by image and video models. Endpoint returns `torch.Tensor`. + With `partial_postprocess=True` the tensor is postprocessed `uint8` image tensor. + + Recommendations: + `"pt"` with `partial_postprocess=True` is the smallest transfer for full quality. `"pt"` with + `partial_postprocess=False` is the most compatible with third party code. `"pil"` with + `image_format="jpg"` is the smallest transfer overall. + + return_type (`"mp4"` or `"pil"` or `"pt", default `"pil"): + **Function** return type. + + `"mp4": Function returns `bytes` of video. `"pil"`: Function returns `PIL.Image.Image`. + With `output_type="pil" no further processing is applied. With `output_type="pt" a `PIL.Image.Image` is + created. + `partial_postprocess=False` `processor` is required. `partial_postprocess=True` `processor` is + **not** required. + `"pt"`: Function returns `torch.Tensor`. + `processor` is **not** required. `partial_postprocess=False` tensor is `float16` or `bfloat16`, without + denormalization. `partial_postprocess=True` tensor is `uint8`, denormalized. + + image_format (`"png"` or `"jpg"`, default `jpg`): + Used with `output_type="pil"`. Endpoint returns `jpg` or `png`. + + partial_postprocess (`bool`, default `False`): + Used with `output_type="pt"`. `partial_postprocess=False` tensor is `float16` or `bfloat16`, without + denormalization. `partial_postprocess=True` tensor is `uint8`, denormalized. + + input_tensor_type (`"binary"`, default `"binary"`): + Tensor transfer type. + + output_tensor_type (`"binary"`, default `"binary"`): + Tensor transfer type. + + height (`int`, **optional**): + Required for `"packed"` latents. + + width (`int`, **optional**): + Required for `"packed"` latents. + + Returns: + output (`Image.Image` or `list[Image.Image]` or `bytes` or `torch.Tensor`). + """ + if input_tensor_type == "base64": + deprecate( + "input_tensor_type='base64'", + "1.0.0", + "input_tensor_type='base64' is deprecated. Using `binary`.", + standard_warn=False, + ) + input_tensor_type = "binary" + if output_tensor_type == "base64": + deprecate( + "output_tensor_type='base64'", + "1.0.0", + "output_tensor_type='base64' is deprecated. Using `binary`.", + standard_warn=False, + ) + output_tensor_type = "binary" + check_inputs_decode( + endpoint, + tensor, + processor, + do_scaling, + scaling_factor, + shift_factor, + output_type, + return_type, + image_format, + partial_postprocess, + input_tensor_type, + output_tensor_type, + height, + width, + ) + kwargs = prepare_decode( + tensor=tensor, + processor=processor, + do_scaling=do_scaling, + scaling_factor=scaling_factor, + shift_factor=shift_factor, + output_type=output_type, + image_format=image_format, + partial_postprocess=partial_postprocess, + height=height, + width=width, + ) + response = requests.post(endpoint, **kwargs) + if not response.ok: + raise RuntimeError(response.json()) + output = postprocess_decode( + response=response, + processor=processor, + output_type=output_type, + return_type=return_type, + partial_postprocess=partial_postprocess, + ) + return output + + +def check_inputs_encode( + endpoint: str, + image: "torch.Tensor" | Image.Image, + scaling_factor: float | None = None, + shift_factor: float | None = None, +): + pass + + +def postprocess_encode( + response: requests.Response, +): + output_tensor = response.content + parameters = response.headers + shape = json.loads(parameters["shape"]) + dtype = parameters["dtype"] + torch_dtype = DTYPE_MAP[dtype] + output_tensor = torch.frombuffer(bytearray(output_tensor), dtype=torch_dtype).reshape(shape) + return output_tensor + + +def prepare_encode( + image: "torch.Tensor" | Image.Image, + scaling_factor: float | None = None, + shift_factor: float | None = None, +): + headers = {} + parameters = {} + if scaling_factor is not None: + parameters["scaling_factor"] = scaling_factor + if shift_factor is not None: + parameters["shift_factor"] = shift_factor + if isinstance(image, torch.Tensor): + data = safetensors.torch._to_ndarray(image.contiguous())[0].tobytes() + parameters["shape"] = list(image.shape) + parameters["dtype"] = str(image.dtype).split(".")[-1] + else: + buffer = io.BytesIO() + image.save(buffer, format="PNG") + data = buffer.getvalue() + return {"data": data, "params": parameters, "headers": headers} + + +def remote_encode( + endpoint: str, + image: "torch.Tensor" | Image.Image, + scaling_factor: float | None = None, + shift_factor: float | None = None, +) -> "torch.Tensor": + """ + Hugging Face Hybrid Inference that allow running VAE encode remotely. + + Args: + endpoint (`str`): + Endpoint for Remote Decode. + image (`torch.Tensor` or `PIL.Image.Image`): + Image to be encoded. + scaling_factor (`float`, *optional*): + Scaling is applied when passed e.g. [`latents * self.vae.config.scaling_factor`]. + - SD v1: 0.18215 + - SD XL: 0.13025 + - Flux: 0.3611 + If `None`, input must be passed with scaling applied. + shift_factor (`float`, *optional*): + Shift is applied when passed e.g. `latents - self.vae.config.shift_factor`. + - Flux: 0.1159 + If `None`, input must be passed with scaling applied. + + Returns: + output (`torch.Tensor`). + """ + check_inputs_encode( + endpoint, + image, + scaling_factor, + shift_factor, + ) + kwargs = prepare_encode( + image=image, + scaling_factor=scaling_factor, + shift_factor=shift_factor, + ) + response = requests.post(endpoint, **kwargs) + if not response.ok: + raise RuntimeError(response.json()) + output = postprocess_encode( + response=response, + ) + return output diff --git a/diffusers/utils/source_code_parsing_utils.py b/diffusers/utils/source_code_parsing_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..5f94711c21d825f4baa0b32bbc37fd94c89aba03 --- /dev/null +++ b/diffusers/utils/source_code_parsing_utils.py @@ -0,0 +1,52 @@ +import ast +import importlib +import inspect +import textwrap + + +class ReturnNameVisitor(ast.NodeVisitor): + """Thanks to ChatGPT for pairing.""" + + def __init__(self): + self.return_names = [] + + def visit_Return(self, node): + # Check if the return value is a tuple. + if isinstance(node.value, ast.Tuple): + for elt in node.value.elts: + if isinstance(elt, ast.Name): + self.return_names.append(elt.id) + else: + try: + self.return_names.append(ast.unparse(elt)) + except Exception: + self.return_names.append(str(elt)) + else: + if isinstance(node.value, ast.Name): + self.return_names.append(node.value.id) + else: + try: + self.return_names.append(ast.unparse(node.value)) + except Exception: + self.return_names.append(str(node.value)) + self.generic_visit(node) + + def _determine_parent_module(self, cls): + from diffusers import DiffusionPipeline + from diffusers.models.modeling_utils import ModelMixin + + if issubclass(cls, DiffusionPipeline): + return "pipelines" + elif issubclass(cls, ModelMixin): + return "models" + else: + raise NotImplementedError + + def get_ast_tree(self, cls, attribute_name="encode_prompt"): + parent_module_name = self._determine_parent_module(cls) + main_module = importlib.import_module(f"diffusers.{parent_module_name}") + current_cls_module = getattr(main_module, cls.__name__) + source_code = inspect.getsource(getattr(current_cls_module, attribute_name)) + source_code = textwrap.dedent(source_code) + tree = ast.parse(source_code) + return tree diff --git a/diffusers/utils/state_dict_utils.py b/diffusers/utils/state_dict_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..cdfdb7b1df0450bb702d921f30430f84d8c15c64 --- /dev/null +++ b/diffusers/utils/state_dict_utils.py @@ -0,0 +1,545 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +State dict utilities: utility methods for converting state dicts easily +""" + +import enum +import json + +from .import_utils import is_torch_available +from .logging import get_logger + + +if is_torch_available(): + import torch + + +logger = get_logger(__name__) + + +class StateDictType(enum.Enum): + """ + The mode to use when converting state dicts. + """ + + DIFFUSERS_OLD = "diffusers_old" + KOHYA_SS = "kohya_ss" + PEFT = "peft" + DIFFUSERS = "diffusers" + + +# We need to define a proper mapping for Unet since it uses different output keys than text encoder +# e.g. to_q_lora -> q_proj / to_q +UNET_TO_DIFFUSERS = { + ".to_out_lora.up": ".to_out.0.lora_B", + ".to_out_lora.down": ".to_out.0.lora_A", + ".to_q_lora.down": ".to_q.lora_A", + ".to_q_lora.up": ".to_q.lora_B", + ".to_k_lora.down": ".to_k.lora_A", + ".to_k_lora.up": ".to_k.lora_B", + ".to_v_lora.down": ".to_v.lora_A", + ".to_v_lora.up": ".to_v.lora_B", + ".lora.up": ".lora_B", + ".lora.down": ".lora_A", + ".to_out.lora_magnitude_vector": ".to_out.0.lora_magnitude_vector", +} + +CONTROL_LORA_TO_DIFFUSERS = { + ".to_q.down": ".to_q.lora_A.weight", + ".to_q.up": ".to_q.lora_B.weight", + ".to_k.down": ".to_k.lora_A.weight", + ".to_k.up": ".to_k.lora_B.weight", + ".to_v.down": ".to_v.lora_A.weight", + ".to_v.up": ".to_v.lora_B.weight", + ".to_out.0.down": ".to_out.0.lora_A.weight", + ".to_out.0.up": ".to_out.0.lora_B.weight", + ".ff.net.0.proj.down": ".ff.net.0.proj.lora_A.weight", + ".ff.net.0.proj.up": ".ff.net.0.proj.lora_B.weight", + ".ff.net.2.down": ".ff.net.2.lora_A.weight", + ".ff.net.2.up": ".ff.net.2.lora_B.weight", + ".proj_in.down": ".proj_in.lora_A.weight", + ".proj_in.up": ".proj_in.lora_B.weight", + ".proj_out.down": ".proj_out.lora_A.weight", + ".proj_out.up": ".proj_out.lora_B.weight", + ".conv.down": ".conv.lora_A.weight", + ".conv.up": ".conv.lora_B.weight", + **{f".conv{i}.down": f".conv{i}.lora_A.weight" for i in range(1, 3)}, + **{f".conv{i}.up": f".conv{i}.lora_B.weight" for i in range(1, 3)}, + "conv_in.down": "conv_in.lora_A.weight", + "conv_in.up": "conv_in.lora_B.weight", + ".conv_shortcut.down": ".conv_shortcut.lora_A.weight", + ".conv_shortcut.up": ".conv_shortcut.lora_B.weight", + **{f".linear_{i}.down": f".linear_{i}.lora_A.weight" for i in range(1, 3)}, + **{f".linear_{i}.up": f".linear_{i}.lora_B.weight" for i in range(1, 3)}, + "time_emb_proj.down": "time_emb_proj.lora_A.weight", + "time_emb_proj.up": "time_emb_proj.lora_B.weight", +} + +DIFFUSERS_TO_PEFT = { + ".q_proj.lora_linear_layer.up": ".q_proj.lora_B", + ".q_proj.lora_linear_layer.down": ".q_proj.lora_A", + ".k_proj.lora_linear_layer.up": ".k_proj.lora_B", + ".k_proj.lora_linear_layer.down": ".k_proj.lora_A", + ".v_proj.lora_linear_layer.up": ".v_proj.lora_B", + ".v_proj.lora_linear_layer.down": ".v_proj.lora_A", + ".out_proj.lora_linear_layer.up": ".out_proj.lora_B", + ".out_proj.lora_linear_layer.down": ".out_proj.lora_A", + ".lora_linear_layer.up": ".lora_B", + ".lora_linear_layer.down": ".lora_A", + "text_projection.lora.down.weight": "text_projection.lora_A.weight", + "text_projection.lora.up.weight": "text_projection.lora_B.weight", +} + +DIFFUSERS_OLD_TO_PEFT = { + ".to_q_lora.up": ".q_proj.lora_B", + ".to_q_lora.down": ".q_proj.lora_A", + ".to_k_lora.up": ".k_proj.lora_B", + ".to_k_lora.down": ".k_proj.lora_A", + ".to_v_lora.up": ".v_proj.lora_B", + ".to_v_lora.down": ".v_proj.lora_A", + ".to_out_lora.up": ".out_proj.lora_B", + ".to_out_lora.down": ".out_proj.lora_A", + ".lora_linear_layer.up": ".lora_B", + ".lora_linear_layer.down": ".lora_A", +} + +PEFT_TO_DIFFUSERS = { + ".q_proj.lora_B": ".q_proj.lora_linear_layer.up", + ".q_proj.lora_A": ".q_proj.lora_linear_layer.down", + ".k_proj.lora_B": ".k_proj.lora_linear_layer.up", + ".k_proj.lora_A": ".k_proj.lora_linear_layer.down", + ".v_proj.lora_B": ".v_proj.lora_linear_layer.up", + ".v_proj.lora_A": ".v_proj.lora_linear_layer.down", + ".out_proj.lora_B": ".out_proj.lora_linear_layer.up", + ".out_proj.lora_A": ".out_proj.lora_linear_layer.down", + "to_k.lora_A": "to_k.lora.down", + "to_k.lora_B": "to_k.lora.up", + "to_q.lora_A": "to_q.lora.down", + "to_q.lora_B": "to_q.lora.up", + "to_v.lora_A": "to_v.lora.down", + "to_v.lora_B": "to_v.lora.up", + "to_out.0.lora_A": "to_out.0.lora.down", + "to_out.0.lora_B": "to_out.0.lora.up", +} + +DIFFUSERS_OLD_TO_DIFFUSERS = { + ".to_q_lora.up": ".q_proj.lora_linear_layer.up", + ".to_q_lora.down": ".q_proj.lora_linear_layer.down", + ".to_k_lora.up": ".k_proj.lora_linear_layer.up", + ".to_k_lora.down": ".k_proj.lora_linear_layer.down", + ".to_v_lora.up": ".v_proj.lora_linear_layer.up", + ".to_v_lora.down": ".v_proj.lora_linear_layer.down", + ".to_out_lora.up": ".out_proj.lora_linear_layer.up", + ".to_out_lora.down": ".out_proj.lora_linear_layer.down", + ".to_k.lora_magnitude_vector": ".k_proj.lora_magnitude_vector", + ".to_v.lora_magnitude_vector": ".v_proj.lora_magnitude_vector", + ".to_q.lora_magnitude_vector": ".q_proj.lora_magnitude_vector", + ".to_out.lora_magnitude_vector": ".out_proj.lora_magnitude_vector", +} + +PEFT_TO_KOHYA_SS = { + "lora_A": "lora_down", + "lora_B": "lora_up", + # This is not a comprehensive dict as kohya format requires replacing `.` with `_` in keys, + # adding prefixes and adding alpha values + # Check `convert_state_dict_to_kohya` for more +} + +PEFT_STATE_DICT_MAPPINGS = { + StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_PEFT, + StateDictType.DIFFUSERS: DIFFUSERS_TO_PEFT, +} + +DIFFUSERS_STATE_DICT_MAPPINGS = { + StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_DIFFUSERS, + StateDictType.PEFT: PEFT_TO_DIFFUSERS, +} + +KOHYA_STATE_DICT_MAPPINGS = {StateDictType.PEFT: PEFT_TO_KOHYA_SS} + +KEYS_TO_ALWAYS_REPLACE = { + ".processor.": ".", +} + + +def convert_state_dict(state_dict, mapping): + r""" + Simply iterates over the state dict and replaces the patterns in `mapping` with the corresponding values. