File size: 33,216 Bytes
1ad8ca8 693939a 1ad8ca8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 |
import inspect
from typing import Any, Dict, List, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms.functional as TF
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import FromSingleFileMixin
from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import (
EXAMPLE_DOC_STRING,
rescale_noise_cfg,
retrieve_timesteps,
)
from diffusers.schedulers import KarrasDiffusionSchedulers
from diffusers.utils import (
USE_PEFT_BACKEND,
BaseOutput,
deprecate,
logging,
replace_example_docstring,
)
from diffusers.utils.torch_utils import randn_tensor
from PIL import Image
from diffusers import AutoencoderKL, DiffusionPipeline, UNet2DConditionModel
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
from dataclasses import dataclass
def postprocess(
image: torch.FloatTensor,
output_type: str = "pil",
):
"""
Postprocess the image output from tensor to `output_type`.
Args:
image (`torch.FloatTensor`):
The image input, should be a pytorch tensor with shape `B x C x H x W`.
output_type (`str`, *optional*, defaults to `pil`):
The output type of the image, can be one of `pil`, `np`, `pt`, `latent`.
Returns:
`PIL.Image.Image`, `np.ndarray` or `torch.FloatTensor`:
The postprocessed image.
"""
if not isinstance(image, torch.Tensor):
raise ValueError(
f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
)
if output_type not in ["latent", "pt", "np", "pil"]:
deprecation_message = (
f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
"`pil`, `np`, `pt`, `latent`"
)
deprecate(
"Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False
)
output_type = "np"
image = image.detach().cpu()
if output_type == "latent":
return image
# denormalize the image
image = image.clamp(-1, 1) * 0.5 + 0.5
materials = []
for i in range(image.shape[0]):
material = MatForgerMaterial()
material.init_from_tensor(image[i])
if output_type == "pt":
material.to_pt()
if output_type == "np":
material.to_np()
if output_type == "pil":
material.to_pil()
materials.append(material)
return materials
@dataclass
class MatForgerMaterial:
def __init__(
self,
basecolor: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
normal: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
height: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
roughness: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
metallic: Optional[Union[Image.Image, np.ndarray, torch.FloatTensor]] = None,
):
self.basecolor = basecolor
self.normal = normal
self.height = height
self.roughness = roughness
self.metallic = metallic
def _to_numpy(self, image):
if image is None:
return None
if isinstance(image, Image.Image):
image = np.array(image)
elif isinstance(image, torch.FloatTensor):
image = image.cpu().numpy()
return image
def _to_pil(self, image):
if image is None:
return None
if isinstance(image, np.ndarray):
image = Image.fromarray(image)
elif isinstance(image, torch.FloatTensor):
image = TF.to_pil_image(image)
return image
def _to_pt(self, image):
if image is None:
return None
if isinstance(image, np.ndarray):
image = torch.from_numpy(image)
elif isinstance(image, Image.Image):
image = TF.to_tensor(image)
return image
def compute_normal_map_z_component(self, normal: torch.FloatTensor):
"""
Compute the z-component of the normal map for a tensor of shape (2, H, W).
Parameters:
- normal_map (torch.Tensor): A tensor of shape (2, H, W) containing the x and y components of the normal map.
Returns:
- A tensor of shape (1, H, W) containing the z-component of the normal map.
"""
# Normalize the normal map to the range [-1, 1]
normal = normal * 2 - 1
# Square the x and y components
squared = normal**2
# Sum along the first dimension (x^2 + y^2)
sum_squared = squared.sum(dim=0, keepdim=True)
# Compute z-component: sqrt(1 - (x^2 + y^2))
z_component = torch.sqrt(1 - sum_squared).clamp(
min=0
) # Clamp to avoid negative values under sqrt
normal = torch.cat([normal, z_component], dim=0)
normal = normal * 0.5 + 0.5 # Denormalize to [0, 1]
return normal
def init_from_tensor(self, image: torch.FloatTensor):
assert image.shape[0] >= 8, "Input tensor should have at least 8 channels"
self.basecolor = image[:3]
self.normal = self.compute_normal_map_z_component(image[3:5])
self.height = image[5:6]
self.roughness = image[6:7]
self.metallic = image[7:8]
def to_pt(self):
# convert to pytorch tensor
self.basecolor = self._to_pt(self.basecolor)
self.normal = self._to_pt(self.normal)
self.height = self._to_pt(self.height)
self.roughness = self._to_pt(self.roughness)
self.metallic = self._to_pt(self.metallic)
def to_np(self):
# convert to numpy
self.basecolor = self._to_numpy(self.basecolor)
self.normal = self._to_numpy(self.normal)
self.height = self._to_numpy(self.height)
self.roughness = self._to_numpy(self.roughness)
self.metallic = self._to_numpy(self.metallic)
def to_pil(self):
# convert to PIL image
self.basecolor = self._to_pil(self.basecolor)
self.normal = self._to_pil(self.normal)
self.height = self._to_pil(self.height)
self.roughness = self._to_pil(self.roughness)
self.metallic = self._to_pil(self.metallic)
def as_dict(self):
return {
"basecolor": self.basecolor,
"normal": self.normal,
"height": self.height,
"roughness": self.roughness,
"metallic": self.metallic,
}
@dataclass
class MatForgerPipelineOutput(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)`.
