Spaces:
Running on Zero
Running on Zero
File size: 13,765 Bytes
3db1cbc 7f70ff0 3db1cbc | 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 | # Copyright 2026 Black Forest Labs 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 torch
from ...models import Flux3Transformer2DModel
from ...utils import logging
from ..modular_pipeline import ModularPipelineBlocks, PipelineState
from ..modular_pipeline_utils import ComponentSpec, InputParam, OutputParam
from .modular_pipeline import Flux3ModularPipeline
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
# Copied from diffusers.modular_pipelines.qwenimage.inputs.repeat_tensor_to_batch_size
def repeat_tensor_to_batch_size(
input_name: str,
input_tensor: torch.Tensor,
batch_size: int,
num_images_per_prompt: int = 1,
) -> torch.Tensor:
"""Repeat tensor elements to match the final batch size.
This function expands a tensor's batch dimension to match the final batch size (batch_size * num_images_per_prompt)
by repeating each element along dimension 0.
The input tensor must have batch size 1 or batch_size. The function will:
- If batch size is 1: repeat each element (batch_size * num_images_per_prompt) times
- If batch size equals batch_size: repeat each element num_images_per_prompt times
Args:
input_name (str): Name of the input tensor (used for error messages)
input_tensor (torch.Tensor): The tensor to repeat. Must have batch size 1 or batch_size.
batch_size (int): The base batch size (number of prompts)
num_images_per_prompt (int, optional): Number of images to generate per prompt. Defaults to 1.
Returns:
torch.Tensor: The repeated tensor with final batch size (batch_size * num_images_per_prompt)
Raises:
ValueError: If input_tensor is not a torch.Tensor or has invalid batch size
Examples:
tensor = torch.tensor([[1, 2, 3]]) # shape: [1, 3] repeated = repeat_tensor_to_batch_size("image", tensor,
batch_size=2, num_images_per_prompt=2) repeated # tensor([[1, 2, 3], [1, 2, 3], [1, 2, 3], [1, 2, 3]]) - shape:
[4, 3]
tensor = torch.tensor([[1, 2, 3], [4, 5, 6]]) # shape: [2, 3] repeated = repeat_tensor_to_batch_size("image",
tensor, batch_size=2, num_images_per_prompt=2) repeated # tensor([[1, 2, 3], [1, 2, 3], [4, 5, 6], [4, 5, 6]])
- shape: [4, 3]
"""
# make sure input is a tensor
if not isinstance(input_tensor, torch.Tensor):
raise ValueError(f"`{input_name}` must be a tensor")
# make sure input tensor e.g. image_latents has batch size 1 or batch_size same as prompts
if input_tensor.shape[0] == 1:
repeat_by = batch_size * num_images_per_prompt
elif input_tensor.shape[0] == batch_size:
repeat_by = num_images_per_prompt
else:
raise ValueError(f"`{input_name}` must have batch size 1 or {batch_size}, but got {input_tensor.shape[0]}")
# expand the tensor to match the batch_size * num_images_per_prompt
input_tensor = input_tensor.repeat_interleave(repeat_by, dim=0)
return input_tensor
class Flux3TextInputStep(ModularPipelineBlocks):
model_name = "flux3"
@property
def description(self) -> str:
return (
"This step:\n"
" 1. Determines `batch_size` from `prompt_embeds` and the compute `dtype` from the transformer\n"
" 2. Ensures all text embeddings have consistent batch sizes (batch_size * num_images_per_prompt)"
)
@property
def expected_components(self) -> list[ComponentSpec]:
return [ComponentSpec("transformer", Flux3Transformer2DModel)]
@property
def inputs(self) -> list[InputParam]:
return [
InputParam.template("num_images_per_prompt"),
InputParam(
"prompt_embeds",
required=True,
kwargs_type="denoiser_input_fields",
type_hint=torch.Tensor,
description="Pre-generated text embeddings. Can be generated from text_encoder step.",
),
InputParam(
"negative_prompt_embeds",
required=False,
kwargs_type="denoiser_input_fields",
type_hint=torch.Tensor,
description="Pre-generated negative text embeddings. Can be generated from text_encoder step.",
),
]
@property
def intermediate_outputs(self) -> list[OutputParam]:
return [
OutputParam(
"batch_size",
type_hint=int,
description="Number of prompts, the final batch size of model inputs should be batch_size * num_images_per_prompt",
),
OutputParam(
"dtype",
type_hint=torch.dtype,
description="Data type of model tensor inputs (the transformer's dtype)",
),
OutputParam(
"prompt_embeds",
type_hint=torch.Tensor,
kwargs_type="denoiser_input_fields",
description="Text embeddings used to guide the image generation",
),
OutputParam(
"negative_prompt_embeds",
type_hint=torch.Tensor,
kwargs_type="denoiser_input_fields",
description="Negative text embeddings used to guide the image generation",
),
]
@torch.no_grad()
def __call__(self, components: Flux3ModularPipeline, state: PipelineState) -> PipelineState:
block_state = self.get_block_state(state)
block_state.batch_size = block_state.prompt_embeds.shape[0]
# the denoiser consumes every tensor this chain prepares, so its dtype is the compute dtype:
# the text encoder may run in a different precision (e.g. a bf16 encoder with an fp32 transformer)
block_state.dtype = components.transformer.dtype
block_state.prompt_embeds = repeat_tensor_to_batch_size(
"prompt_embeds", block_state.prompt_embeds, block_state.batch_size, block_state.num_images_per_prompt
).to(block_state.dtype)
if block_state.negative_prompt_embeds is not None:
block_state.negative_prompt_embeds = repeat_tensor_to_batch_size(
"negative_prompt_embeds",
block_state.negative_prompt_embeds,
block_state.batch_size,
block_state.num_images_per_prompt,
).to(block_state.dtype)
elif components.requires_unconditional_embeds:
raise ValueError(
"The guider needs unconditional embeddings but `negative_prompt_embeds` is None. Pass `negative_prompt` "
"(an empty string for plain classifier-free guidance) to the text encoder step, or disable the guider."