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + mapping (`dict[str, str]`): + The mapping to use for conversion, the mapping should be a dictionary with the following structure: + - key: the pattern to replace + - value: the pattern to replace with + + Returns: + converted_state_dict (`dict`) + The converted state dict. + """ + converted_state_dict = {} + for k, v in state_dict.items(): + # First, filter out the keys that we always want to replace + for pattern in KEYS_TO_ALWAYS_REPLACE.keys(): + if pattern in k: + new_pattern = KEYS_TO_ALWAYS_REPLACE[pattern] + k = k.replace(pattern, new_pattern) + + for pattern in mapping.keys(): + if pattern in k: + new_pattern = mapping[pattern] + k = k.replace(pattern, new_pattern) + break + converted_state_dict[k] = v + return converted_state_dict + + +def convert_state_dict_to_peft(state_dict, original_type=None, **kwargs): + r""" + Converts a state dict to the PEFT format The state dict can be from previous diffusers format (`OLD_DIFFUSERS`), or + new diffusers format (`DIFFUSERS`). The method only supports the conversion from diffusers old/new to PEFT for now. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + original_type (`StateDictType`, *optional*): + The original type of the state dict, if not provided, the method will try to infer it automatically. + """ + if original_type is None: + # Old diffusers to PEFT + if any("to_out_lora" in k for k in state_dict.keys()): + original_type = StateDictType.DIFFUSERS_OLD + elif any("lora_linear_layer" in k for k in state_dict.keys()): + original_type = StateDictType.DIFFUSERS + else: + raise ValueError("Could not automatically infer state dict type") + + if original_type not in PEFT_STATE_DICT_MAPPINGS.keys(): + raise ValueError(f"Original type {original_type} is not supported") + + mapping = PEFT_STATE_DICT_MAPPINGS[original_type] + return convert_state_dict(state_dict, mapping) + + +def convert_state_dict_to_diffusers(state_dict, original_type=None, **kwargs): + r""" + Converts a state dict to new diffusers format. The state dict can be from previous diffusers format + (`OLD_DIFFUSERS`), or PEFT format (`PEFT`) or new diffusers format (`DIFFUSERS`). In the last case the method will + return the state dict as is. + + The method only supports the conversion from diffusers old, PEFT to diffusers new for now. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + original_type (`StateDictType`, *optional*): + The original type of the state dict, if not provided, the method will try to infer it automatically. + kwargs (`dict`, *args*): + Additional arguments to pass to the method. + + - **adapter_name**: For example, in case of PEFT, some keys will be prepended + with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in + `get_peft_model_state_dict` method: + https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92 + but we add it here in case we don't want to rely on that method. + """ + peft_adapter_name = kwargs.pop("adapter_name", None) + if peft_adapter_name is not None: + peft_adapter_name = "." + peft_adapter_name + else: + peft_adapter_name = "" + + if original_type is None: + # Old diffusers to PEFT + if any("to_out_lora" in k for k in state_dict.keys()): + original_type = StateDictType.DIFFUSERS_OLD + elif any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()): + original_type = StateDictType.PEFT + elif any("lora_linear_layer" in k for k in state_dict.keys()): + # nothing to do + return state_dict + else: + raise ValueError("Could not automatically infer state dict type") + + if original_type not in DIFFUSERS_STATE_DICT_MAPPINGS.keys(): + raise ValueError(f"Original type {original_type} is not supported") + + mapping = DIFFUSERS_STATE_DICT_MAPPINGS[original_type] + return convert_state_dict(state_dict, mapping) + + +def convert_unet_state_dict_to_peft(state_dict): + r""" + Converts a state dict from UNet format to diffusers format - i.e. by removing some keys + """ + mapping = UNET_TO_DIFFUSERS + return convert_state_dict(state_dict, mapping) + + +def convert_sai_sd_control_lora_state_dict_to_peft(state_dict): + def _convert_controlnet_to_diffusers(state_dict): + is_sdxl = "input_blocks.11.0.in_layers.0.weight" not in state_dict + logger.info(f"Using ControlNet lora ({'SDXL' if is_sdxl else 'SD15'})") + + # Retrieves the keys for the input blocks only + num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in state_dict if "input_blocks" in layer}) + input_blocks = { + layer_id: [key for key in state_dict if f"input_blocks.{layer_id}" in key] + for layer_id in range(num_input_blocks) + } + layers_per_block = 2 + + # op blocks + op_blocks = [key for key in state_dict if "0.op" in key] + + converted_state_dict = {} + # Conv in layers + for key in input_blocks[0]: + diffusers_key = key.replace("input_blocks.0.0", "conv_in") + converted_state_dict[diffusers_key] = state_dict.get(key) + + # controlnet time embedding blocks + time_embedding_blocks = [key for key in state_dict if "time_embed" in key] + for key in time_embedding_blocks: + diffusers_key = key.replace("time_embed.0", "time_embedding.linear_1").replace( + "time_embed.2", "time_embedding.linear_2" + ) + converted_state_dict[diffusers_key] = state_dict.get(key) + + # controlnet label embedding blocks + label_embedding_blocks = [key for key in state_dict if "label_emb" in key] + for key in label_embedding_blocks: + diffusers_key = key.replace("label_emb.0.0", "add_embedding.linear_1").replace( + "label_emb.0.2", "add_embedding.linear_2" + ) + converted_state_dict[diffusers_key] = state_dict.get(key) + + # Down blocks + for i in range(1, num_input_blocks): + block_id = (i - 1) // (layers_per_block + 1) + layer_in_block_id = (i - 1) % (layers_per_block + 1) + + resnets = [ + key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key + ] + for key in resnets: + diffusers_key = ( + key.replace("in_layers.0", "norm1") + .replace("in_layers.2", "conv1") + .replace("out_layers.0", "norm2") + .replace("out_layers.3", "conv2") + .replace("emb_layers.1", "time_emb_proj") + .replace("skip_connection", "conv_shortcut") + ) + diffusers_key = diffusers_key.replace( + f"input_blocks.{i}.0", f"down_blocks.{block_id}.resnets.{layer_in_block_id}" + ) + converted_state_dict[diffusers_key] = state_dict.get(key) + + if f"input_blocks.{i}.0.op.bias" in state_dict: + for key in [key for key in op_blocks if f"input_blocks.{i}.0.op" in key]: + diffusers_key = key.replace( + f"input_blocks.{i}.0.op", f"down_blocks.{block_id}.downsamplers.0.conv" + ) + converted_state_dict[diffusers_key] = state_dict.get(key) + + attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key] + if attentions: + for key in attentions: + diffusers_key = key.replace( + f"input_blocks.{i}.1", f"down_blocks.{block_id}.attentions.{layer_in_block_id}" + ) + converted_state_dict[diffusers_key] = state_dict.get(key) + + # controlnet down blocks + for i in range(num_input_blocks): + converted_state_dict[f"controlnet_down_blocks.{i}.weight"] = state_dict.get(f"zero_convs.{i}.0.weight") + converted_state_dict[f"controlnet_down_blocks.{i}.bias"] = state_dict.get(f"zero_convs.{i}.0.bias") + + # Retrieves the keys for the middle blocks only + num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in state_dict if "middle_block" in layer}) + middle_blocks = { + layer_id: [key for key in state_dict if f"middle_block.{layer_id}" in key] + for layer_id in range(num_middle_blocks) + } + + # Mid blocks + for key in middle_blocks.keys(): + diffusers_key = max(key - 1, 0) + if key % 2 == 0: + for k in middle_blocks[key]: + diffusers_key_hf = ( + k.replace("in_layers.0", "norm1") + .replace("in_layers.2", "conv1") + .replace("out_layers.0", "norm2") + .replace("out_layers.3", "conv2") + .replace("emb_layers.1", "time_emb_proj") + .replace("skip_connection", "conv_shortcut") + ) + diffusers_key_hf = diffusers_key_hf.replace( + f"middle_block.{key}", f"mid_block.resnets.{diffusers_key}" + ) + converted_state_dict[diffusers_key_hf] = state_dict.get(k) + else: + for k in middle_blocks[key]: + diffusers_key_hf = k.replace(f"middle_block.{key}", f"mid_block.attentions.{diffusers_key}") + converted_state_dict[diffusers_key_hf] = state_dict.get(k) + + # mid block + converted_state_dict["controlnet_mid_block.weight"] = state_dict.get("middle_block_out.0.weight") + converted_state_dict["controlnet_mid_block.bias"] = state_dict.get("middle_block_out.0.bias") + + # controlnet cond embedding blocks + cond_embedding_blocks = { + ".".join(layer.split(".")[:2]) + for layer in state_dict + if "input_hint_block" in layer + and ("input_hint_block.0" not in layer) + and ("input_hint_block.14" not in layer) + } + num_cond_embedding_blocks = len(cond_embedding_blocks) + + for idx in range(1, num_cond_embedding_blocks + 1): + diffusers_idx = idx - 1 + cond_block_id = 2 * idx + + converted_state_dict[f"controlnet_cond_embedding.blocks.{diffusers_idx}.weight"] = state_dict.get( + f"input_hint_block.{cond_block_id}.weight" + ) + converted_state_dict[f"controlnet_cond_embedding.blocks.{diffusers_idx}.bias"] = state_dict.get( + f"input_hint_block.{cond_block_id}.bias" + ) + + for key in [key for key in state_dict if "input_hint_block.0" in key]: + diffusers_key = key.replace("input_hint_block.0", "controlnet_cond_embedding.conv_in") + converted_state_dict[diffusers_key] = state_dict.get(key) + + for key in [key for key in state_dict if "input_hint_block.14" in key]: + diffusers_key = key.replace("input_hint_block.14", "controlnet_cond_embedding.conv_out") + converted_state_dict[diffusers_key] = state_dict.get(key) + + return converted_state_dict + + state_dict = _convert_controlnet_to_diffusers(state_dict) + mapping = CONTROL_LORA_TO_DIFFUSERS + return convert_state_dict(state_dict, mapping) + + +def convert_all_state_dict_to_peft(state_dict): + r""" + Attempts to first `convert_state_dict_to_peft`, and if it doesn't detect `lora_linear_layer` for a valid + `DIFFUSERS` LoRA for example, attempts to exclusively convert the Unet `convert_unet_state_dict_to_peft` + """ + try: + peft_dict = convert_state_dict_to_peft(state_dict) + except Exception as e: + if str(e) == "Could not automatically infer state dict type": + peft_dict = convert_unet_state_dict_to_peft(state_dict) + else: + raise + + if not any("lora_A" in key or "lora_B" in key for key in peft_dict.keys()): + raise ValueError("Your LoRA was not converted to PEFT") + + return peft_dict + + +def convert_state_dict_to_kohya(state_dict, original_type=None, **kwargs): + r""" + Converts a `PEFT` state dict to `Kohya` format that can be used in AUTOMATIC1111, ComfyUI, SD.Next, InvokeAI, etc. + The method only supports the conversion from PEFT to Kohya for now. + + Args: + state_dict (`dict[str, torch.Tensor]`): + The state dict to convert. + original_type (`StateDictType`, *optional*): + The original type of the state dict, if not provided, the method will try to infer it automatically. + kwargs (`dict`, *args*): + Additional arguments to pass to the method. + + - **adapter_name**: For example, in case of PEFT, some keys will be prepended + with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in + `get_peft_model_state_dict` method: + https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92 + but we add it here in case we don't want to rely on that method. + """ + try: + import torch + except ImportError: + logger.error("Converting PEFT state dicts to Kohya requires torch to be installed.") + raise + + peft_adapter_name = kwargs.pop("adapter_name", None) + if peft_adapter_name is not None: + peft_adapter_name = "." + peft_adapter_name + else: + peft_adapter_name = "" + + if original_type is None: + if any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()): + original_type = StateDictType.PEFT + + if original_type not in KOHYA_STATE_DICT_MAPPINGS.keys(): + raise ValueError(f"Original type {original_type} is not supported") + + # Use the convert_state_dict function with the appropriate mapping + kohya_ss_partial_state_dict = convert_state_dict(state_dict, KOHYA_STATE_DICT_MAPPINGS[StateDictType.PEFT]) + kohya_ss_state_dict = {} + + # Additional logic for replacing header, alpha parameters `.` with `_` in all keys + for kohya_key, weight in kohya_ss_partial_state_dict.items(): + if "text_encoder_2." in kohya_key: + kohya_key = kohya_key.replace("text_encoder_2.", "lora_te2.") + elif "text_encoder." in kohya_key: + kohya_key = kohya_key.replace("text_encoder.", "lora_te1.") + elif "unet" in kohya_key: + kohya_key = kohya_key.replace("unet", "lora_unet") + elif "lora_magnitude_vector" in kohya_key: + kohya_key = kohya_key.replace("lora_magnitude_vector", "dora_scale") + + kohya_key = kohya_key.replace(".", "_", kohya_key.count(".") - 2) + kohya_key = kohya_key.replace(peft_adapter_name, "") # Kohya doesn't take names + kohya_ss_state_dict[kohya_key] = weight + if "lora_down" in kohya_key: + alpha_key = f"{kohya_key.split('.')