Update pipeline.py
Browse files- pipeline.py +294 -325
pipeline.py
CHANGED
@@ -6,13 +6,10 @@ import numpy as np
|
|
6 |
import torch
|
7 |
|
8 |
import PIL
|
9 |
-
from diffusers
|
10 |
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
11 |
-
from diffusers.
|
12 |
-
from diffusers.
|
13 |
-
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
|
14 |
-
from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
|
15 |
-
from diffusers.utils import deprecate, is_accelerate_available, logging
|
16 |
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
|
17 |
|
18 |
|
@@ -124,7 +121,7 @@ def parse_prompt_attention(text):
|
|
124 |
return res
|
125 |
|
126 |
|
127 |
-
def get_prompts_with_weights(pipe:
|
128 |
r"""
|
129 |
Tokenize a list of prompts and return its tokens with weights of each token.
|
130 |
|
@@ -185,7 +182,7 @@ def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, no_boseos_midd
|
|
185 |
|
186 |
|
187 |
def get_unweighted_text_embeddings(
|
188 |
-
pipe:
|
189 |
text_input: torch.Tensor,
|
190 |
chunk_length: int,
|
191 |
no_boseos_middle: Optional[bool] = True,
|
@@ -225,10 +222,10 @@ def get_unweighted_text_embeddings(
|
|
225 |
|
226 |
|
227 |
def get_weighted_text_embeddings(
|
228 |
-
pipe:
|
229 |
prompt: Union[str, List[str]],
|
230 |
uncond_prompt: Optional[Union[str, List[str]]] = None,
|
231 |
-
max_embeddings_multiples: Optional[int] =
|
232 |
no_boseos_middle: Optional[bool] = False,
|
233 |
skip_parsing: Optional[bool] = False,
|
234 |
skip_weighting: Optional[bool] = False,
|
@@ -242,14 +239,14 @@ def get_weighted_text_embeddings(
|
|
242 |
Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean.
|
243 |
|
244 |
Args:
|
245 |
-
pipe (`
|
246 |
Pipe to provide access to the tokenizer and the text encoder.
|
247 |
prompt (`str` or `List[str]`):
|
248 |
The prompt or prompts to guide the image generation.
|
249 |
uncond_prompt (`str` or `List[str]`):
|
250 |
The unconditional prompt or prompts for guide the image generation. If unconditional prompt
|
251 |
is provided, the embeddings of prompt and uncond_prompt are concatenated.
|
252 |
-
max_embeddings_multiples (`int`, *optional*, defaults to `
|
253 |
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
254 |
no_boseos_middle (`bool`, *optional*, defaults to `False`):
|
255 |
If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and
|
@@ -358,18 +355,18 @@ def get_weighted_text_embeddings(
|
|
358 |
def preprocess_image(image):
|
359 |
w, h = image.size
|
360 |
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
361 |
-
image = image.resize((w, h), resample=
|
362 |
image = np.array(image).astype(np.float32) / 255.0
|
363 |
image = image[None].transpose(0, 3, 1, 2)
|
364 |
image = torch.from_numpy(image)
|
365 |
return 2.0 * image - 1.0
|
366 |
|
367 |
|
368 |
-
def preprocess_mask(mask):
|
369 |
mask = mask.convert("L")
|
370 |
w, h = mask.size
|
371 |
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
372 |
-
mask = mask.resize((w //
|
373 |
mask = np.array(mask).astype(np.float32) / 255.0
|
374 |
mask = np.tile(mask, (4, 1, 1))
|
375 |
mask = mask[None].transpose(0, 1, 2, 3) # what does this step do?
|
@@ -378,7 +375,7 @@ def preprocess_mask(mask):
|
|
378 |
return mask
|
379 |
|
380 |
|
381 |
-
class StableDiffusionLongPromptWeightingPipeline(
|
382 |
r"""
|
383 |
Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing
|
384 |
weighting in prompt.
|
@@ -398,7 +395,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
398 |
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
|
399 |
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
|
400 |
scheduler ([`SchedulerMixin`]):
|
401 |
-
A scheduler to be used in combination with `unet` to denoise the encoded image
|
402 |
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
|
403 |
safety_checker ([`StableDiffusionSafetyChecker`]):
|
404 |
Classification module that estimates whether generated images could be considered offensive or harmful.