"""
images: List[MatForgerMaterial]
class MatForgerPipeline(DiffusionPipeline, FromSingleFileMixin):
model_cpu_offload_seq = "prompt_encoder->unet->vae"
def __init__(
self,
vae: AutoencoderKL,
unet: UNet2DConditionModel,
prompt_encoder: nn.Module,
scheduler: KarrasDiffusionSchedulers,
):
super().__init__()
self.register_modules(
vae=vae,
unet=unet,
prompt_encoder=prompt_encoder,
scheduler=scheduler,
)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
def enable_vae_slicing(self):
r"""
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
"""
self.vae.enable_slicing()
def disable_vae_slicing(self):
r"""
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_slicing()
def enable_vae_tiling(self):
r"""
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
processing larger images.
"""
self.vae.enable_tiling()
def disable_vae_tiling(self):
r"""
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
computing decoding in one step.
"""
self.vae.disable_tiling()
def encode_prompt(
self,
prompt,
device,
num_images_per_prompt,
do_classifier_free_guidance,
negative_prompt=None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = 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.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.
"""
if (
prompt is not None
and isinstance(prompt, str)
or isinstance(prompt, Image.Image)
):
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:
prompt_embeds = self.prompt_encoder.encode_prompt(prompt)
if self.prompt_encoder is not None:
prompt_embeds_dtype = self.prompt_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
)
if do_classifier_free_guidance and negative_prompt_embeds is None:
negative_prompt_embeds = self.prompt_encoder.encode_prompt(
[""] * batch_size # TODO: Make this customizable
)
# get unconditional embeddings for classifier free guidance
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
)
return prompt_embeds, negative_prompt_embeds
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://arxiv.org/abs/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=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 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 prepare_latents(
self,
batch_size,
num_channels_latents,
height,
width,
dtype,
device,
generator,
latents=None,
):
shape = (
batch_size,
num_channels_latents,
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
def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
r"""Enables the FreeU mechanism as in https://arxiv.org/abs/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()
# Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
"""
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
Args:
timesteps (`torch.Tensor`):
generate embedding vectors at these timesteps
embedding_dim (`int`, *optional*, defaults to 512):
dimension of the embeddings to generate
dtype:
data type of the generated embeddings
Returns:
`torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), 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
# def patch image
def patch_image(
self,
image: torch.FloatTensor,
patch_size: int,
overlap: float = 0.5,
) -> torch.FloatTensor:
r"""
Patch the input image into smaller patches.
Args:
image (`torch.Tensor`):
The input image tensor to be patched. The tensor should have shape `(B, C, H, W)`.
patch_size (`int`):
The size of the patch.
overlap (`float`, *optional*, defaults to `0.25`):
The overlap between patches.
Returns:
`torch.Tensor`:
The patched image tensor.
"""
# Get the number of channels
B, C, H, W = image.shape
# Calculate the stride for unfolding
stride = int(patch_size * (1 - overlap))
# Calculate required padding for height and width
pad_height = (H - patch_size) % stride
pad_width = (W - patch_size) % stride
# Adjust padding to fully cover the image dimensions
if pad_height > 0:
pad_height = stride - pad_height
if pad_width > 0:
pad_width = stride - pad_width
# Apply padding symmetrically to the bottom and right sides
image = F.pad(image, (0, pad_width, 0, pad_height), mode="circular", value=0)
H_padded, W_padded = image.shape[-2:]
# Unfold the padded image tensor into patches
image = image.unfold(2, patch_size, stride).unfold(3, patch_size, stride)
image = image.permute(0, 2, 3, 1, 4, 5)
image = image.reshape(-1, C, patch_size, patch_size)
return image, (H_padded, W_padded)
# def unpatch image with overlap
def unpatch_image(
self,
patches: torch.FloatTensor,
batch_size: int,
output_size: Tuple[int, int],
patch_size: int,
crop_size: Optional[Tuple[int, int]] = None,
overlap: float = 0.25,
) -> torch.FloatTensor:
"""
Reconstruct the original image from its patches using fold, averaging the overlaps.