[0]}.alpha" + kohya_ss_state_dict[alpha_key] = torch.tensor(len(weight)) + + return kohya_ss_state_dict + + +def state_dict_all_zero(state_dict, filter_str=None): + if filter_str is not None: + if isinstance(filter_str, str): + filter_str = [filter_str] + state_dict = {k: v for k, v in state_dict.items() if any(f in k for f in filter_str)} + + return all(torch.all(param == 0).item() for param in state_dict.values()) + + +def _load_sft_state_dict_metadata(model_file: str): + import safetensors.torch + + from ..loaders.lora_base import LORA_ADAPTER_METADATA_KEY + + with safetensors.torch.safe_open(model_file, framework="pt", device="cpu") as f: + metadata = f.metadata() or {} + + metadata.pop("format", None) + if metadata: + raw = metadata.get(LORA_ADAPTER_METADATA_KEY) + return json.loads(raw) if raw else None + else: + return None diff --git a/diffusers/utils/testing_utils.py b/diffusers/utils/testing_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..b9f15b1b8d5de8f58bf6f5c69e2d5f068ac656f6 --- /dev/null +++ b/diffusers/utils/testing_utils.py @@ -0,0 +1,1620 @@ +import functools +import glob +import importlib +import importlib.metadata +import inspect +import io +import logging +import multiprocessing +import os +import random +import re +import struct +import sys +import tempfile +import time +import unittest +import urllib.parse +from collections import UserDict +from contextlib import contextmanager +from io import BytesIO, StringIO +from pathlib import Path +from typing import TYPE_CHECKING, Any, Callable, Set + +import numpy as np +import PIL.Image +import PIL.ImageOps +import requests +from numpy.linalg import norm +from packaging import version + +from .constants import DIFFUSERS_REQUEST_TIMEOUT +from .deprecation_utils import deprecate +from .import_utils import ( + BACKENDS_MAPPING, + is_accelerate_available, + is_bitsandbytes_available, + is_compel_available, + is_gguf_available, + is_kernels_available, + is_note_seq_available, + is_nvidia_modelopt_available, + is_onnx_available, + is_opencv_available, + is_optimum_quanto_available, + is_peft_available, + is_sdnq_available, + is_timm_available, + is_torch_available, + is_torch_neuronx_available, + is_torch_version, + is_torchao_available, + is_torchsde_available, + is_transformers_available, +) +from .logging import get_logger + + +if is_torch_available(): + import torch + + IS_ROCM_SYSTEM = torch.version.hip is not None + IS_CUDA_SYSTEM = torch.version.cuda is not None + IS_XPU_SYSTEM = getattr(torch.version, "xpu", None) is not None +else: + IS_ROCM_SYSTEM = False + IS_CUDA_SYSTEM = False + IS_XPU_SYSTEM = False + +global_rng = random.Random() + +logger = get_logger(__name__) +deprecate( + "diffusers.utils.testing_utils", + "1.0.0", + "diffusers.utils.testing_utils is deprecated and will be removed in a future version. " + "Determinism and device backend utilities have been moved to `diffusers.utils.torch_utils`. ", +) +_required_peft_version = is_peft_available() and version.parse( + version.parse(importlib.metadata.version("peft")).base_version +) > version.parse("0.5") +_required_transformers_version = is_transformers_available() and version.parse( + version.parse(importlib.metadata.version("transformers")).base_version +) > version.parse("4.33") + +USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version +BIG_GPU_MEMORY = int(os.getenv("BIG_GPU_MEMORY", 40)) + +if is_torch_available(): + import torch + + # Set a backend environment variable for any extra module import required for a custom accelerator + if "DIFFUSERS_TEST_BACKEND" in os.environ: + backend = os.environ["DIFFUSERS_TEST_BACKEND"] + try: + _ = importlib.import_module(backend) + except ModuleNotFoundError as e: + raise ModuleNotFoundError( + f"Failed to import `DIFFUSERS_TEST_BACKEND` '{backend}'! This should be the name of an installed module \ + to enable a specified backend.):\n{e}" + ) from e + + if "DIFFUSERS_TEST_DEVICE" in os.environ: + torch_device = os.environ["DIFFUSERS_TEST_DEVICE"] + try: + # try creating device to see if provided device is valid + _ = torch.device(torch_device) + except RuntimeError as e: + raise RuntimeError( + f"Unknown testing device specified by environment variable `DIFFUSERS_TEST_DEVICE`: {torch_device}" + ) from e + logger.info(f"torch_device overrode to {torch_device}") + else: + if torch.cuda.is_available(): + torch_device = "cuda" + elif torch.xpu.is_available(): + torch_device = "xpu" + elif is_torch_neuronx_available() and hasattr(torch, "neuron") and torch.neuron.is_available(): + torch_device = torch.neuron.current_device() + else: + torch_device = "cpu" + is_torch_higher_equal_than_1_12 = version.parse( + version.parse(torch.__version__).base_version + ) >= version.parse("1.12") + + if is_torch_higher_equal_than_1_12: + # Some builds of torch 1.12 don't have the mps backend registered. See #892 for more details + mps_backend_registered = hasattr(torch.backends, "mps") + torch_device = "mps" if (mps_backend_registered and torch.backends.mps.is_available()) else torch_device + + from .torch_utils import get_torch_cuda_device_capability + + +def torch_all_close(a, b, *args, **kwargs): + if not is_torch_available(): + raise ValueError("PyTorch needs to be installed to use this function.") + if not torch.allclose(a, b, *args, **kwargs): + assert False, f"Max diff is absolute {(a - b).abs().max()}. Diff tensor is {(a - b).abs()}." + return True + + +def numpy_cosine_similarity_distance(a, b): + similarity = np.dot(a, b) / (norm(a) * norm(b)) + distance = 1.0 - similarity.mean() + + return distance + + +def check_if_dicts_are_equal(dict1, dict2): + dict1, dict2 = dict1.copy(), dict2.copy() + + for key, value in dict1.items(): + if isinstance(value, set): + dict1[key] = sorted(value) + for key, value in dict2.items(): + if isinstance(value, set): + dict2[key] = sorted(value) + + for key in dict1: + if key not in dict2: + return False + if dict1[key] != dict2[key]: + return False + + for key in dict2: + if key not in dict1: + return False + + return True + + +def print_tensor_test( + tensor, + limit_to_slices=None, + max_torch_print=None, + filename="test_corrections.txt", + expected_tensor_name="expected_slice", +): + if max_torch_print: + torch.set_printoptions(threshold=10_000) + + test_name = os.environ.get("PYTEST_CURRENT_TEST") + if not torch.is_tensor(tensor): + tensor = torch.from_numpy(tensor) + if limit_to_slices: + tensor = tensor[0, -3:, -3:, -1] + + tensor_str = str(tensor.detach().cpu().flatten().to(torch.float32)).replace("\n", "") + # format is usually: + # expected_slice = np.array([-0.5713, -0.3018, -0.9814, 0.04663, -0.879, 0.76, -1.734, 0.1044, 1.161]) + output_str = tensor_str.replace("tensor", f"{expected_tensor_name} = np.array") + test_file, test_class, test_fn = test_name.split("::") + test_fn = test_fn.split()[0] + with open(filename, "a") as f: + print("::".join([test_file, test_class, test_fn, output_str]), file=f) + + +def get_tests_dir(append_path=None): + """ + Args: + append_path: optional path to append to the tests dir path + Return: + The full path to the `tests` dir, so that the tests can be invoked from anywhere. Optionally `append_path` is + joined after the `tests` dir the former is provided. + """ + # this function caller's __file__ + caller__file__ = inspect.stack()[1][1] + tests_dir = os.path.abspath(os.path.dirname(caller__file__)) + + while not tests_dir.endswith("tests"): + tests_dir = os.path.dirname(tests_dir) + + if append_path: + return Path(tests_dir, append_path).as_posix() + else: + return tests_dir + + +# Taken from the following PR: +# https://github.com/huggingface/accelerate/pull/1964 +def str_to_bool(value) -> int: + """ + Converts a string representation of truth to `True` (1) or `False` (0). True values are `y`, `yes`, `t`, `true`, + `on`, and `1`; False value are `n`, `no`, `f`, `false`, `off`, and `0`; + """ + value = value.lower() + if value in ("y", "yes", "t", "true", "on", "1"): + return 1 + elif value in ("n", "no", "f", "false", "off", "0"): + return 0 + else: + raise ValueError(f"invalid truth value {value}") + + +def parse_flag_from_env(key, default=False): + try: + value = os.environ[key] + except KeyError: + # KEY isn't set, default to `default`. + _value = default + else: + # KEY is set, convert it to True or False. + try: + _value = str_to_bool(value) + except ValueError: + # More values are supported, but let's keep the message simple. + raise ValueError(f"If set, {key} must be yes or no.") + return _value + + +_run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False) +_run_nightly_tests = parse_flag_from_env("RUN_NIGHTLY", default=False) +_run_compile_tests = parse_flag_from_env("RUN_COMPILE", default=False) + + +def floats_tensor(shape, scale=1.0, rng=None, name=None): + """Creates a random float32 tensor""" + if rng is None: + rng = global_rng + + total_dims = 1 + for dim in shape: + total_dims *= dim + + values = [] + for _ in range(total_dims): + values.append(rng.random() * scale) + + return torch.tensor(data=values, dtype=torch.float).view(shape).contiguous() + + +def slow(test_case): + """ + Decorator marking a test as slow. + + Slow tests are skipped by default. Set the RUN_SLOW environment variable to a truthy value to run them. + + """ + return unittest.skipUnless(_run_slow_tests, "test is slow")(test_case) + + +def nightly(test_case): + """ + Decorator marking a test that runs nightly in the diffusers CI. + + Slow tests are skipped by default. Set the RUN_NIGHTLY environment variable to a truthy value to run them. + + """ + return unittest.skipUnless(_run_nightly_tests, "test is nightly")(test_case) + + +def is_torch_compile(test_case): + """ + Decorator marking a test that runs compile tests in the diffusers CI. + + Compile tests are skipped by default. Set the RUN_COMPILE environment variable to a truthy value to run them. + + """ + return unittest.skipUnless(_run_compile_tests, "test is torch compile")(test_case) + + +def require_torch(test_case): + """ + Decorator marking a test that requires PyTorch. These tests are skipped when PyTorch isn't installed. + """ + return unittest.skipUnless(is_torch_available(), "test requires PyTorch")(test_case) + + +def require_torch_2(test_case): + """ + Decorator marking a test that requires PyTorch 2. These tests are skipped when it isn't installed. + """ + return unittest.skipUnless(is_torch_available() and is_torch_version(">=", "2.0.0"), "test requires PyTorch 2")( + test_case + ) + + +def require_torch_version_greater_equal(torch_version): + """Decorator marking a test that requires torch with a specific version or greater.""" + + def decorator(test_case): + correct_torch_version = is_torch_available() and is_torch_version(">=", torch_version) + return unittest.skipUnless( + correct_torch_version, f"test requires torch with the version greater than or equal to {torch_version}" + )(test_case) + + return decorator + + +def require_torch_version_greater(torch_version): + """Decorator marking a test that requires torch with a specific version greater.""" + + def decorator(test_case): + correct_torch_version = is_torch_available() and is_torch_version(">", torch_version) + return unittest.skipUnless( + correct_torch_version, f"test requires torch with the version greater than {torch_version}" + )(test_case) + + return decorator + + +def require_torch_gpu(test_case): + """Decorator marking a test that requires CUDA and PyTorch.""" + return unittest.skipUnless(is_torch_available() and torch_device == "cuda", "test requires PyTorch+CUDA")( + test_case + ) + + +def require_torch_cuda_compatibility(expected_compute_capability): + def decorator(test_case): + if torch.cuda.is_available(): + current_compute_capability = get_torch_cuda_device_capability() + return unittest.skipUnless( + float(current_compute_capability) == float(expected_compute_capability), + "Test not supported for this compute capability.", + ) + + return decorator + + +# These decorators are for accelerator-specific behaviours that are not GPU-specific +def require_torch_accelerator(test_case): + """Decorator marking a test that requires an accelerator backend and PyTorch.""" + return unittest.skipUnless(is_torch_available() and torch_device != "cpu", "test requires accelerator+PyTorch")( + test_case + ) + + +def require_torch_multi_gpu(test_case): + """ + Decorator marking a test that requires a multi-GPU setup (in PyTorch). These tests are skipped on a machine without + multiple GPUs. To run *only* the multi_gpu tests, assuming all test names contain multi_gpu: $ pytest -sv ./tests + -k "multi_gpu" + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + return unittest.skipUnless(torch.cuda.device_count() > 1, "test requires multiple GPUs")(test_case) + + +def require_torch_multi_accelerator(test_case): + """ + Decorator marking a test that requires a multi-accelerator setup (in PyTorch). These tests are skipped on a machine + without multiple hardware accelerators. + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + return unittest.skipUnless( + torch.cuda.device_count() > 1 or torch.xpu.device_count() > 1, "test requires multiple hardware accelerators" + )(test_case) + + +def require_torch_accelerator_with_fp16(test_case): + """Decorator marking a test that requires an accelerator with support for the FP16 data type.""" + return unittest.skipUnless(_is_torch_fp16_available(torch_device), "test requires accelerator with fp16 support")( + test_case + ) + + +def require_torch_accelerator_with_fp64(test_case): + """Decorator marking a test that requires an accelerator with support for the FP64 data type.""" + return unittest.skipUnless(_is_torch_fp64_available(torch_device), "test requires accelerator with fp64 support")( + test_case + ) + + +def require_big_gpu_with_torch_cuda(test_case): + """ + Decorator marking a test that requires a bigger GPU (24GB) for execution. Some example pipelines: Flux, SD3, Cog, + etc. + """ + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + if not torch.cuda.is_available(): + return unittest.skip("test requires PyTorch CUDA")(test_case) + + device_properties = torch.cuda.get_device_properties(0) + total_memory = device_properties.total_memory / (1024**3) + return unittest.skipUnless( + total_memory >= BIG_GPU_MEMORY, f"test requires a GPU with at least {BIG_GPU_MEMORY} GB memory" + )(test_case) + + +def require_big_accelerator(test_case): + """ + Decorator marking a test that requires a bigger hardware accelerator (24GB) for execution. Some example pipelines: + Flux, SD3, Cog, etc. + """ + import pytest + + test_case = pytest.mark.big_accelerator(test_case) + + if not is_torch_available(): + return unittest.skip("test requires PyTorch")(test_case) + + import torch + + if not (torch.cuda.is_available() or torch.xpu.is_available()): + return unittest.skip("test requires PyTorch CUDA")(test_case) + + if torch.xpu.is_available(): + device_properties = torch.xpu.get_device_properties(0) + else: + device_properties = torch.cuda.get_device_properties(0) + + total_memory = device_properties.total_memory / (1024**3) + return unittest.skipUnless( + total_memory >= BIG_GPU_MEMORY, + f"test requires a hardware accelerator with at least {BIG_GPU_MEMORY} GB memory", + )(test_case) + + +def require_torch_accelerator_with_training(test_case): + """Decorator marking a test that requires an accelerator with support for training.""" + return