)
self.set_block_state(state, block_state)
return components, state
class Flux3DistilledTextInputStep(ModularPipelineBlocks):
model_name = "flux3-distilled"
@property
def description(self) -> str:
return (
"This step:\n"
" 1. Determines `batch_size` from `prompt_embeds` and the compute `dtype` from the transformer\n"
" 2. Repeats the text embeddings to the batch size (batch_size * num_images_per_prompt)\n"
"For the step-distilled FLUX.3 model, which samples without classifier-free guidance (no negative "
"embeddings)."
)
@property
def expected_components(self) -> list[ComponentSpec]:
return [ComponentSpec("transformer", Flux3Transformer2DModel)]
@property
def inputs(self) -> list[InputParam]:
return [
InputParam.template("num_images_per_prompt"),
InputParam(
"prompt_embeds",
required=True,
kwargs_type="denoiser_input_fields",
type_hint=torch.Tensor,
description="Pre-generated text embeddings. Can be generated from text_encoder step.",
),
]
@property
def intermediate_outputs(self) -> list[OutputParam]:
return [
OutputParam(
"batch_size",
type_hint=int,
description="Number of prompts, the final batch size of model inputs should be batch_size * num_images_per_prompt",
),
OutputParam(
"dtype",
type_hint=torch.dtype,
description="Data type of model tensor inputs (the transformer's dtype)",
),
OutputParam(
"prompt_embeds",
type_hint=torch.Tensor,
kwargs_type="denoiser_input_fields",
description="Text embeddings used to guide the image generation",
),
]
@torch.no_grad()
def __call__(self, components: Flux3ModularPipeline, state: PipelineState) -> PipelineState:
block_state = self.get_block_state(state)
block_state.batch_size = block_state.prompt_embeds.shape[0]
# the denoiser consumes every tensor this chain prepares, so its dtype is the compute dtype
block_state.dtype = components.transformer.dtype
block_state.prompt_embeds = repeat_tensor_to_batch_size(
"prompt_embeds", block_state.prompt_embeds, block_state.batch_size, block_state.num_images_per_prompt
).to(block_state.dtype)
self.set_block_state(state, block_state)
return components, state
class Flux3ImageCondInputStep(ModularPipelineBlocks):
model_name = "flux3"
def __init__(self, default_height: int = 1024, default_width: int = 1024):
# the generation size defaults to this area at the first reference image's aspect ratio
self.default_height = default_height
self.default_width = default_width
super().__init__()
@property
def description(self) -> str:
return (
"Step that expands the reference image latents to the final batch size and sets the generation "
f"`height`/`width` when not provided: {self.default_height} x {self.default_width} pixels of area at the "
"first reference image's aspect ratio, on the 16-pixel grid (a given side overrides its default). The "
"reference images keep their own resolution, the generation size is independent of them."
)
@property
def inputs(self) -> list[InputParam]:
return [
InputParam(
"image_cond_latents",
required=True,
type_hint=list[torch.Tensor],
description="Normalized packed latents of the reference images, one tensor of shape (B, C, height_i//16, width_i//16) per image",
),
InputParam(
"height",
type_hint=int,
description="The height in pixels of the generated image; defaults to the first reference image's "
"aspect ratio at the default area",
),
InputParam(
"width",
type_hint=int,
description="The width in pixels of the generated image; defaults to the first reference image's "
"aspect ratio at the default area",
),
InputParam.template("num_images_per_prompt"),
InputParam("batch_size", required=True, type_hint=int),
]
@property
def intermediate_outputs(self) -> list[OutputParam]:
return [
OutputParam(
"image_cond_latents",
type_hint=list[torch.Tensor],
description="Normalized packed latents of the reference images, expanded to the final batch size",
),
OutputParam("height", type_hint=int, description="The height in pixels of the generated image"),
OutputParam("width", type_hint=int, description="The width in pixels of the generated image"),
]
@torch.no_grad()
def __call__(self, components: Flux3ModularPipeline, state: PipelineState) -> PipelineState:
block_state = self.get_block_state(state)
# the generation size defaults to the default area at the first reference's aspect ratio, rounded to a
# multiple of 16 pixels (one token); a given `height` or `width` overrides its side
if block_state.height is None or block_state.width is None:
multiple_of = components.vae_scale_factor * 2
latent_height, latent_width = block_state.image_cond_latents[0].shape[-2:]
scale = math.sqrt(self.default_height * self.default_width / (latent_height * latent_width)) / multiple_of
block_state.height = block_state.height or max(multiple_of, round(latent_height * scale) * multiple_of)
block_state.width = block_state.width or max(multiple_of, round(latent_width * scale) * multiple_of)
block_state.image_cond_latents = [
repeat_tensor_to_batch_size(
"image_cond_latents", latents, block_state.batch_size, block_state.num_images_per_prompt
)
for latents in block_state.image_cond_latents
]
self.set_block_state(state, block_state)
return components, state
|