File size: 18,000 Bytes
b67e8f3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 | """Hub custom pipeline: MiniT2ITextToImagePipeline.
Load with native Hugging Face diffusers and trust_remote_code=True.
"""
from __future__ import annotations
from diffusers.image_processor import VaeImageProcessor
from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers
from diffusers.utils import BaseOutput
from diffusers.utils.torch_utils import randn_tensor
# Copyright 2025 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 json
import os
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple, Union
os.environ.setdefault("USE_FLAX", "0")
os.environ.setdefault("TRANSFORMERS_NO_FLAX", "1")
import torch
from huggingface_hub import snapshot_download
from PIL import Image
from transformers import AutoTokenizer, T5EncoderModel
from transformers import logging as transformers_logging
transformers_logging.set_verbosity_error()
DEFAULT_NUM_INFERENCE_STEPS = 100
NOISE_INIT_SCALE = 2.0
EXAMPLE_DOC_STRING = """
Examples:
```py
>>> from pathlib import Path
>>> import torch
>>> from diffusers import DiffusionPipeline, FlowMatchEulerDiscreteScheduler
>>> model_dir = Path("./minit2i-diffusers").resolve()
>>> pipe = DiffusionPipeline.from_pretrained(
... str(model_dir),
... local_files_only=True,
... custom_pipeline=str(model_dir / "pipeline.py"),
... trust_remote_code=True,
... torch_dtype=torch.bfloat16,
... )
>>> pipe.to("cuda")
>>> pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe.scheduler.config)
>>> generator = torch.Generator(device="cuda").manual_seed(42)
>>> image = pipe(
... "a cinematic portrait of a robot musician",
... num_inference_steps=100,
... guidance_scale=6.0,
... generator=generator,
... ).images[0]
>>> image.save("demo.png")
```
"""
MODEL_ALIASES: Dict[str, str] = {
"b": "minit2i-b-16",
"b16": "minit2i-b-16",
"b-16": "minit2i-b-16",
"base": "minit2i-b-16",
"minit2i-b16": "minit2i-b-16",
"minit2i-b-16": "minit2i-b-16",
"minit2i-b/16": "minit2i-b-16",
"l": "minit2i-l-16",
"l16": "minit2i-l-16",
"l-16": "minit2i-l-16",
"large": "minit2i-l-16",
"minit2i-l16": "minit2i-l-16",
"minit2i-l-16": "minit2i-l-16",
"minit2i-l/16": "minit2i-l-16",
}
def resolve_model_type(model_type: str) -> str:
key = model_type.lower().replace("_", "-")
if key not in MODEL_ALIASES:
choices = ", ".join(sorted(set(MODEL_ALIASES)))
raise ValueError(f"Unknown model_type={model_type!r}. Expected one of: {choices}")
return MODEL_ALIASES[key]
class MiniT2ITextToImagePipeline(DiffusionPipeline):
r"""
Text-to-image pipeline for MiniT2I pixel-space flow matching.
Parameters:
transformer ([`MiniT2IMMJiTModel`]):
MiniT2I MM-JiT transformer that predicts flow-matching velocity in pixel space.
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
Flow-matching Euler scheduler. Other [`KarrasDiffusionSchedulers`] can be swapped at inference time.
tokenizer ([`AutoTokenizer`], *optional*):
Tokenizer for the text encoder.
text_encoder ([`T5EncoderModel`], *optional*):
Text encoder used to embed prompts.