unittest.skipUnless( + is_torch_available() and backend_supports_training(torch_device), + "test requires accelerator with training support", + )(test_case) + + +def skip_mps(test_case): + """Decorator marking a test to skip if torch_device is 'mps'""" + return unittest.skipUnless(torch_device != "mps", "test requires non 'mps' device")(test_case) + + +def require_compel(test_case): + """ + Decorator marking a test that requires compel: https://github.com/damian0815/compel. These tests are skipped when + the library is not installed. + """ + return unittest.skipUnless(is_compel_available(), "test requires compel")(test_case) + + +def require_onnxruntime(test_case): + """ + Decorator marking a test that requires onnxruntime. These tests are skipped when onnxruntime isn't installed. + """ + return unittest.skipUnless(is_onnx_available(), "test requires onnxruntime")(test_case) + + +def require_note_seq(test_case): + """ + Decorator marking a test that requires note_seq. These tests are skipped when note_seq isn't installed. + """ + return unittest.skipUnless(is_note_seq_available(), "test requires note_seq")(test_case) + + +def require_accelerator(test_case): + """ + Decorator marking a test that requires a hardware accelerator backend. These tests are skipped when there are no + hardware accelerator available. + """ + return unittest.skipUnless(torch_device != "cpu", "test requires a hardware accelerator")(test_case) + + +def require_torchsde(test_case): + """ + Decorator marking a test that requires torchsde. These tests are skipped when torchsde isn't installed. + """ + return unittest.skipUnless(is_torchsde_available(), "test requires torchsde")(test_case) + + +def require_peft_backend(test_case): + """ + Decorator marking a test that requires PEFT backend, this would require some specific versions of PEFT and + transformers. + """ + return unittest.skipUnless(USE_PEFT_BACKEND, "test requires PEFT backend")(test_case) + + +def require_timm(test_case): + """ + Decorator marking a test that requires timm. These tests are skipped when timm isn't installed. + """ + return unittest.skipUnless(is_timm_available(), "test requires timm")(test_case) + + +def require_bitsandbytes(test_case): + """ + Decorator marking a test that requires bitsandbytes. These tests are skipped when bitsandbytes isn't installed. + """ + return unittest.skipUnless(is_bitsandbytes_available(), "test requires bitsandbytes")(test_case) + + +def require_quanto(test_case): + """ + Decorator marking a test that requires quanto. These tests are skipped when quanto isn't installed. + """ + return unittest.skipUnless(is_optimum_quanto_available(), "test requires quanto")(test_case) + + +def require_sdnq(test_case): + """ + Decorator marking a test that requires sdnq. These tests are skipped when sdnq isn't installed. + """ + return unittest.skipUnless(is_sdnq_available(), "test requires sdnq")(test_case) + + +def require_accelerate(test_case): + """ + Decorator marking a test that requires accelerate. These tests are skipped when accelerate isn't installed. + """ + return unittest.skipUnless(is_accelerate_available(), "test requires accelerate")(test_case) + + +def require_peft_version_greater(peft_version): + """ + Decorator marking a test that requires PEFT backend with a specific version, this would require some specific + versions of PEFT and transformers. + """ + + def decorator(test_case): + correct_peft_version = is_peft_available() and version.parse( + version.parse(importlib.metadata.version("peft")).base_version + ) > version.parse(peft_version) + return unittest.skipUnless( + correct_peft_version, f"test requires PEFT backend with the version greater than {peft_version}" + )(test_case) + + return decorator + + +def require_transformers_version_greater(transformers_version): + """ + Decorator marking a test that requires transformers with a specific version, this would require some specific + versions of PEFT and transformers. + """ + + def decorator(test_case): + correct_transformers_version = is_transformers_available() and version.parse( + version.parse(importlib.metadata.version("transformers")).base_version + ) > version.parse(transformers_version) + return unittest.skipUnless( + correct_transformers_version, + f"test requires transformers with the version greater than {transformers_version}", + )(test_case) + + return decorator + + +def require_accelerate_version_greater(accelerate_version): + def decorator(test_case): + correct_accelerate_version = is_accelerate_available() and version.parse( + version.parse(importlib.metadata.version("accelerate")).base_version + ) > version.parse(accelerate_version) + return unittest.skipUnless( + correct_accelerate_version, f"Test requires accelerate with the version greater than {accelerate_version}." + )(test_case) + + return decorator + + +def require_bitsandbytes_version_greater(bnb_version): + def decorator(test_case): + correct_bnb_version = is_bitsandbytes_available() and version.parse( + version.parse(importlib.metadata.version("bitsandbytes")).base_version + ) > version.parse(bnb_version) + return unittest.skipUnless( + correct_bnb_version, f"Test requires bitsandbytes with the version greater than {bnb_version}." + )(test_case) + + return decorator + + +def require_hf_hub_version_greater(hf_hub_version): + def decorator(test_case): + correct_hf_hub_version = version.parse( + version.parse(importlib.metadata.version("huggingface_hub")).base_version + ) > version.parse(hf_hub_version) + return unittest.skipUnless( + correct_hf_hub_version, f"Test requires huggingface_hub with the version greater than {hf_hub_version}." + )(test_case) + + return decorator + + +def require_gguf_version_greater_or_equal(gguf_version): + def decorator(test_case): + correct_gguf_version = is_gguf_available() and version.parse( + version.parse(importlib.metadata.version("gguf")).base_version + ) >= version.parse(gguf_version) + return unittest.skipUnless( + correct_gguf_version, f"Test requires gguf with the version greater than {gguf_version}." + )(test_case) + + return decorator + + +def require_torchao_version_greater_or_equal(torchao_version): + def decorator(test_case): + correct_torchao_version = is_torchao_available() and version.parse( + version.parse(importlib.metadata.version("torchao")).base_version + ) >= version.parse(torchao_version) + return unittest.skipUnless( + correct_torchao_version, f"Test requires torchao with version greater than {torchao_version}." + )(test_case) + + return decorator + + +def require_modelopt_version_greater_or_equal(modelopt_version): + def decorator(test_case): + correct_nvidia_modelopt_version = is_nvidia_modelopt_available() and version.parse( + version.parse(importlib.metadata.version("nvidia-modelopt")).base_version + ) >= version.parse(modelopt_version) + return unittest.skipUnless( + correct_nvidia_modelopt_version, f"Test requires modelopt with version greater than {modelopt_version}." + )(test_case) + + return decorator + + +def require_kernels_version_greater_or_equal(kernels_version): + def decorator(test_case): + correct_kernels_version = is_kernels_available() and version.parse( + version.parse(importlib.metadata.version("kernels")).base_version + ) >= version.parse(kernels_version) + return unittest.skipUnless( + correct_kernels_version, f"Test requires kernels with version greater than {kernels_version}." + )(test_case) + + return decorator + + +def deprecate_after_peft_backend(test_case): + """ + Decorator marking a test that will be skipped after PEFT backend + """ + return unittest.skipUnless(not USE_PEFT_BACKEND, "test skipped in favor of PEFT backend")(test_case) + + +def get_python_version(): + sys_info = sys.version_info + major, minor = sys_info.major, sys_info.minor + return major, minor + + +def load_numpy(arry: str | np.ndarray, local_path: str | None = None) -> np.ndarray: + if isinstance(arry, str): + if local_path is not None: + # local_path can be passed to correct images of tests + return Path(local_path, arry.split("/")[-5], arry.split("/")[-2], arry.split("/")[-1]).as_posix() + elif arry.startswith("http://") or arry.startswith("https://"): + response = requests.get(arry, timeout=DIFFUSERS_REQUEST_TIMEOUT) + response.raise_for_status() + arry = np.load(BytesIO(response.content)) + elif os.path.isfile(arry): + arry = np.load(arry) + else: + raise ValueError( + f"Incorrect path or url, URLs must start with `http://` or `https://`, and {arry} is not a valid path" + ) + elif isinstance(arry, np.ndarray): + pass + else: + raise ValueError( + "Incorrect format used for numpy ndarray. Should be an url linking to an image, a local path, or a" + " ndarray." + ) + + return arry + + +def load_pt(url: str, map_location: str | None = None, weights_only: bool | None = True): + response = requests.get(url, timeout=DIFFUSERS_REQUEST_TIMEOUT) + response.raise_for_status() + arry = torch.load(BytesIO(response.content), map_location=map_location, weights_only=weights_only) + return arry + + +def load_image(image: str | PIL.Image.Image) -> PIL.Image.Image: + """ + Loads `image` to a PIL Image. + + Args: + image (`str` or `PIL.Image.Image`): + The image to convert to the PIL Image format. + Returns: + `PIL.Image.Image`: + A PIL Image. + """ + if isinstance(image, str): + if image.startswith("http://") or image.startswith("https://"): + image = PIL.Image.open(requests.get(image, stream=True, timeout=DIFFUSERS_REQUEST_TIMEOUT).raw) + elif os.path.isfile(image): + image = PIL.Image.open(image) + else: + raise ValueError( + f"Incorrect path or url, URLs must start with `http://` or `https://`, and {image} is not a valid path" + ) + elif isinstance(image, PIL.Image.Image): + image = image + else: + raise ValueError( + "Incorrect format used for image. Should be an url linking to an image, a local path, or a PIL image." + ) + image = PIL.ImageOps.exif_transpose(image) + image = image.convert("RGB") + return image + + +def preprocess_image(image: PIL.Image, batch_size: int): + w, h = image.size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + image = image.resize((w, h), resample=PIL.Image.LANCZOS) + image = np.array(image).astype(np.float32) / 255.0 + image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def export_to_gif(image: list[PIL.Image.Image], output_gif_path: str = None) -> str: + if output_gif_path is None: + output_gif_path = tempfile.NamedTemporaryFile(suffix=".gif").name + + image[0].save( + output_gif_path, + save_all=True, + append_images=image[1:], + optimize=False, + duration=100, + loop=0, + ) + return output_gif_path + + +@contextmanager +def buffered_writer(raw_f): + f = io.BufferedWriter(raw_f) + yield f + f.flush() + + +def export_to_ply(mesh, output_ply_path: str = None): + """ + Write a PLY file for a mesh. + """ + if output_ply_path is None: + output_ply_path = tempfile.NamedTemporaryFile(suffix=".ply").name + + coords = mesh.verts.detach().cpu().numpy() + faces = mesh.faces.cpu().numpy() + rgb = np.stack([mesh.vertex_channels[x].detach().cpu().numpy() for x in "RGB"], axis=1) + + with buffered_writer(open(output_ply_path, "wb")) as f: + f.write(b"ply\n") + f.write(b"format binary_little_endian 1.0\n") + f.write(bytes(f"element vertex {len(coords)}\n", "ascii")) + f.write(b"property float x\n") + f.write(b"property float y\n") + f.write(b"property float z\n") + if rgb is not None: + f.write(b"property uchar red\n") + f.write(b"property uchar green\n") + f.write(b"property uchar blue\n") + if faces is not None: + f.write(bytes(f"element face {len(faces)}\n", "ascii")) + f.write(b"property list uchar int vertex_index\n") + f.write(b"end_header\n") + + if rgb is not None: + rgb = (rgb * 255.499).round().astype(int) + vertices = [ + (*coord, *rgb) + for coord, rgb in zip( + coords.tolist(), + rgb.tolist(), + ) + ] + format = struct.Struct("<3f3B") + for item in vertices: + f.write(format.pack(*item)) + else: + format = struct.Struct("<3f") + for vertex in coords.tolist(): + f.write(format.pack(*vertex)) + + if faces is not None: + for tri in faces.tolist(): + f.write(format.pack(len(tri), *tri)) + format = struct.Struct(" str: + if is_opencv_available(): + import cv2 + else: + raise ImportError(BACKENDS_MAPPING["opencv"][1].format("export_to_video")) + if output_video_path is None: + output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name + + fourcc = cv2.VideoWriter_fourcc(*"mp4v") + h, w, c = video_frames[0].shape + video_writer = cv2.VideoWriter(output_video_path, fourcc, fps=8, frameSize=(w, h)) + for i in range(len(video_frames)): + img = cv2.cvtColor(video_frames[i], cv2.COLOR_RGB2BGR) + video_writer.write(img) + return output_video_path + + +def load_hf_numpy(path) -> np.ndarray: + base_url = "https://huggingface.co/datasets/fusing/diffusers-testing/resolve/main" + + if not path.startswith("http://") and not path.startswith("https://"): + path = os.path.join(base_url, urllib.parse.quote(path)) + + return load_numpy(path) + + +# --- pytest conf functions --- # + +# to avoid multiple invocation from tests/conftest.py and examples/conftest.py - make sure it's called only once +pytest_opt_registered = {} + + +def pytest_addoption_shared(parser): + """ + This function is to be called from `conftest.py` via `pytest_addoption` wrapper that has to be defined there. + + It allows loading both `conftest.py` files at once without causing a failure due to adding the same `pytest` + option. + + """ + option = "--make-reports" + if option not in pytest_opt_registered: + parser.addoption( + option, + action="store", + default=False, + help="generate report files. The value of this option is used as a prefix to report names", + ) + pytest_opt_registered[option] = 1 + + +def pytest_terminal_summary_main(tr, id): + """ + Generate multiple reports at the end of test suite run - each report goes into a dedicated file in the current + directory. The report files are prefixed with the test suite name. + + This function