|
@@ -413,50 +410,12 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
413 |
text_encoder: CLIPTextModel,
|
414 |
tokenizer: CLIPTokenizer,
|
415 |
unet: UNet2DConditionModel,
|
416 |
-
scheduler:
|
417 |
safety_checker: StableDiffusionSafetyChecker,
|
418 |
feature_extractor: CLIPFeatureExtractor,
|
|
|
419 |
):
|
420 |
-
super().__init__(
|
421 |
-
|
422 |
-
if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
|
423 |
-
deprecation_message = (
|
424 |
-
f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
|
425 |
-
f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
|
426 |
-
"to update the config accordingly as leaving `steps_offset` might led to incorrect results"
|
427 |
-
" in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
|
428 |
-
" it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
|
429 |
-
" file"
|
430 |
-
)
|
431 |
-
deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
|
432 |
-
new_config = dict(scheduler.config)
|
433 |
-
new_config["steps_offset"] = 1
|
434 |
-
scheduler._internal_dict = FrozenDict(new_config)
|
435 |
-
|
436 |
-
if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
|
437 |
-
deprecation_message = (
|
438 |
-
f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
|
439 |
-
" `clip_sample` should be set to False in the configuration file. Please make sure to update the"
|
440 |
-
" config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
|
441 |
-
" future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
|
442 |
-
" nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
|
443 |
-
)
|
444 |
-
deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
|
445 |
-
new_config = dict(scheduler.config)
|
446 |
-
new_config["clip_sample"] = False
|
447 |
-
scheduler._internal_dict = FrozenDict(new_config)
|
448 |
-
|
449 |
-
if safety_checker is None:
|
450 |
-
logger.warn(
|
451 |
-
f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
|
452 |
-
" that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
|
453 |
-
" results in services or applications open to the public. Both the diffusers team and Hugging Face"
|
454 |
-
" strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
|
455 |
-
" it only for use-cases that involve analyzing network behavior or auditing its results. For more"
|
456 |
-
" information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
|
457 |
-
)
|
458 |
-
|
459 |
-
self.register_modules(
|
460 |
vae=vae,
|
461 |
text_encoder=text_encoder,
|
462 |
tokenizer=tokenizer,
|
@@ -464,76 +423,178 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
464 |
scheduler=scheduler,
|
465 |
safety_checker=safety_checker,
|
466 |
feature_extractor=feature_extractor,
|
|
|
467 |
)
|
468 |
|
469 |
-
def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
470 |
r"""
|
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 |
-
slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
|
496 |
-
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
|
497 |
-
a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case,
|
498 |
-
`attention_head_dim` must be a multiple of `slice_size`.
|
499 |
-
"""
|
500 |
-
if slice_size == "auto":
|
501 |
-
# half the attention head size is usually a good trade-off between
|
502 |
-
# speed and memory
|
503 |
-
slice_size = self.unet.config.attention_head_dim // 2
|
504 |
-
self.unet.set_attention_slice(slice_size)
|
505 |
|
506 |
-
def
|
507 |
-
|
508 |
-
|
509 |
-
back to computing attention in one step.
|
510 |
-
"""
|
511 |
-
# set slice_size = `None` to disable `attention slicing`
|
512 |
-
self.enable_attention_slicing(None)
|
513 |
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
if
|
521 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
522 |
else:
|
523 |
-
|
|
|
524 |
|
525 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
526 |
|
527 |
-
|
528 |
-
|
529 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
530 |
|
531 |
@torch.no_grad()
|
532 |
def __call__(
|
533 |
self,
|
534 |
prompt: Union[str, List[str]],
|
535 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
536 |
-
|
537 |
mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
538 |
height: int = 512,
|
539 |
width: int = 512,
|
@@ -561,11 +622,11 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
561 |
negative_prompt (`str` or `List[str]`, *optional*):
|
562 |
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
563 |
if `guidance_scale` is less than `1`).
|
564 |
-
|
565 |
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
566 |
process.
|
567 |
mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
568 |
-
`Image`, or tensor representing an image batch, to mask `
|
569 |
replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
|
570 |
PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
|
571 |
contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
|
@@ -583,11 +644,11 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
583 |
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
584 |
usually at the expense of lower image quality.
|
585 |
strength (`float`, *optional*, defaults to 0.8):
|
586 |
-
Conceptually, indicates how much to transform the reference `
|
587 |
-
`
|
588 |
number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
|
589 |
noise will be maximum and the denoising process will run for the full number of iterations specified in
|
590 |
-
`num_inference_steps`. A value of 1, therefore, essentially ignores `
|
591 |
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
592 |
The number of images to generate per prompt.