"""
model_cpu_offload_seq = "text_encoder->transformer"
_optional_components = ["tokenizer", "text_encoder"]
def __init__(
self,
transformer,
scheduler,
tokenizer=None,
text_encoder=None,
text_encoder_name: str = "google/flan-t5-large",
model_type: str = "b16",
repo_id_or_path: Optional[str] = None,
default_num_inference_steps: int = DEFAULT_NUM_INFERENCE_STEPS,
):
super().__init__()
if scheduler is None:
scheduler = self._default_inference_scheduler()
self.register_modules(
transformer=transformer,
scheduler=scheduler,
tokenizer=tokenizer,
text_encoder=text_encoder,
)
self.register_to_config(
text_encoder_name=text_encoder_name,
model_type=model_type,
repo_id_or_path=repo_id_or_path,
default_num_inference_steps=int(default_num_inference_steps),
)
self._variant_transformers: Dict[str, MiniT2IMMJiTModel] = {}
self._active_model_type = resolve_model_type(model_type)
@staticmethod
def _default_inference_scheduler() -> FlowMatchEulerDiscreteScheduler:
return FlowMatchEulerDiscreteScheduler(
num_train_timesteps=1000,
shift=1.0,
stochastic_sampling=False,
)
@classmethod
def _load_scheduler_from_dir(
cls,
scheduler_dir: Path,
model_kwargs: Dict[str, Any],
) -> Tuple[KarrasDiffusionSchedulers, int]:
config_path = scheduler_dir / "scheduler_config.json"
if not config_path.exists():
return cls._default_inference_scheduler(), DEFAULT_NUM_INFERENCE_STEPS
config = json.loads(config_path.read_text(encoding="utf-8"))
class_name = config.get("_class_name", "")
default_steps = int(config.get("num_inference_steps", DEFAULT_NUM_INFERENCE_STEPS))
if class_name == "MiniT2IFlowMatchScheduler":
return cls._default_inference_scheduler(), default_steps
schedulers_pkg = _hf["schedulers"]
if hasattr(schedulers_pkg, class_name):
scheduler_cls = getattr(schedulers_pkg, class_name)
return scheduler_cls.from_pretrained(str(scheduler_dir), **model_kwargs), default_steps
return cls._default_inference_scheduler(), default_steps
@staticmethod
def _resolve_transformer_path(root: Path, variant_dir: str) -> Path:
variant_transformer = root / variant_dir / "transformer"
if variant_transformer.exists():
return variant_transformer
root_transformer = root / "transformer"
if root_transformer.exists():
return root_transformer
raise FileNotFoundError(
f"Could not find transformer weights under {root}. "
f"Tried {variant_transformer} and {root_transformer}."
)
def _get_transformer(
self,
model_type: Optional[str],
repo_id_or_path: Optional[str],
torch_dtype: Optional[torch.dtype] = None,
variant: Optional[str] = None,
) -> MiniT2IMMJiTModel:
active_type = resolve_model_type(model_type or self.config.model_type)
if active_type == self._active_model_type and self.transformer is not None:
return self.transformer
if active_type in self._variant_transformers:
return self._variant_transformers[active_type]
repo = repo_id_or_path or self.config.repo_id_or_path
if repo is None:
raise ValueError("model_type switching requires repo_id_or_path to be set on the pipeline.")
root = Path(repo)
if not root.exists():
root = Path(snapshot_download(repo_id=str(repo)))
transformer = MiniT2IMMJiTModel.from_pretrained(
self._resolve_transformer_path(root, active_type),
torch_dtype=torch_dtype,
variant=variant,
)
self._variant_transformers[active_type] = transformer
if active_type == resolve_model_type(self.config.model_type):
self.transformer = transformer
self._active_model_type = active_type
return transformer
@staticmethod
def prepare_extra_step_kwargs(
scheduler,
generator: Optional[Union[torch.Generator, List[torch.Generator]]],
) -> Dict[str, Any]:
kwargs: Dict[str, Any] = {}
step_params = set(inspect.signature(scheduler.step).parameters.keys())
if "generator" in step_params:
kwargs["generator"] = generator
return kwargs
def check_inputs(
self,
prompt: Union[str, List[str]],
guidance_scale: float,
num_inference_steps: int,
output_type: str,
) -> None:
if not isinstance(prompt, str) and not (isinstance(prompt, list) and all(isinstance(p, str) for p in prompt)):
raise TypeError(f"`prompt` must be a string or list of strings, got {type(prompt)}.")