Args:
patches (torch.Tensor): The patches of the image with shape `(B, C, H, W)`,
where `B` is the effective batch size (number of patches),
`C` is the channel depth, and `H`, `W` are the patch height and width.
batch_size (int): The effective batch size (number of patches).
output_size (tuple): The height and width of the original image before patching.
patch_size (int): The height and width of each patch (assuming square patches).
crop_size (tuple, *optional*): The height and width of the cropped image.
overlap (`float`, *optional*, defaults to `0.25`):
The overlap between patches.
Returns:
torch.Tensor: The reconstructed images of shape `(B, C, H, W)`.
"""
# Set crop size if not provided
if crop_size is None:
crop_size = output_size
# Calculate the stride for folding
stride = int(patch_size * (1 - overlap))
# Calculate the number of patches per image
num_patches_per_image = patches.shape[0] // batch_size
patches = patches.view(
batch_size, num_patches_per_image, patches.shape[1], patch_size, patch_size
)
patches = patches.permute(0, 2, 3, 4, 1).contiguous()
patches = patches.view(
batch_size, patches.shape[1] * patch_size * patch_size, -1
)
# Use fold to reconstruct the images
reconstructed = F.fold(
patches, output_size=output_size, kernel_size=patch_size, stride=stride
)
# For averaging the overlaps, create a tensor of ones and fold it
mask = torch.ones_like(patches)
mask = F.fold(
mask, output_size=output_size, kernel_size=patch_size, stride=stride
)
# Average the accumulated values in the overlaps
reconstructed /= mask
# Crop the reconstructed image to the desired size
reconstructed = reconstructed[..., : crop_size[0], : crop_size[1]]
return reconstructed
@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://arxiv.org/pdf/2205.11487.pdf . `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: Union[
str, List[str], PipelineImageInput, List[PipelineImageInput]
] = None,
height: Optional[int] = None,
width: Optional[int] = None,
tileable: bool = True,
patched: bool = True,
num_inference_steps: int = 50,
timesteps: List[int] = None,
guidance_scale: float = 7.5,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
latents: Optional[torch.FloatTensor] = None,
prompt_embeds: Optional[torch.FloatTensor] = None,
negative_prompt_embeds: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
guidance_rescale: float = 0.0,
**kwargs,
):
# 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,
height,
width,
negative_prompt,
prompt_embeds,
negative_prompt_embeds,
)
self._guidance_scale = guidance_scale
self._guidance_rescale = guidance_rescale
self._cross_attention_kwargs = cross_attention_kwargs
self._interrupt = False
# 2. Define call parameters
if prompt is not None and (
isinstance(prompt, str) or isinstance(prompt, Image.Image)
):
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 = 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,
)
# 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
timesteps, num_inference_steps = retrieve_timesteps(
self.scheduler, num_inference_steps, device, 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,
prompt_embeds.dtype,
device,
generator,
latents,
)
# 6. Prepare extra step kwargs.
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. Denoising loop
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
# If patched diffusion
if patched:
B = latents.shape[0]
# patch the latents
latents, size_padded = self.patch_image(
latents, patch_size=32, overlap=0.0
)
# TODO: Improve prompt repeat when patching
Bp = latents.shape[0]
if prompt_embeds.shape[0] != Bp * 2:
prompt_embeds = prompt_embeds.repeat_interleave(Bp // B, dim=0)
# 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,
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://arxiv.org/pdf/2305.08891.pdf
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 patched:
# unpatch the latents
latents = self.unpatch_image(
latents, B, size_padded, patch_size=32, overlap=0.0
)
# noise rolling, baby!
# Based on 5.1. in https://arxiv.org/pdf/2309.01700.pdf
if tileable:
roll_h = torch.randint(0, height, (1,)).item()
roll_w = torch.randint(0, width, (1,)).item()
latents = torch.roll(latents, shifts=(roll_h, roll_w), dims=(2, 3))
# call the callback, if provided
if i == len(timesteps) - 1 or (i + 1) % self.scheduler.order == 0:
progress_bar.update()
if not output_type == "latent":
if tileable:
# decode padded latent to preserve tileability
l_height = height // self.vae_scale_factor
l_width = width // self.vae_scale_factor
latents = TF.center_crop(
latents.repeat(1, 1, 3, 3), (l_height + 4, l_width + 4)
)
# decode the latents
image = self.vae.decode(
latents / self.vae.config.scaling_factor,
return_dict=False,
generator=generator,
)[0]
# crop to original size
image = TF.center_crop(image, (height, width))
else:
image = latents
image.to(torch.float32)
image = postprocess(image, output_type=output_type)
# Offload all models
self.maybe_free_model_hooks()
if not return_dict:
return image
return MatForgerPipelineOutput(images=image)
|