|
593 |
eta (`float`, *optional*, defaults to 0.0):
|
@@ -626,222 +687,115 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
626 |
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
627 |
(nsfw) content, according to the `safety_checker`.
|
628 |
"""
|
|
|
|
|
|
|
629 |
|
630 |
-
|
631 |
-
|
632 |
-
|
633 |
-
elif isinstance(prompt, list):
|
634 |
-
batch_size = len(prompt)
|
635 |
-
else:
|
636 |
-
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
637 |
|
638 |
-
|
639 |
-
|
640 |
-
|
641 |
-
if height % 8 != 0 or width % 8 != 0:
|
642 |
-
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
643 |
-
|
644 |
-
if (callback_steps is None) or (
|
645 |
-
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
646 |
-
):
|
647 |
-
raise ValueError(
|
648 |
-
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
649 |
-
f" {type(callback_steps)}."
|
650 |
-
)
|
651 |
-
|
652 |
-
# get prompt text embeddings
|
653 |
|
|
|
|
|
|
|
654 |
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
655 |
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
656 |
# corresponds to doing no classifier free guidance.
|
657 |
do_classifier_free_guidance = guidance_scale > 1.0
|
658 |
-
# get unconditional embeddings for classifier free guidance
|
659 |
-
if negative_prompt is None:
|
660 |
-
negative_prompt = [""] * batch_size
|
661 |
-
elif isinstance(negative_prompt, str):
|
662 |
-
negative_prompt = [negative_prompt] * batch_size
|
663 |
-
if batch_size != len(negative_prompt):
|
664 |
-
raise ValueError(
|
665 |
-
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
666 |
-
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
667 |
-
" the batch size of `prompt`."
|
668 |
-
)
|
669 |
|
670 |
-
|
671 |
-
|
672 |
-
prompt
|
673 |
-
|
674 |
-
|
675 |
-
|
|
|
|
|
676 |
)
|
677 |
-
|
678 |
-
|
679 |
-
|
680 |
-
|
681 |
-
|
682 |
-
|
683 |
-
|
684 |
-
|
685 |
-
|
686 |
-
|
687 |
-
|
688 |
-
|
689 |
-
|
690 |
-
latents_dtype = text_embeddings.dtype
|
691 |
-
init_latents_orig = None
|
692 |
-
mask = None
|
693 |
-
noise = None
|
694 |
-
|
695 |
-
if init_image is None:
|
696 |
-
# get the initial random noise unless the user supplied it
|
697 |
-
|
698 |
-
# Unlike in other pipelines, latents need to be generated in the target device
|
699 |
-
# for 1-to-1 results reproducibility with the CompVis implementation.
|
700 |
-
# However this currently doesn't work in `mps`.
|
701 |
-
latents_shape = (
|
702 |
-
batch_size * num_images_per_prompt,
|
703 |
-
self.unet.in_channels,
|
704 |
-
height // 8,
|
705 |
-
width // 8,
|
706 |
-
)
|
707 |
-
|
708 |
-
if latents is None:
|
709 |
-
if self.device.type == "mps":
|
710 |
-
# randn does not exist on mps
|
711 |
-
latents = torch.randn(
|
712 |
-
latents_shape,
|
713 |
-
generator=generator,
|
714 |
-
device="cpu",
|
715 |
-
dtype=latents_dtype,
|
716 |
-
).to(self.device)
|
717 |
-
else:
|
718 |
-
latents = torch.randn(
|
719 |
-
latents_shape,
|
720 |
-
generator=generator,
|
721 |
-
device=self.device,
|
722 |
-
dtype=latents_dtype,
|
723 |
-
)
|
724 |
-
else:
|
725 |
-
if latents.shape != latents_shape:
|
726 |
-
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
|
727 |
-
latents = latents.to(self.device)
|
728 |
-
|
729 |
-
timesteps = self.scheduler.timesteps.to(self.device)
|
730 |
-
|
731 |
-
# scale the initial noise by the standard deviation required by the scheduler
|
732 |
-
latents = latents * self.scheduler.init_noise_sigma
|
733 |
else:
|
734 |
-
|
735 |
-
|
736 |
-
|
737 |
-
|
738 |
-
|
739 |
-
|
740 |
-
|
741 |
-
|
742 |
-
|
743 |
-
|
744 |
-
|
745 |
-
|
746 |
-
|
747 |
-
|
748 |
-
|
749 |
-
|
750 |
-
|
751 |
-
|
752 |
-
|
753 |
-
raise ValueError("The mask and init_image should be the same size!")