if guidance_scale < 0:
raise ValueError(f"`guidance_scale` must be non-negative, got {guidance_scale}.")
if num_inference_steps <= 0:
raise ValueError(f"`num_inference_steps` must be positive, got {num_inference_steps}.")
if output_type not in {"pil", "np", "pt", "latent"}:
raise ValueError(f"Unsupported `output_type`: {output_type}")
def prepare_latents(
self,
batch_size: int,
image_size: int,
in_channels: int,
device: torch.device,
dtype: torch.dtype,
generator: Optional[torch.Generator] = None,
latents: Optional[torch.Tensor] = None,
) -> torch.Tensor:
shape = (batch_size, in_channels, image_size, image_size)
if latents is None:
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
latents = latents * NOISE_INIT_SCALE
else:
latents = latents.to(device=device, dtype=dtype)
if tuple(latents.shape) != shape:
raise ValueError(f"Invalid `latents` shape: {tuple(latents.shape)}. Expected {shape}.")
return latents
def _encode_prompt(
self,
prompt: Union[str, List[str]],
device: torch.device,
transformer = None,
) -> Tuple[torch.Tensor, torch.Tensor]:
if isinstance(prompt, str):
prompt = [prompt]
transformer = transformer or self.transformer
if self.tokenizer is None:
self.tokenizer = AutoTokenizer.from_pretrained(self.config.text_encoder_name)
if self.text_encoder is None:
self.text_encoder = T5EncoderModel.from_pretrained(self.config.text_encoder_name)
if next(self.text_encoder.parameters()).device != device:
self.text_encoder.to(device)
cfg = transformer.mmjit_config
tokens = self.tokenizer(
prompt,
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=cfg.prompt_length,
)
input_ids = tokens.input_ids.to(device)
attn = tokens.attention_mask.to(device)
text = self.text_encoder(input_ids=input_ids, attention_mask=attn).last_hidden_state
return text, attn
@staticmethod
def _cfg_velocity(
transformer,
x: torch.Tensor,
t: torch.Tensor,
text: torch.Tensor,
mask: torch.Tensor,
cfg_scale: float,
) -> torch.Tensor:
batch_size = x.shape[0]
doubled_x = torch.cat([x, x], dim=0)
doubled_t = torch.cat([t, t], dim=0)
doubled_text = torch.cat([text, text], dim=0)
null_mask = torch.zeros_like(mask)
doubled_mask = torch.cat([mask, null_mask], dim=0)
velocity = transformer.pred_velocity(doubled_x, doubled_t, doubled_text, doubled_mask)
cond, uncond = velocity[:batch_size], velocity[batch_size:]
cfg_interval = transformer.mmjit_config.cfg_interval
use_cfg = ((t >= cfg_interval[0]) & (t <= cfg_interval[1])).to(velocity.dtype)
scale = torch.where(
use_cfg[:, None, None, None] > 0,
torch.tensor(cfg_scale, device=x.device, dtype=velocity.dtype),
torch.tensor(1.0, device=x.device, dtype=velocity.dtype),
)
return uncond + (cond - uncond) * scale
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]],
num_images_per_prompt: int = 1,
guidance_scale: float = 6.0,
num_inference_steps: Optional[int] = None,
generator: Optional[torch.Generator] = None,
latents: Optional[torch.Tensor] = None,
output_type: str = "pil",
return_dict: bool = True,
progress: bool = True,
model_type: Optional[str] = None,
repo_id_or_path: Optional[str] = None,
variant: Optional[str] = None,
torch_dtype: Optional[torch.dtype] = None,
) -> Union[ImagePipelineOutput, Tuple]:
r"""
Generate images from text prompts with MiniT2I.