emulates --duration and -rA pytest arguments. + + This function is to be called from `conftest.py` via `pytest_terminal_summary` wrapper that has to be defined + there. + + Args: + - tr: `terminalreporter` passed from `conftest.py` + - id: unique id like `tests` or `examples` that will be incorporated into the final reports filenames - this is + needed as some jobs have multiple runs of pytest, so we can't have them overwrite each other. + + NB: this functions taps into a private _pytest API and while unlikely, it could break should + pytest do internal changes - also it calls default internal methods of terminalreporter which + can be hijacked by various `pytest-` plugins and interfere. + + """ + from _pytest.config import create_terminal_writer + + if not len(id): + id = "tests" + + config = tr.config + orig_writer = config.get_terminal_writer() + orig_tbstyle = config.option.tbstyle + orig_reportchars = tr.reportchars + + dir = "reports" + Path(dir).mkdir(parents=True, exist_ok=True) + report_files = { + k: f"{dir}/{id}_{k}.txt" + for k in [ + "durations", + "errors", + "failures_long", + "failures_short", + "failures_line", + "passes", + "stats", + "summary_short", + "warnings", + ] + } + + # custom durations report + # note: there is no need to call pytest --durations=XX to get this separate report + # adapted from https://github.com/pytest-dev/pytest/blob/897f151e/src/_pytest/runner.py#L66 + dlist = [] + for replist in tr.stats.values(): + for rep in replist: + if hasattr(rep, "duration"): + dlist.append(rep) + if dlist: + dlist.sort(key=lambda x: x.duration, reverse=True) + with open(report_files["durations"], "w") as f: + durations_min = 0.05 # sec + f.write("slowest durations\n") + for i, rep in enumerate(dlist): + if rep.duration < durations_min: + f.write(f"{len(dlist) - i} durations < {durations_min} secs were omitted") + break + f.write(f"{rep.duration:02.2f}s {rep.when:<8} {rep.nodeid}\n") + + def summary_failures_short(tr): + # expecting that the reports were --tb=long (default) so we chop them off here to the last frame + reports = tr.getreports("failed") + if not reports: + return + tr.write_sep("=", "FAILURES SHORT STACK") + for rep in reports: + msg = tr._getfailureheadline(rep) + tr.write_sep("_", msg, red=True, bold=True) + # chop off the optional leading extra frames, leaving only the last one + longrepr = re.sub(r".*_ _ _ (_ ){10,}_ _ ", "", rep.longreprtext, 0, re.M | re.S) + tr._tw.line(longrepr) + # note: not printing out any rep.sections to keep the report short + + # use ready-made report funcs, we are just hijacking the filehandle to log to a dedicated file each + # adapted from https://github.com/pytest-dev/pytest/blob/897f151e/src/_pytest/terminal.py#L814 + # note: some pytest plugins may interfere by hijacking the default `terminalreporter` (e.g. + # pytest-instafail does that) + + # report failures with line/short/long styles + config.option.tbstyle = "auto" # full tb + with open(report_files["failures_long"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_failures() + + # config.option.tbstyle = "short" # short tb + with open(report_files["failures_short"], "w") as f: + tr._tw = create_terminal_writer(config, f) + summary_failures_short(tr) + + config.option.tbstyle = "line" # one line per error + with open(report_files["failures_line"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_failures() + + with open(report_files["errors"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_errors() + + with open(report_files["warnings"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_warnings() # normal warnings + tr.summary_warnings() # final warnings + + tr.reportchars = "wPpsxXEf" # emulate -rA (used in summary_passes() and short_test_summary()) + with open(report_files["passes"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_passes() + + with open(report_files["summary_short"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.short_test_summary() + + with open(report_files["stats"], "w") as f: + tr._tw = create_terminal_writer(config, f) + tr.summary_stats() + + # restore: + tr._tw = orig_writer + tr.reportchars = orig_reportchars + config.option.tbstyle = orig_tbstyle + + +# Adapted from https://github.com/huggingface/transformers/blob/000e52aec8850d3fe2f360adc6fd256e5b47fe4c/src/transformers/testing_utils.py#L1905 +def is_flaky(max_attempts: int = 5, wait_before_retry: float | None = None, description: str | None = None): + """ + To decorate flaky tests (methods or entire classes). They will be retried on failures. + + Args: + max_attempts (`int`, *optional*, defaults to 5): + The maximum number of attempts to retry the flaky test. + wait_before_retry (`float`, *optional*): + If provided, will wait that number of seconds before retrying the test. + description (`str`, *optional*): + A string to describe the situation (what / where / why is flaky, link to GH issue/PR comments, errors, + etc.) + """ + + def decorator(obj): + # If decorating a class, wrap each test method on it + if inspect.isclass(obj): + for attr_name, attr_value in list(obj.__dict__.items()): + if callable(attr_value) and attr_name.startswith("test"): + # recursively decorate the method + setattr(obj, attr_name, decorator(attr_value)) + return obj + + # Otherwise we're decorating a single test function / method + @functools.wraps(obj) + def wrapper(*args, **kwargs): + retry_count = 1 + while retry_count < max_attempts: + try: + return obj(*args, **kwargs) + except Exception as err: + msg = ( + f"[FLAKY] {description or obj.__name__!r} " + f"failed on attempt {retry_count}/{max_attempts}: {err}" + ) + print(msg, file=sys.stderr) + if wait_before_retry is not None: + time.sleep(wait_before_retry) + retry_count += 1 + + return obj(*args, **kwargs) + + return wrapper + + return decorator + + +# Taken from: https://github.com/huggingface/transformers/blob/3658488ff77ff8d45101293e749263acf437f4d5/src/transformers/testing_utils.py#L1787 +def run_test_in_subprocess(test_case, target_func, inputs=None, timeout=None): + """ + To run a test in a subprocess. In particular, this can avoid (GPU) memory issue. + + Args: + test_case (`unittest.TestCase`): + The test that will run `target_func`. + target_func (`Callable`): + The function implementing the actual testing logic. + inputs (`dict`, *optional*, defaults to `None`): + The inputs that will be passed to `target_func` through an (input) queue. + timeout (`int`, *optional*, defaults to `None`): + The timeout (in seconds) that will be passed to the input and output queues. If not specified, the env. + variable `PYTEST_TIMEOUT` will be checked. If still `None`, its value will be set to `600`. + """ + if timeout is None: + timeout = int(os.environ.get("PYTEST_TIMEOUT", 600)) + + start_methohd = "spawn" + ctx = multiprocessing.get_context(start_methohd) + + input_queue = ctx.Queue(1) + output_queue = ctx.JoinableQueue(1) + + # We can't send `unittest.TestCase` to the child, otherwise we get issues regarding pickle. + input_queue.put(inputs, timeout=timeout) + + process = ctx.Process(target=target_func, args=(input_queue, output_queue, timeout)) + process.start() + # Kill the child process if we can't get outputs from it in time: otherwise, the hanging subprocess prevents + # the test to exit properly. + try: + results = output_queue.get(timeout=timeout) + output_queue.task_done() + except Exception as e: + process.terminate() + test_case.fail(e) + process.join(timeout=timeout) + + if results["error"] is not None: + test_case.fail(f"{results['error']}") + + +class CaptureLogger: + """ + Args: + Context manager to capture `logging` streams + logger: 'logging` logger object + Returns: + The captured output is available via `self.out` + Example: + ```python + >>> from diffusers import logging + >>> from diffusers.testing_utils import CaptureLogger + + >>> msg = "Testing 1, 2, 3" + >>> logging.set_verbosity_info() + >>> logger = logging.get_logger("diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.py") + >>> with CaptureLogger(logger) as cl: + ... logger.info(msg) + >>> assert cl.out, msg + "\n" + ``` + """ + + def __init__(self, logger): + self.logger = logger + self.io = StringIO() + self.sh = logging.StreamHandler(self.io) + self.out = "" + + def __enter__(self): + self.logger.addHandler(self.sh) + return self + + def __exit__(self, *exc): + self.logger.removeHandler(self.sh) + self.out = self.io.getvalue() + + def __repr__(self): + return f"captured: {self.out}\n" + + +def enable_full_determinism(): + """ + Helper function for reproducible behavior during distributed training. See + - https://pytorch.org/docs/stable/notes/randomness.html for pytorch + """ + from .torch_utils import enable_full_determinism as _enable_full_determinism + + logger.warning( + "enable_full_determinism has been moved to diffusers.utils.torch_utils. " + "Importing from diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _enable_full_determinism() + + +def disable_full_determinism(): + from .torch_utils import disable_full_determinism as _disable_full_determinism + + logger.warning( + "disable_full_determinism has been moved to diffusers.utils.torch_utils. " + "Importing from diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _disable_full_determinism() + + +# Utils for custom and alternative accelerator devices +def _is_torch_fp16_available(device): + if not is_torch_available(): + return False + + import torch + + device = torch.device(device) + + try: + x = torch.zeros((2, 2), dtype=torch.float16).to(device) + _ = torch.mul(x, x) + return True + + except Exception as e: + if device.type == "cuda": + raise ValueError( + f"You have passed a device of type 'cuda' which should work with 'fp16', but 'cuda' does not seem to be correctly installed on your machine: {e}" + ) + + return False + + +def _is_torch_fp64_available(device): + if not is_torch_available(): + return False + + import torch + + device = torch.device(device) + + try: + x = torch.zeros((2, 2), dtype=torch.float64).to(device) + _ = torch.mul(x, x) + return True + + except Exception as e: + if device.type == "cuda": + raise ValueError( + f"You have passed a device of type 'cuda' which should work with 'fp64', but 'cuda' does not seem to be correctly installed on your machine: {e}" + ) + + return False + + +# Guard these lookups for when Torch is not used - alternative accelerator support is for PyTorch +if is_torch_available(): + # Behaviour flags + BACKEND_SUPPORTS_TRAINING = {"cuda": True, "xpu": True, "cpu": True, "mps": False, "default": True} + + # Function definitions + BACKEND_EMPTY_CACHE = { + "cuda": torch.cuda.empty_cache, + "xpu": torch.xpu.empty_cache, + "cpu": None, + "mps": torch.mps.empty_cache, + "default": None, + } + BACKEND_DEVICE_COUNT = { + "cuda": torch.cuda.device_count, + "xpu": torch.xpu.device_count, + "cpu": lambda: 0, + "mps": lambda: 0, + "default": 0, + } + BACKEND_MANUAL_SEED = { + "cuda": torch.cuda.manual_seed, + "xpu": torch.xpu.manual_seed, + "cpu": torch.manual_seed, + "mps": torch.mps.manual_seed, + "default": torch.manual_seed, + } + BACKEND_RESET_PEAK_MEMORY_STATS = { + "cuda": torch.cuda.reset_peak_memory_stats, + "xpu": getattr(torch.xpu, "reset_peak_memory_stats", None), + "cpu": None, + "mps": None, + "default": None, + } + BACKEND_RESET_MAX_MEMORY_ALLOCATED = { + "cuda": torch.cuda.reset_max_memory_allocated, + "xpu": getattr(torch.xpu, "reset_peak_memory_stats", None), + "cpu": None, + "mps": None, + "default": None, + } + BACKEND_MAX_MEMORY_ALLOCATED = { + "cuda": torch.cuda.max_memory_allocated, + "xpu": getattr(torch.xpu, "max_memory_allocated", None), + "cpu": 0, + "mps": 0, + "default": 0, + } + BACKEND_SYNCHRONIZE = { + "cuda": torch.cuda.synchronize, + "xpu": getattr(torch.xpu, "synchronize", None), + "cpu": None, + "mps": None, + "default": None, + } + + +# This dispatches a defined function according to the accelerator from the function definitions. +def _device_agnostic_dispatch(device: str, dispatch_table: dict[str, Callable], *args, **kwargs): + if device not in dispatch_table: + return dispatch_table["default"](*args, **kwargs) + + fn = dispatch_table[device] + + # Some device agnostic functions return values. Need to guard against 'None' instead at + # user level + if not callable(fn): + return fn + + return fn(*args, **kwargs) + + +# These are callables which automatically dispatch the function specific to the accelerator +def backend_manual_seed(device: str, seed: int): + from .torch_utils import backend_manual_seed as _backend_manual_seed + + logger.warning( + "backend_manual_seed has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_manual_seed(device, seed) + + +def backend_synchronize(device: str): + from .torch_utils import backend_synchronize as _backend_synchronize + + logger.warning( + "backend_synchronize has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_synchronize(device) + + +def backend_empty_cache(device: str): + from .torch_utils import backend_empty_cache as _backend_empty_cache + + logger.warning( + "backend_empty_cache has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_empty_cache(device) + + +def backend_device_count(device: str): + from .torch_utils import backend_device_count as _backend_device_count + + logger.warning( + "backend_device_count has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_device_count(device) + + +def backend_reset_peak_memory_stats(device: str): + from .torch_utils import backend_reset_peak_memory_stats as _backend_reset_peak_memory_stats + + logger.warning( + "backend_reset_peak_memory_stats has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_reset_peak_memory_stats(device) + + +def backend_reset_max_memory_allocated(device: str): + from .torch_utils import backend_reset_max_memory_allocated as _backend_reset_max_memory_allocated + + logger.warning( + "backend_reset_max_memory_allocated has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_reset_max_memory_allocated(device) + + +def