|
754 |
-
|
755 |
-
# get the original timestep using init_timestep
|
756 |
-
offset = self.scheduler.config.get("steps_offset", 0)
|
757 |
-
init_timestep = int(num_inference_steps * strength) + offset
|
758 |
-
init_timestep = min(init_timestep, num_inference_steps)
|
759 |
-
|
760 |
-
timesteps = self.scheduler.timesteps[-init_timestep]
|
761 |
-
timesteps = torch.tensor([timesteps] * batch_size * num_images_per_prompt, device=self.device)
|
762 |
-
|
763 |
-
# add noise to latents using the timesteps
|
764 |
-
if self.device.type == "mps":
|
765 |
-
# randn does not exist on mps
|
766 |
-
noise = torch.randn(
|
767 |
-
init_latents.shape,
|
768 |
-
generator=generator,
|
769 |
-
device="cpu",
|
770 |
-
dtype=latents_dtype,
|
771 |
-
).to(self.device)
|
772 |
-
else:
|
773 |
-
noise = torch.randn(
|
774 |
-
init_latents.shape,
|
775 |
-
generator=generator,
|
776 |
-
device=self.device,
|
777 |
-
dtype=latents_dtype,
|
778 |
-
)
|
779 |
-
latents = self.scheduler.add_noise(init_latents, noise, timesteps)
|
780 |
-
|
781 |
-
t_start = max(num_inference_steps - init_timestep + offset, 0)
|
782 |
-
timesteps = self.scheduler.timesteps[t_start:].to(self.device)
|
783 |
-
|
784 |
-
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
785 |
-
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
786 |
-
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
787 |
-
# and should be between [0, 1]
|
788 |
-
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
789 |
-
extra_step_kwargs = {}
|
790 |
-
if accepts_eta:
|
791 |
-
extra_step_kwargs["eta"] = eta
|
792 |
-
|
793 |
-
for i, t in enumerate(self.progress_bar(timesteps)):
|
794 |
-
# expand the latents if we are doing classifier free guidance
|
795 |
-
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
796 |
-
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
797 |
-
|
798 |
-
# predict the noise residual
|
799 |
-
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
800 |
-
|
801 |
-
# perform guidance
|
802 |
-
if do_classifier_free_guidance:
|
803 |
-
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
804 |
-
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
805 |
-
|
806 |
-
# compute the previous noisy sample x_t -> x_t-1
|
807 |
-
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
808 |
-
|
809 |
-
if mask is not None:
|
810 |
-
# masking
|
811 |
-
init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t]))
|
812 |
-
latents = (init_latents_proper * mask) + (latents * (1 - mask))
|
813 |
-
|
814 |
-
# call the callback, if provided
|
815 |
-
if i % callback_steps == 0:
|
816 |
-
if callback is not None:
|
817 |
-
callback(i, t, latents)
|
818 |
-
if is_cancelled_callback is not None and is_cancelled_callback():
|
819 |
-
return None
|
820 |
-
|
821 |
-
latents = 1 / 0.18215 * latents
|
822 |
-
image = self.vae.decode(latents).sample
|
823 |
-
|
824 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
825 |
-
|
826 |
-
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
|
827 |
-
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
828 |
-
|
829 |
-
if self.safety_checker is not None:
|
830 |
-
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(
|
831 |
-
self.device
|
832 |
-
)
|
833 |
-
image, has_nsfw_concept = self.safety_checker(
|
834 |
-
images=image,
|
835 |
-
clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype),
|
836 |
-
)
|
837 |
-
else:
|
838 |
-
has_nsfw_concept = None
|
839 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
840 |
if output_type == "pil":
|
841 |
image = self.numpy_to_pil(image)
|
842 |
|
843 |
if not return_dict:
|
844 |
-
return
|
845 |
|
846 |
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
847 |
|
@@ -861,6 +815,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
861 |
output_type: Optional[str] = "pil",
|
862 |
return_dict: bool = True,
|
863 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
|
|
864 |
callback_steps: Optional[int] = 1,
|
865 |
**kwargs,
|
866 |
):
|
@@ -908,6 +863,9 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
908 |
callback (`Callable`, *optional*):
|
909 |
A function that will be called every `callback_steps` steps during inference. The function will be
|
910 |
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
|
|
|
|
|
|
911 |
callback_steps (`int`, *optional*, defaults to 1):
|
912 |
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
913 |
called at every step.