Args:
prompt (`str` or `list[str]`):
Text prompt or batch of prompts.
num_images_per_prompt (`int`, defaults to `1`):
Number of images to generate per prompt.
guidance_scale (`float`, defaults to `6.0`):
Classifier-free guidance scale. CFG is active when `guidance_scale != 1.0`.
num_inference_steps (`int`, *optional*):
Number of denoising steps. Defaults to the pipeline config value.
generator (`torch.Generator`, *optional*):
RNG for reproducibility.
latents (`torch.Tensor`, *optional*):
Pre-generated pixel latents with shape `(batch, channels, height, width)`.
output_type (`str`, defaults to `"pil"`):
`"pil"`, `"np"`, `"pt"`, or `"latent"`.
return_dict (`bool`, defaults to `True`):
Return [`ImagePipelineOutput`] if True.
progress (`bool`, defaults to `True`):
Whether to show a progress bar during denoising.
model_type (`str`, *optional*):
MiniT2I variant alias such as `"b16"` or `"l16"`.
repo_id_or_path (`str`, *optional*):
Hub id or local path used when switching `model_type`.
variant (`str`, *optional*):
Weight variant passed to `from_pretrained`.
torch_dtype (`torch.dtype`, *optional*):
Optional dtype override when loading a different transformer variant.
"""
num_inference_steps = int(num_inference_steps or self.config.default_num_inference_steps)
self.check_inputs(prompt, guidance_scale, num_inference_steps, output_type)
transformer = self._get_transformer(model_type, repo_id_or_path, torch_dtype=torch_dtype, variant=variant)
device = self._execution_device
transformer = transformer.to(device)
if isinstance(prompt, str):
prompt_batch = [prompt] * num_images_per_prompt
else:
prompt_batch = []
for entry in prompt:
prompt_batch.extend([entry] * num_images_per_prompt)
batch_size = len(prompt_batch)
mmjit_cfg = transformer.mmjit_config
model_dtype = next(transformer.parameters()).dtype
text, attn = self._encode_prompt(prompt_batch, device, transformer=transformer)
text = text.to(dtype=model_dtype)
attn = attn.to(dtype=model_dtype)
if getattr(self.scheduler.config, "stochastic_sampling", False):
raise ValueError(
"MiniT2I expects deterministic FlowMatchEulerDiscreteScheduler stepping "
"(scheduler.config.stochastic_sampling=False)."
)
extra_step_kwargs = self.prepare_extra_step_kwargs(self.scheduler, generator=generator)
self.scheduler.set_timesteps(num_inference_steps, device=device)
num_train_timesteps = self.scheduler.config.num_train_timesteps
latents = self.prepare_latents(
batch_size=batch_size,
image_size=mmjit_cfg.image_size,
in_channels=mmjit_cfg.in_channels,
device=device,
dtype=model_dtype,
generator=generator,
latents=latents,
)
timesteps = self.scheduler.timesteps
if progress:
timesteps = self.progress_bar(timesteps)
using_cfg = guidance_scale != 1.0
for timestep in timesteps:
flow_time = 1.0 - float(timestep) / num_train_timesteps
t = torch.full((batch_size,), flow_time, device=device, dtype=model_dtype)
if using_cfg:
velocity = self._cfg_velocity(transformer, latents, t, text, attn, guidance_scale)
else:
velocity = transformer.pred_velocity(latents, t, text, attn)
# MiniT2I integrates velocity from noise (t=0) to data (t=1); flip sign for
# FlowMatchEulerDiscreteScheduler sigma decreasing from 1 to 0.
latents = self.scheduler.step(-velocity, timestep, latents, **extra_step_kwargs).prev_sample
if output_type == "latent":
images = latents
else:
images = (latents.clamp(-1, 1) * 127.5 + 128.0).clamp(0, 255).to(torch.uint8)
if output_type == "pt":
images = images.float() / 255.0
else:
images = images.permute(0, 2, 3, 1).cpu().numpy()
if output_type == "pil":
images = [Image.fromarray(image) for image in images]
self.maybe_free_model_hooks()
if not return_dict:
return (images,)
return ImagePipelineOutput(images=images) |