backend_max_memory_allocated(device: str): + from .torch_utils import backend_max_memory_allocated as _backend_max_memory_allocated + + logger.warning( + "backend_max_memory_allocated has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_max_memory_allocated(device) + + +# These are callables which return boolean behaviour flags and can be used to specify some +# device agnostic alternative where the feature is unsupported. +def backend_supports_training(device: str): + from .torch_utils import backend_supports_training as _backend_supports_training + + logger.warning( + "backend_supports_training has been moved to diffusers.utils.torch_utils. " + "diffusers.utils.testing_utils is deprecated and will be removed in a future version." + ) + return _backend_supports_training(device) + + +# Guard for when Torch is not available +if is_torch_available(): + # Update device function dict mapping + def update_mapping_from_spec(device_fn_dict: dict[str, Callable], attribute_name: str): + try: + # Try to import the function directly + spec_fn = getattr(device_spec_module, attribute_name) + device_fn_dict[torch_device] = spec_fn + except AttributeError as e: + # If the function doesn't exist, and there is no default, throw an error + if "default" not in device_fn_dict: + raise AttributeError( + f"`{attribute_name}` not found in '{device_spec_path}' and no default fallback function found." + ) from e + + if "DIFFUSERS_TEST_DEVICE_SPEC" in os.environ: + device_spec_path = os.environ["DIFFUSERS_TEST_DEVICE_SPEC"] + if not Path(device_spec_path).is_file(): + raise ValueError(f"Specified path to device specification file is not found. Received {device_spec_path}") + + try: + import_name = device_spec_path[: device_spec_path.index(".py")] + except ValueError as e: + raise ValueError(f"Provided device spec file is not a Python file! Received {device_spec_path}") from e + + device_spec_module = importlib.import_module(import_name) + + try: + device_name = device_spec_module.DEVICE_NAME + except AttributeError: + raise AttributeError("Device spec file did not contain `DEVICE_NAME`") + + if "DIFFUSERS_TEST_DEVICE" in os.environ and torch_device != device_name: + msg = f"Mismatch between environment variable `DIFFUSERS_TEST_DEVICE` '{torch_device}' and device found in spec '{device_name}'\n" + msg += "Either unset `DIFFUSERS_TEST_DEVICE` or ensure it matches device spec name." + raise ValueError(msg) + + torch_device = device_name + + # Add one entry here for each `BACKEND_*` dictionary. + update_mapping_from_spec(BACKEND_MANUAL_SEED, "MANUAL_SEED_FN") + update_mapping_from_spec(BACKEND_EMPTY_CACHE, "EMPTY_CACHE_FN") + update_mapping_from_spec(BACKEND_DEVICE_COUNT, "DEVICE_COUNT_FN") + update_mapping_from_spec(BACKEND_SUPPORTS_TRAINING, "SUPPORTS_TRAINING") + update_mapping_from_spec(BACKEND_RESET_PEAK_MEMORY_STATS, "RESET_PEAK_MEMORY_STATS_FN") + update_mapping_from_spec(BACKEND_RESET_MAX_MEMORY_ALLOCATED, "RESET_MAX_MEMORY_ALLOCATED_FN") + update_mapping_from_spec(BACKEND_MAX_MEMORY_ALLOCATED, "MAX_MEMORY_ALLOCATED_FN") + + +# Modified from https://github.com/huggingface/transformers/blob/cdfb018d0300fef3b07d9220f3efe9c2a9974662/src/transformers/testing_utils.py#L3090 + +# Type definition of key used in `Expectations` class. +DeviceProperties = tuple[str | None, int | None] + + +@functools.lru_cache +def get_device_properties() -> DeviceProperties: + """ + Get environment device properties. + """ + if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM: + import torch + + major, _ = torch.cuda.get_device_capability() + if IS_ROCM_SYSTEM: + return ("rocm", major) + else: + return ("cuda", major) + elif IS_XPU_SYSTEM: + import torch + + # To get more info of the architecture meaning and bit allocation, refer to https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/ext/oneapi/experimental/device_architecture.def + arch = torch.xpu.get_device_capability()["architecture"] + gen_mask = 0x000000FF00000000 + gen = (arch & gen_mask) >> 32 + return ("xpu", gen) + else: + return (torch_device, None) + + +if TYPE_CHECKING: + DevicePropertiesUserDict = UserDict[DeviceProperties, Any] +else: + DevicePropertiesUserDict = UserDict + +if is_torch_available(): + from diffusers.hooks._common import _GO_LC_SUPPORTED_PYTORCH_LAYERS + from diffusers.hooks.group_offloading import ( + _GROUP_ID_LAZY_LEAF, + _compute_group_hash, + _find_parent_module_in_module_dict, + _gather_buffers_with_no_group_offloading_parent, + _gather_parameters_with_no_group_offloading_parent, + ) + + def _get_expected_safetensors_files( + module: torch.nn.Module, + offload_to_disk_path: str, + offload_type: str, + num_blocks_per_group: int | None = None, + ) -> Set[str]: + expected_files = set() + + def get_hashed_filename(group_id: str) -> str: + short_hash = _compute_group_hash(group_id) + return os.path.join(offload_to_disk_path, f"group_{short_hash}.safetensors") + + if offload_type == "block_level": + if num_blocks_per_group is None: + raise ValueError("num_blocks_per_group must be provided for 'block_level' offloading.") + + # Handle groups of ModuleList and Sequential blocks + unmatched_modules = [] + for name, submodule in module.named_children(): + if not isinstance(submodule, (torch.nn.ModuleList, torch.nn.Sequential)): + unmatched_modules.append(module) + continue + + for i in range(0, len(submodule), num_blocks_per_group): + current_modules = submodule[i : i + num_blocks_per_group] + if not current_modules: + continue + group_id = f"{name}_{i}_{i + len(current_modules) - 1}" + expected_files.add(get_hashed_filename(group_id)) + + # Handle the group for unmatched top-level modules and parameters + for module in unmatched_modules: + expected_files.add(get_hashed_filename(f"{module.__class__.__name__}_unmatched_group")) + + elif offload_type == "leaf_level": + # Handle leaf-level module groups + for name, submodule in module.named_modules(): + if isinstance(submodule, _GO_LC_SUPPORTED_PYTORCH_LAYERS): + # These groups will always have parameters, so a file is expected + expected_files.add(get_hashed_filename(name)) + + # Handle groups for non-leaf parameters/buffers + modules_with_group_offloading = { + name for name, sm in module.named_modules() if isinstance(sm, _GO_LC_SUPPORTED_PYTORCH_LAYERS) + } + parameters = _gather_parameters_with_no_group_offloading_parent(module, modules_with_group_offloading) + buffers = _gather_buffers_with_no_group_offloading_parent(module, modules_with_group_offloading) + + all_orphans = parameters + buffers + if all_orphans: + parent_to_tensors = {} + module_dict = dict(module.named_modules()) + for tensor_name, _ in all_orphans: + parent_name = _find_parent_module_in_module_dict(tensor_name, module_dict) + if parent_name not in parent_to_tensors: + parent_to_tensors[parent_name] = [] + parent_to_tensors[parent_name].append(tensor_name) + + for parent_name in parent_to_tensors: + # A file is expected for each parent that gathers orphaned tensors + expected_files.add(get_hashed_filename(parent_name)) + expected_files.add(get_hashed_filename(_GROUP_ID_LAZY_LEAF)) + + else: + raise ValueError(f"Unsupported offload_type: {offload_type}") + + return expected_files + + def _check_safetensors_serialization( + module: torch.nn.Module, + offload_to_disk_path: str, + offload_type: str, + num_blocks_per_group: int | None = None, + ) -> bool: + if not os.path.isdir(offload_to_disk_path): + return False, None, None + + expected_files = _get_expected_safetensors_files( + module, offload_to_disk_path, offload_type, num_blocks_per_group + ) + actual_files = set(glob.glob(os.path.join(offload_to_disk_path, "*.safetensors"))) + missing_files = expected_files - actual_files + extra_files = actual_files - expected_files + + is_correct = not missing_files and not extra_files + return is_correct, extra_files, missing_files + + +class Expectations(DevicePropertiesUserDict): + def get_expectation(self) -> Any: + """ + Find best matching expectation based on environment device properties. + """ + return self.find_expectation(get_device_properties()) + + @staticmethod + def is_default(key: DeviceProperties) -> bool: + return all(p is None for p in key) + + @staticmethod + def score(key: DeviceProperties, other: DeviceProperties) -> int: + """ + Returns score indicating how similar two instances of the `Properties` tuple are. Points are calculated using + bits, but documented as int. Rules are as follows: + * Matching `type` gives 8 points. + * Semi-matching `type`, for example cuda and rocm, gives 4 points. + * Matching `major` (compute capability major version) gives 2 points. + * Default expectation (if present) gives 1 points. + """ + (device_type, major) = key + (other_device_type, other_major) = other + + score = 0b0 + if device_type == other_device_type: + score |= 0b1000 + elif device_type in ["cuda", "rocm"] and other_device_type in ["cuda", "rocm"]: + score |= 0b100 + + if major == other_major and other_major is not None: + score |= 0b10 + + if Expectations.is_default(other): + score |= 0b1 + + return int(score) + + def find_expectation(self, key: DeviceProperties = (None, None)) -> Any: + """ + Find best matching expectation based on provided device properties. + """ + (result_key, result) = max(self.data.items(), key=lambda x: Expectations.score(key, x[0])) + + if Expectations.score(key, result_key) == 0: + raise ValueError(f"No matching expectation found for {key}") + + return result + + def __repr__(self): + return f"{self.data}" diff --git a/diffusers/utils/torch_utils.py b/diffusers/utils/torch_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..981e05a7c055774d09c63f546b68acb79210639d --- /dev/null +++ b/diffusers/utils/torch_utils.py @@ -0,0 +1,409 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +PyTorch utilities: Utilities related to PyTorch +""" + +from __future__ import annotations + +import functools +import os +from typing import Callable, ParamSpec, TypeVar + +from . import logging +from .import_utils import ( + is_torch_available, + is_torch_mlu_available, + is_torch_neuronx_available, + is_torch_npu_available, + is_torch_version, +) + + +T = TypeVar("T") +P = ParamSpec("P") + + +if is_torch_available(): + import torch + from torch.fft import fftn, fftshift, ifftn, ifftshift + + BACKEND_SUPPORTS_TRAINING = { + "cuda": True, + "xpu": True, + "cpu": True, + "mps": False, + "neuron": False, + "default": True, + } + BACKEND_EMPTY_CACHE = { + "cuda": torch.cuda.empty_cache, + "xpu": torch.xpu.empty_cache, + "cpu": None, + "mps": torch.mps.empty_cache, + "neuron": None, + "default": None, + } + BACKEND_DEVICE_COUNT = { + "cuda": torch.cuda.device_count, + "xpu": torch.xpu.device_count, + "cpu": lambda: 0, + "mps": lambda: 0, + "neuron": lambda: getattr(getattr(torch, "neuron", None), "device_count", lambda: 0)(), + "default": 0, + } + BACKEND_MANUAL_SEED = { + "cuda": torch.cuda.manual_seed, + "xpu": torch.xpu.manual_seed, + "cpu": torch.manual_seed, + "mps": torch.mps.manual_seed, + "neuron": torch.manual_seed, + "default": torch.manual_seed, + } + BACKEND_RESET_PEAK_MEMORY_STATS = { + "cuda": torch.cuda.reset_peak_memory_stats, + "xpu": getattr(torch.xpu, "reset_peak_memory_stats", None), + "cpu": None, + "mps": None, + "neuron": None, + "default": None, + } + BACKEND_RESET_MAX_MEMORY_ALLOCATED = { + "cuda": torch.cuda.reset_max_memory_allocated, + "xpu": getattr(torch.xpu, "reset_peak_memory_stats", None), + "cpu": None, + "mps": None, + "neuron": None, + "default": None, + } + BACKEND_MAX_MEMORY_ALLOCATED = { + "cuda": torch.cuda.max_memory_allocated, + "xpu": getattr(torch.xpu, "max_memory_allocated", None), + "cpu": 0, + "mps": 0, + "neuron": 0, + "default": 0, + } + BACKEND_SYNCHRONIZE = { + "cuda": torch.cuda.synchronize, + "xpu": getattr(torch.xpu, "synchronize", None), + "cpu": None, + "mps": None, + "neuron": getattr(getattr(torch, "neuron", None), "synchronize", None), + "default": None, + } + + _FP64_UNSUPPORTED_DEVICES = frozenset({"mps", "npu", "neuron"}) + _INT64_UNSUPPORTED_DEVICES = frozenset({"mps", "npu", "neuron"}) + _DTYPE_DOWNCAST = {torch.float64: torch.float32, torch.int64: torch.int32} + _DTYPE_UNSUPPORTED_DEVICES = {torch.float64: _FP64_UNSUPPORTED_DEVICES, torch.int64: _INT64_UNSUPPORTED_DEVICES} + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +try: + from torch._dynamo import allow_in_graph as maybe_allow_in_graph +except (ImportError, ModuleNotFoundError): + + def maybe_allow_in_graph(cls): + return cls + + +# This dispatches a defined function according to the accelerator from the function definitions. +def _device_agnostic_dispatch(device: str, dispatch_table: dict[str, callable], *args, **kwargs): + if device not in dispatch_table: + return dispatch_table["default"](*args, **kwargs) + + fn = dispatch_table[device] + + # Some device agnostic functions return values. Need to guard against 'None' instead at + # user level + if not callable(fn): + return fn + + return fn(*args, **kwargs) + + +# These are callables which automatically dispatch the function specific to the accelerator +def backend_manual_seed(device: str, seed: int): + return _device_agnostic_dispatch(device, BACKEND_MANUAL_SEED, seed) + + +def backend_synchronize(device: str): + return _device_agnostic_dispatch(device, BACKEND_SYNCHRONIZE) + + +def backend_empty_cache(device: str): + return _device_agnostic_dispatch(device, BACKEND_EMPTY_CACHE) + + +def backend_device_count(device: str): + return _device_agnostic_dispatch(device, BACKEND_DEVICE_COUNT) + + +def backend_reset_peak_memory_stats(device: str): + return _device_agnostic_dispatch(device, BACKEND_RESET_PEAK_MEMORY_STATS) + + +def backend_reset_max_memory_allocated(device: str): + return _device_agnostic_dispatch(device, BACKEND_RESET_MAX_MEMORY_ALLOCATED) + + +def backend_max_memory_allocated(device: str): + return _device_agnostic_dispatch(device, BACKEND_MAX_MEMORY_ALLOCATED) + + +# These are