|
@@ -933,13 +891,14 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
933 |
output_type=output_type,
|
934 |
return_dict=return_dict,
|
935 |
callback=callback,
|
|
|
936 |
callback_steps=callback_steps,
|
937 |
**kwargs,
|
938 |
)
|
939 |
|
940 |
def img2img(
|
941 |
self,
|
942 |
-
|
943 |
prompt: Union[str, List[str]],
|
944 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
945 |
strength: float = 0.8,
|
@@ -952,13 +911,14 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
952 |
output_type: Optional[str] = "pil",
|
953 |
return_dict: bool = True,
|
954 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
|
|
955 |
callback_steps: Optional[int] = 1,
|
956 |
**kwargs,
|
957 |
):
|
958 |
r"""
|
959 |
Function for image-to-image generation.
|
960 |
Args:
|
961 |
-
|
962 |
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
963 |
process.
|
964 |
prompt (`str` or `List[str]`):
|
@@ -967,11 +927,11 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
967 |
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
968 |
if `guidance_scale` is less than `1`).
|
969 |
strength (`float`, *optional*, defaults to 0.8):
|
970 |
-
Conceptually, indicates how much to transform the reference `
|
971 |
-
`
|
972 |
number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
|
973 |
noise will be maximum and the denoising process will run for the full number of iterations specified in
|
974 |
-
`num_inference_steps`. A value of 1, therefore, essentially ignores `
|
975 |
num_inference_steps (`int`, *optional*, defaults to 50):
|
976 |
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
977 |
expense of slower inference. This parameter will be modulated by `strength`.
|
@@ -1000,6 +960,9 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1000 |
callback (`Callable`, *optional*):
|
1001 |
A function that will be called every `callback_steps` steps during inference. The function will be
|
1002 |
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
|
|
|
|
|
|
1003 |
callback_steps (`int`, *optional*, defaults to 1):
|
1004 |
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
1005 |
called at every step.
|
@@ -1013,7 +976,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1013 |
return self.__call__(
|
1014 |
prompt=prompt,
|
1015 |
negative_prompt=negative_prompt,
|
1016 |
-
|
1017 |
num_inference_steps=num_inference_steps,
|
1018 |
guidance_scale=guidance_scale,
|
1019 |
strength=strength,
|
@@ -1024,13 +987,14 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1024 |
output_type=output_type,
|
1025 |
return_dict=return_dict,
|
1026 |
callback=callback,
|
|
|
1027 |
callback_steps=callback_steps,
|
1028 |
**kwargs,
|
1029 |
)
|
1030 |
|
1031 |
def inpaint(
|
1032 |
self,
|
1033 |
-
|
1034 |
mask_image: Union[torch.FloatTensor, PIL.Image.Image],
|
1035 |
prompt: Union[str, List[str]],
|
1036 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
@@ -1044,17 +1008,18 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1044 |
output_type: Optional[str] = "pil",
|
1045 |
return_dict: bool = True,
|
1046 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
|
|
1047 |
callback_steps: Optional[int] = 1,
|
1048 |
**kwargs,
|
1049 |
):
|
1050 |
r"""
|
1051 |
Function for inpaint.
|
1052 |
Args:
|
1053 |
-
|
1054 |
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
1055 |
process. This is the image whose masked region will be inpainted.
|
1056 |
mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
1057 |
-
`Image`, or tensor representing an image batch, to mask `
|
1058 |
replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
|
1059 |
PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
|
1060 |
contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
|
@@ -1066,7 +1031,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1066 |
strength (`float`, *optional*, defaults to 0.8):
|
1067 |
Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength`
|
1068 |
is 1, the denoising process will be run on the masked area for the full number of iterations specified
|
1069 |
-
in `num_inference_steps`. `
|
1070 |
noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur.
|
1071 |
num_inference_steps (`int`, *optional*, defaults to 50):
|
1072 |
The reference number of denoising steps. More denoising steps usually lead to a higher quality image at
|
@@ -1096,6 +1061,9 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1096 |
callback (`Callable`, *optional*):
|
1097 |
A function that will be called every `callback_steps` steps during inference. The function will be
|
1098 |
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
|
|
|
|
|
|
1099 |
callback_steps (`int`, *optional*, defaults to 1):
|
1100 |
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
1101 |
called at every step.