callables which return boolean behaviour flags and can be used to specify some +# device agnostic alternative where the feature is unsupported. +def backend_supports_training(device: str): + if not is_torch_available(): + return False + + if device not in BACKEND_SUPPORTS_TRAINING: + device = "default" + + return BACKEND_SUPPORTS_TRAINING[device] + + +def maybe_adjust_dtype_for_device(dtype: "torch.dtype", device: "torch.device") -> "torch.dtype": + unsupported = _DTYPE_UNSUPPORTED_DEVICES.get(dtype) + return _DTYPE_DOWNCAST[dtype] if unsupported and device.type in unsupported else dtype + + +def randn_tensor( + shape: tuple | list, + generator: list["torch.Generator"] | "torch.Generator" | None = None, + device: str | "torch.device" | None = None, + dtype: "torch.dtype" | None = None, + layout: "torch.layout" | None = None, +): + """A helper function to create random tensors on the desired `device` with the desired `dtype`. When + passing a list of generators, you can seed each batch size individually. If CPU generators are passed, the tensor + is always created on the CPU. + """ + # device on which tensor is created defaults to device + if isinstance(device, str): + device = torch.device(device) + rand_device = device + batch_size = shape[0] + + layout = layout or torch.strided + device = device or torch.device("cpu") + + # Neuron (XLA) does not support creating random tensors directly on device; always use CPU + if device.type == "neuron": + rand_device = torch.device("cpu") + + if generator is not None: + gen_device_type = generator.device.type if not isinstance(generator, list) else generator[0].device.type + if gen_device_type != device.type and gen_device_type == "cpu": + rand_device = "cpu" + if device.type not in ("mps", "neuron"): + logger.info( + f"The passed generator was created on 'cpu' even though a tensor on {device} was expected." + f" Tensors will be created on 'cpu' and then moved to {device}. Note that one can probably" + f" slightly speed up this function by passing a generator that was created on the {device} device." + ) + elif gen_device_type != device.type and gen_device_type == "cuda": + raise ValueError(f"Cannot generate a {device} tensor from a generator of type {gen_device_type}.") + + # make sure generator list of length 1 is treated like a non-list + if isinstance(generator, list) and len(generator) == 1: + generator = generator[0] + + if isinstance(generator, list): + shape = (1,) + shape[1:] + latents = [ + torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype, layout=layout) + for i in range(batch_size) + ] + latents = torch.cat(latents, dim=0).to(device) + else: + latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype, layout=layout).to(device) + + return latents + + +def is_compiled_module(module) -> bool: + """Check whether the module was compiled with torch.compile()""" + if is_torch_version("<", "2.0.0") or not hasattr(torch, "_dynamo"): + return False + return isinstance(module, torch._dynamo.eval_frame.OptimizedModule) + + +def unwrap_module(module): + """Unwraps a module if it was compiled with torch.compile()""" + return module._orig_mod if is_compiled_module(module) else module + + +def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor": + """Fourier filter as introduced in FreeU (https://huggingface.co/papers/2309.11497). + + This version of the method comes from here: + https://github.com/huggingface/diffusers/pull/5164#issuecomment-1732638706 + """ + x = x_in + B, C, H, W = x.shape + + # Non-power of 2 images must be float32 + if (W & (W - 1)) != 0 or (H & (H - 1)) != 0: + x = x.to(dtype=torch.float32) + # fftn does not support bfloat16, and produces the experimental ComplexHalf + # dtype (torch.complex32) when given float16, which is numerically unstable + # and triggers a UserWarning. Upcast any non-float32 dtype to float32. + elif x.dtype != torch.float32: + x = x.to(dtype=torch.float32) + + # FFT + x_freq = fftn(x, dim=(-2, -1)) + x_freq = fftshift(x_freq, dim=(-2, -1)) + + B, C, H, W = x_freq.shape + mask = torch.ones((B, C, H, W), device=x.device) + + crow, ccol = H // 2, W // 2 + mask[..., crow - threshold : crow + threshold, ccol - threshold : ccol + threshold] = scale + x_freq = x_freq * mask + + # IFFT + x_freq = ifftshift(x_freq, dim=(-2, -1)) + x_filtered = ifftn(x_freq, dim=(-2, -1)).real + + return x_filtered.to(dtype=x_in.dtype) + + +def apply_freeu( + resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs +) -> tuple["torch.Tensor", "torch.Tensor"]: + """Applies the FreeU mechanism as introduced in https: + //arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU. + + Args: + resolution_idx (`int`): Integer denoting the UNet block where FreeU is being applied. + hidden_states (`torch.Tensor`): Inputs to the underlying block. + res_hidden_states (`torch.Tensor`): Features from the skip block corresponding to the underlying block. + s1 (`float`): Scaling factor for stage 1 to attenuate the contributions of the skip features. + s2 (`float`): Scaling factor for stage 2 to attenuate the contributions of the skip features. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if resolution_idx == 0: + num_half_channels = hidden_states.shape[1] // 2 + hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b1"] + res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s1"]) + if resolution_idx == 1: + num_half_channels = hidden_states.shape[1] // 2 + hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b2"] + res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s2"]) + + return hidden_states, res_hidden_states + + +def get_torch_cuda_device_capability(): + if torch.cuda.is_available(): + device = torch.device("cuda") + compute_capability = torch.cuda.get_device_capability(device) + compute_capability = f"{compute_capability[0]}.{compute_capability[1]}" + return float(compute_capability) + else: + return None + + +@functools.lru_cache +def get_device(): + if torch.cuda.is_available(): + return "cuda" + elif is_torch_npu_available(): + return "npu" + elif hasattr(torch, "xpu") and torch.xpu.is_available(): + return "xpu" + elif torch.backends.mps.is_available(): + return "mps" + elif is_torch_mlu_available(): + return "mlu" + elif is_torch_neuronx_available() and hasattr(torch, "neuron") and torch.neuron.is_available(): + return "neuron" + else: + return "cpu" + + +def empty_device_cache(device_type: str | None = None): + if device_type is None: + device_type = get_device() + if device_type in ["cpu"]: + return + device_mod = getattr(torch, device_type, torch.cuda) + device_mod.empty_cache() + + +def device_synchronize(device_type: str | None = None): + if device_type is None: + device_type = get_device() + device_mod = getattr(torch, device_type, torch.cuda) + device_mod.synchronize() + + +def enable_full_determinism(): + """ + Helper function for reproducible behavior during distributed training. See + - https://pytorch.org/docs/stable/notes/randomness.html for pytorch + """ + # Enable PyTorch deterministic mode. This potentially requires either the environment + # variable 'CUDA_LAUNCH_BLOCKING' or 'CUBLAS_WORKSPACE_CONFIG' to be set, + # depending on the CUDA version, so we set them both here + os.environ["CUDA_LAUNCH_BLOCKING"] = "1" + os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8" + torch.use_deterministic_algorithms(True) + + # Enable CUDNN deterministic mode + torch.backends.cudnn.deterministic = True + torch.backends.cudnn.benchmark = False + torch.backends.cuda.matmul.allow_tf32 = False + + +def disable_full_determinism(): + os.environ["CUDA_LAUNCH_BLOCKING"] = "0" + os.environ["CUBLAS_WORKSPACE_CONFIG"] = "" + torch.use_deterministic_algorithms(False) + + +@functools.wraps(functools.lru_cache) +def lru_cache_unless_export(maxsize=128, typed=False): + def outer_wrapper(fn: Callable[P, T]): + cached = functools.lru_cache(maxsize=maxsize, typed=typed)(fn) + if is_torch_version("<", "2.7.0"): + return cached + + @functools.wraps(fn) + def inner_wrapper(*args: P.args, **kwargs: P.kwargs): + compiler = getattr(torch, "compiler", None) + is_exporting = bool(compiler and hasattr(compiler, "is_exporting") and compiler.is_exporting()) + is_compiling = bool(compiler and hasattr(compiler, "is_compiling") and compiler.is_compiling()) + + # Fallback for older builds where compiler.is_compiling is unavailable. + if not is_compiling: + dynamo = getattr(torch, "_dynamo", None) + if dynamo is not None and hasattr(dynamo, "is_compiling"): + is_compiling = dynamo.is_compiling() + + if is_exporting or is_compiling: + return fn(*args, **kwargs) + return cached(*args, **kwargs) + + return inner_wrapper + + return outer_wrapper + + +if is_torch_available(): + torch_device = get_device() diff --git a/diffusers/utils/typing_utils.py b/diffusers/utils/typing_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..1599e68d0a1fc802c37719c0c4d780d3ee7e5caf --- /dev/null +++ b/diffusers/utils/typing_utils.py @@ -0,0 +1,91 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Typing utilities: Utilities related to type checking and validation +""" + +from typing import Any, Dict, List, Set, Tuple, Type, Union, get_args, get_origin + + +def _is_valid_type(obj: Any, class_or_tuple: Type | tuple[Type, ...]) -> bool: + """ + Checks if an object is an instance of any of the provided types. For collections, it checks if every element is of + the correct type as well. + """ + if not isinstance(class_or_tuple, tuple): + class_or_tuple = (class_or_tuple,) + + # Unpack unions + unpacked_class_or_tuple = [] + for t in class_or_tuple: + if get_origin(t) is Union: + unpacked_class_or_tuple.extend(get_args(t)) + else: + unpacked_class_or_tuple.append(t) + class_or_tuple = tuple(unpacked_class_or_tuple) + + if Any in class_or_tuple: + return True + + obj_type = type(obj) + # Classes with obj's type + class_or_tuple = {t for t in class_or_tuple if isinstance(obj, get_origin(t) or t)} + + # Singular types (e.g. int, ControlNet, ...) + # Untyped collections (e.g. List, but not List[int]) + elem_class_or_tuple = {get_args(t) for t in class_or_tuple} + if () in elem_class_or_tuple: + return True + # Typed lists or sets + elif obj_type in (list, set): + return any(all(_is_valid_type(x, t) for x in obj) for t in elem_class_or_tuple) + # Typed tuples + elif obj_type is tuple: + return any( + # Tuples with any length and single type (e.g. Tuple[int, ...]) + (len(t) == 2 and t[-1] is Ellipsis and all(_is_valid_type(x, t[0]) for x in obj)) + or + # Tuples with fixed length and any types (e.g. Tuple[int, str]) + (len(obj) == len(t) and all(_is_valid_type(x, tt) for x, tt in zip(obj, t))) + for t in elem_class_or_tuple + ) + # Typed dicts + elif obj_type is dict: + return any( + all(_is_valid_type(k, kt) and _is_valid_type(v, vt) for k, v in obj.items()) + for kt, vt in elem_class_or_tuple + ) + + else: + return False + + +def _get_detailed_type(obj: Any) -> Type: + """ + Gets a detailed type for an object, including nested types for collections. + """ + obj_type = type(obj) + + if obj_type in (list, set): + obj_origin_type = List if obj_type is list else Set + elems_type = Union[tuple({_get_detailed_type(x) for x in obj})] + return obj_origin_type[elems_type] + elif obj_type is tuple: + return Tuple[tuple(_get_detailed_type(x) for x in obj)] + elif obj_type is dict: + keys_type = Union[tuple({_get_detailed_type(k) for k in obj.keys()})] + values_type = Union[tuple({_get_detailed_type(k) for k in obj.values()})] + return Dict[keys_type, values_type] + else: + return obj_type diff --git a/diffusers/utils/versions.py b/diffusers/utils/versions.py new file mode 100644 index 0000000000000000000000000000000000000000..ccede7b5ee5737f7e76fe9e00f9f836eb695d16a --- /dev/null +++ b/diffusers/utils/versions.py @@ -0,0 +1,116 @@ +# Copyright 2020 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Utilities for working with package versions +""" + +import importlib.metadata +import operator +import re +import sys + +from packaging import version + + +ops = { + "<": operator.lt, + "<=": operator.le, + "==": operator.eq, + "!=": operator.ne, + ">=": operator.ge, + ">": operator.gt, +} + + +def _compare_versions(op, got_ver, want_ver, requirement, pkg, hint): + if got_ver is None or want_ver is None: + raise ValueError( + f"Unable to compare versions for {requirement}: need={want_ver} found={got_ver}. This is unusual. Consider" + f" reinstalling {pkg}." + ) + if not ops[op](version.parse(got_ver), version.parse(want_ver)): + raise ImportError( + f"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}" + ) + + +def require_version(requirement: str, hint: str | None = None) -> None: + """ + Perform a runtime check of the dependency versions, using the exact same syntax used by pip. + + The installed module version comes from the *site-packages* dir via *importlib.metadata*. + + Args: + requirement (`str`): pip style definition, e.g., "tokenizers==0.9.4", "tqdm>=4.27", "numpy" + hint (`str`, *optional*): what suggestion to print in case of requirements not being met + + Example: + + ```python + require_version("pandas>1.1.2") + require_version("numpy>1.18.5", "this is important to have for whatever reason") + ```""" + + hint = f"\n{hint}" if hint is not None else "" + + # non-versioned check + if re.match(r"^[\w_\-\d]+$", requirement): + pkg, op, want_ver = requirement, None, None + else: + match = re.findall(r"^([^!=<>\s]+)([\s!=<>]{1,2}.+)", requirement) + if not match: + raise ValueError( + "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23, but" + f" got {requirement}" + ) + pkg, want_full = match[0] + want_range = want_full.split(",") # there could be multiple requirements + wanted = {} + for w in want_range: + match = re.findall(r"^([\s!=<>]{1,2})(.+)", w) + if not match: + raise ValueError( + "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23," + f" but got {requirement}" + ) + op, want_ver = match[0] + wanted[op] = want_ver + if op not in ops: + raise ValueError(f"{requirement}: need one of {list(ops.keys())}, but got {op}") + + # special case + if pkg == "python": + got_ver = ".".join([str(x) for x in sys.version_info[:3]]) + for op, want_ver in wanted.items(): + _compare_versions(op, got_ver, want_ver, requirement, pkg, hint) + return + + # check if any version is installed + try: + got_ver = importlib.metadata.version(pkg) + except importlib.metadata.PackageNotFoundError: + raise importlib.metadata.PackageNotFoundError( + f"The '{requirement}' distribution was not found and is required by this application. {hint}" + ) + + # check that the right version is installed if version number or a range was provided + if want_ver is not None: + for op, want_ver in wanted.items(): + _compare_versions(op, got_ver, want_ver, requirement, pkg, hint) + + +def require_version_core(requirement): + """require_version wrapper which emits a core-specific hint on failure""" + hint = "Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git main" + return require_version(requirement, hint) diff --git a/diffusers/video_processor.py b/diffusers/video_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..3e997d75a7e0a899e0c37ff9bde8a6470dd19fc5 --- /dev/null +++ b/diffusers/video_processor.py @@ -0,0 +1,180 @@ +# Copyright 2026 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import warnings + +import numpy as np +import PIL +import torch +import torch.nn.functional as F + +from .image_processor import VaeImageProcessor, is_valid_image, is_valid_image_imagelist + + +class VideoProcessor(VaeImageProcessor): + r"""Simple video processor.""" + + def preprocess_video(self, video, height: int | None = None, width: int | None = None, **kwargs) -> torch.Tensor: + r""" + Preprocesses input video(s). Keyword arguments will be forwarded to `VaeImageProcessor.preprocess`. + + Args: + video (`list[PIL.Image]`, `list[list[PIL.Image]]`, `torch.Tensor`, `np.array`, `list[torch.Tensor]`, `list[np.array]`): + The input video. It can be one of the following: + * list of the PIL images. + * list of list of PIL images. + * 4D Torch tensors (expected shape for each tensor `(num_frames, num_channels, height, width)`). + * 4D NumPy arrays (expected shape for each array `(num_frames, height, width, num_channels)`). + * list of 4D Torch tensors (expected shape for each tensor `(num_frames, num_channels, height, + width)`). + * list of 4D NumPy arrays (expected shape for each array `(num_frames, height, width, num_channels)`). + * 5D NumPy arrays: expected shape for each array `(batch_size, num_frames, height, width, + num_channels)`. + * 5D Torch tensors: expected shape for each array `(batch_size, num_frames, num_channels, height, + width)`. + height (`int`, *optional*, defaults to `None`): + The height in preprocessed frames of the video. If `None`, will use the `get_default_height_width()` to + get default height. + width (`int`, *optional*`, defaults to `None`): + The width in preprocessed frames of the video. If `None`, will use get_default_height_width()` to get + the default width. + + Returns: + `torch.Tensor` of shape `(batch_size, num_channels, num_frames, height, width)`: + A 5D tensor holding the batched channels-first video(s). + """ + if isinstance(video, list) and isinstance(video[0], np.ndarray) and video[0].ndim == 5: + warnings.warn( + "Passing `video` as a list of 5d np.ndarray is deprecated." + "Please concatenate the list along the batch dimension and pass it as a single 5d np.ndarray", + FutureWarning, + ) + video = np.concatenate(video, axis=0) + if isinstance(video, list) and isinstance(video[0], torch.Tensor) and video[0].ndim == 5: + warnings.warn( + "Passing `video` as a list of 5d torch.Tensor is deprecated." + "Please concatenate the list along the batch dimension and pass it as a single 5d torch.Tensor", + FutureWarning, + ) + video = torch.cat(video, axis=0) + + # ensure the input is a list of videos: + # - if it is a batch of videos (5d torch.Tensor or np.ndarray), it is converted to a list of videos (a list of 4d torch.Tensor or np.ndarray) + # - if it is a single video, it is converted to a list of one video. + if isinstance(video, (np.ndarray, torch.Tensor)) and video.ndim == 5: + video = list(video) + elif isinstance(video, list) and is_valid_image(video[0]) or is_valid_image_imagelist(video): + video = [video] + elif isinstance(video, list) and is_valid_image_imagelist(video[0]): + video = video + else: + raise ValueError( + "Input is in incorrect format. Currently, we only support numpy.ndarray, torch.Tensor, PIL.Image.Image" + ) + + video = torch.stack([self.preprocess(img, height=height, width=width, **kwargs) for img in video], dim=0) + + # move the number of channels before the number of frames. + video = video.permute(0, 2, 1, 3, 4) + + return video + + def postprocess_video( + self, video: torch.Tensor, output_type: str = "np", **kwargs + ) -> np.ndarray | torch.Tensor | list[PIL.Image.Image]: + r""" + Converts a video tensor to a list of frames for export. Keyword arguments will be forwarded to + `VaeImageProcessor.postprocess`. + + Args: + video (`torch.Tensor`): The video as a tensor. + output_type (`str`, defaults to `"np"`): Output type of the postprocessed `video` tensor. + """ + batch_size = video.shape[0] + outputs = [] + for batch_idx in range(batch_size): + batch_vid = video[batch_idx].permute(1, 0, 2, 3) + batch_output = self.postprocess(batch_vid, output_type, **kwargs) + outputs.append(batch_output) + + if output_type == "np": + outputs = np.stack(outputs) + elif output_type == "pt": + outputs = torch.stack(outputs) + elif not output_type == "pil": + raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']") + + return outputs + + @staticmethod + def classify_height_width_bin(height: int, width: int, ratios: dict) -> tuple[int, int]: + r""" + Returns the binned height and width based on the aspect ratio. + + Args: + height (`int`): The height of the image. + width (`int`): The width of the image. + ratios (`dict`): A dictionary where keys are aspect ratios and values are tuples of (height, width). + + Returns: + `tuple[int, int]`: The closest binned height and width. + """ + ar = float(height / width) + closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - ar)) + default_hw = ratios[closest_ratio] + return int(default_hw[0]), int(default_hw[1]) + + @staticmethod + def resize_and_crop_tensor(samples: torch.Tensor, new_width: int, new_height: int) -> torch.Tensor: + r""" + Resizes and crops a tensor of videos to the specified dimensions. + + Args: + samples (`torch.Tensor`): + A tensor of shape (N, C, T, H, W) where N is the batch size, C is the number of channels, T is the + number of frames, H is the height, and W is the width. + new_width (`int`): The desired width of the output videos. + new_height (`int`): The desired height of the output videos. + + Returns: + `torch.Tensor`: A tensor containing the resized and cropped videos. + """ + orig_height, orig_width = samples.shape[3], samples.shape[4] + + # Check if resizing is needed + if orig_height != new_height or orig_width != new_width: + ratio = max(new_height / orig_height, new_width / orig_width) + resized_width = int(orig_width * ratio) + resized_height = int(orig_height * ratio) + + # Reshape to (N*T, C, H, W) for interpolation + n, c, t, h, w = samples.shape + samples = samples.permute(0, 2, 1, 3, 4).reshape(n * t, c, h, w) + + # Resize + samples = F.interpolate( + samples, size=(resized_height, resized_width), mode="bilinear", align_corners=False + ) + + # Center Crop + start_x = (resized_width - new_width) // 2 + end_x = start_x + new_width + start_y = (resized_height - new_height) // 2 + end_y = start_y + new_height + samples = samples[:, :, start_y:end_y, start_x:end_x] + + # Reshape back to (N, C, T, H, W) + samples = samples.reshape(n, t, c, new_height, new_width).permute(0, 2, 1, 3, 4) + + return samples diff --git a/h3_split_blocks.py b/h3_split_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..611bd0fb5aa2076e141945c85c116a587fc077d0 --- /dev/null +++ b/h3_split_blocks.py @@ -0,0 +1,98 @@ +"""The two halves of a **split** MiniMax-H3 deployment. + +MiniMax-H3 is modular-only, and `MiniMaxH3Blocks` is a `SequentialPipelineBlocks` of eight steps: + + setup -> text_encoder -> vae_encoder -> prepare_layout -> prepare_latents -> set_timesteps -> denoise -> decode + +The conditioner (a 62.14 GiB Qwen3-VL) and the denoiser (a 61.73 GiB transformer plus ~20.5 GiB of float32 VAEs) do +not fit on one 95 GiB card unquantized, so this module cuts the sequence in two at the `text_encoder` step: + + * `MiniMaxH3ConditionerBlocks` = `[setup, text_encoder]` — loads `text_encoder` / `tokenizer` / `processor` only, + and emits `prompt_embeds` + `text_token_tags`, which is the whole wire format between the two halves. + * `MiniMaxH3GeneratorBlocks` = everything else — loads `transformer` / `vae` / `audio_vae` / the two + schedulers only, and takes `prompt_embeds` + `text_token_tags` as *inputs*. + +`setup` runs on both sides on purpose. It owns no component (it is PIL and arithmetic), it resolves the canvas, the +`17 * n + 5` frame count and the latent geometry, and it puts the keyframes onto that canvas — which the conditioner +needs to build its vision blocks and the generator needs to encode with the video VAE. Running it twice over the same +inputs is deterministic; the conditioner half returns its resolved `height` / `width` / `num_frames` anyway, so the +caller can pin them explicitly on the generator half. +""" + +from diffusers.modular_pipelines.minimax_h3.before_denoise import ( + MiniMaxH3PrepareLatentsStep, + MiniMaxH3PrepareLayoutStep, + MiniMaxH3SetTimestepsStep, +) +from diffusers.modular_pipelines.minimax_h3.before_encoder import MiniMaxH3SetupStep +from diffusers.modular_pipelines.minimax_h3.denoise import MiniMaxH3DenoiseStep +from diffusers.modular_pipelines.minimax_h3.encoders import MiniMaxH3TextEncoderStep +from diffusers.modular_pipelines.minimax_h3.modular_blocks_minimax_h3 import ( + MiniMaxH3AutoKeyframeVaeEncoderStep, + MiniMaxH3DecodeStep, + _generation_outputs, +) +from diffusers.modular_pipelines.modular_pipeline import SequentialPipelineBlocks +from diffusers.modular_pipelines.modular_pipeline_utils import OutputParam + + +class MiniMaxH3ConditionerBlocks(SequentialPipelineBlocks): + """The conditioner half of a split MiniMax-H3: the request plan plus the Qwen3-VL read at its 50th layer.""" + + model_name = "minimax-h3" + block_classes = [MiniMaxH3SetupStep, MiniMaxH3TextEncoderStep] + block_names = ["setup", "text_encoder"] + + @property + def description(self): + return ( + "The conditioner half of a split MiniMax-H3 deployment: resolves the request plan (canvas, frame count, " + "latent geometry, keyframes on the canvas) and encodes MiniMax-H3's presentation of it into the " + "`prompt_embeds` / `text_token_tags` pair the denoising half consumes." + ) + + @property + def outputs(self): + return [ + OutputParam.template("prompt_embeds"), + OutputParam("text_token_tags", description="The per-row modality tag of every row of `prompt_embeds`."), + OutputParam("height", type_hint=int, description="Resolved height of the generated video in pixels."), + OutputParam("width", type_hint=int, description="Resolved width of the generated video in pixels."), + OutputParam("num_frames", type_hint=int, description="Resolved number of frames, of the form 17 * n + 5."), + ] + + +class MiniMaxH3GeneratorBlocks(SequentialPipelineBlocks): + """The denoising half of a split MiniMax-H3: `MiniMaxH3Blocks` with its `text_encoder` step removed.""" + + model_name = "minimax-h3" + block_classes = [ + MiniMaxH3SetupStep, + MiniMaxH3AutoKeyframeVaeEncoderStep, + MiniMaxH3PrepareLayoutStep, + MiniMaxH3PrepareLatentsStep, + MiniMaxH3SetTimestepsStep, + MiniMaxH3DenoiseStep, + MiniMaxH3DecodeStep, + ] + block_names = [ + "setup", + "vae_encoder", + "prepare_layout", + "prepare_latents", + "set_timesteps", + "denoise", + "decode", + ] + + @property + def description(self): + return ( + "The denoising half of a split MiniMax-H3 deployment: `MiniMaxH3Blocks` without its text-encoder step, so " + "`prompt_embeds` and `text_token_tags` come in as inputs and the 62.14 GiB Qwen3-VL conditioner is never " + "loaded here." + ) + + @property + def outputs(self): + return _generation_outputs() diff --git a/packages.txt b/packages.txt new file mode 100644 index 0000000000000000000000000000000000000000..20645e641240cb419f5fc66c14c1447e91daf669 --- /dev/null +++ b/packages.txt @@ -0,0 +1 @@ +ffmpeg diff --git a/requirements.txt b/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..eed7fb8d805c9053a88bbb3805189d1b27daf301 --- /dev/null +++ b/requirements.txt @@ -0,0 +1,25 @@ +# `diffusers` itself is not installed: the MiniMax-H3 integration branch's `src/diffusers` tree is vendored into this +# Space as a top-level `diffusers/` package and imported from the working directory (`sync.sh` re-syncs it). What +# follows is everything that tree imports. Nothing is quantized in this deployment, so there is no `torchao`. +--extra-index-url https://download.pytorch.org/whl/cu130 +torch==2.11.0 +torchvision==0.26.0 +# Pinned to the version the MiniMax-H3 parity work was verified against: the Qwen3-VL processor decides the vision +# patch count, so a different minor changes the conditioning. +transformers==5.8.0 +accelerate==1.14.0 +# The vendored tree needs `get_cached_repo_tree`, which landed in 1.23; diffusers pins <2. +huggingface-hub==1.24.0 +gradio==6.20.0 +spaces==0.51.1 +# No `kernels` pin on purpose: the Hub attention backends want `kernels>=0.12.3`, and that version breaks +# transformers 5.8.0 at import. cuDNN attention (`_native_cudnn`) is faster than the SDPA default and needs nothing. +# PyAV: muxing the generated soundtrack onto the frames (`encode_video`). +av +pillow +numpy +requests +safetensors>=0.8.0 +filelock +regex +importlib_metadata