|
@@ -1109,7 +1077,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1109 |
return self.__call__(
|
1110 |
prompt=prompt,
|
1111 |
negative_prompt=negative_prompt,
|
1112 |
-
|
1113 |
mask_image=mask_image,
|
1114 |
num_inference_steps=num_inference_steps,
|
1115 |
guidance_scale=guidance_scale,
|
@@ -1121,6 +1089,7 @@ class StableDiffusionLongPromptWeightingPipeline(DiffusionPipeline):
|
|
1121 |
output_type=output_type,
|
1122 |
return_dict=return_dict,
|
1123 |
callback=callback,
|
|
|
1124 |
callback_steps=callback_steps,
|
1125 |
**kwargs,
|
1126 |
)
|
|
|
6 |
import torch
|
7 |
|
8 |
import PIL
|
9 |
+
from diffusers import SchedulerMixin, StableDiffusionPipeline
|
10 |
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
11 |
+
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
|
12 |
+
from diffusers.utils import PIL_INTERPOLATION, deprecate, logging
|
|
|
|
|
|
|
13 |
from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
|
14 |
|
15 |
|
|
|
121 |
return res
|
122 |
|
123 |
|
124 |
+
def get_prompts_with_weights(pipe: StableDiffusionPipeline, prompt: List[str], max_length: int):
|
125 |
r"""
|
126 |
Tokenize a list of prompts and return its tokens with weights of each token.
|
127 |
|
|
|
182 |
|
183 |
|
184 |
def get_unweighted_text_embeddings(
|
185 |
+
pipe: StableDiffusionPipeline,
|
186 |
text_input: torch.Tensor,
|
187 |
chunk_length: int,
|
188 |
no_boseos_middle: Optional[bool] = True,
|
|
|
222 |
|
223 |
|
224 |
def get_weighted_text_embeddings(
|
225 |
+
pipe: StableDiffusionPipeline,
|
226 |
prompt: Union[str, List[str]],
|
227 |
uncond_prompt: Optional[Union[str, List[str]]] = None,
|
228 |
+
max_embeddings_multiples: Optional[int] = 3,
|
229 |
no_boseos_middle: Optional[bool] = False,
|
230 |
skip_parsing: Optional[bool] = False,
|
231 |
skip_weighting: Optional[bool] = False,
|
|
|
239 |
Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean.
|
240 |
|
241 |
Args:
|
242 |
+
pipe (`StableDiffusionPipeline`):
|
243 |
Pipe to provide access to the tokenizer and the text encoder.
|
244 |
prompt (`str` or `List[str]`):
|
245 |
The prompt or prompts to guide the image generation.
|
246 |
uncond_prompt (`str` or `List[str]`):
|
247 |
The unconditional prompt or prompts for guide the image generation. If unconditional prompt
|
248 |
is provided, the embeddings of prompt and uncond_prompt are concatenated.
|
249 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `3`):
|
250 |
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
251 |
no_boseos_middle (`bool`, *optional*, defaults to `False`):
|
252 |
If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and
|
|
|
355 |
def preprocess_image(image):
|
356 |
w, h = image.size
|
357 |
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
358 |
+
image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
|
359 |
image = np.array(image).astype(np.float32) / 255.0
|
360 |
image = image[None].transpose(0, 3, 1, 2)
|
361 |
image = torch.from_numpy(image)
|
362 |
return 2.0 * image - 1.0
|
363 |
|
364 |
|
365 |
+
def preprocess_mask(mask, scale_factor=8):
|
366 |
mask = mask.convert("L")
|
367 |
w, h = mask.size
|
368 |
w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32
|
369 |
+
mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"])
|
370 |
mask = np.array(mask).astype(np.float32) / 255.0
|
371 |
mask = np.tile(mask, (4, 1, 1))
|
372 |
mask = mask[None].transpose(0, 1, 2, 3) # what does this step do?
|
|
|
375 |
return mask
|
376 |
|
377 |
|
378 |
+
class StableDiffusionLongPromptWeightingPipeline(StableDiffusionPipeline):
|
379 |
r"""
|
380 |
Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing
|
381 |
weighting in prompt.
|
|
|
395 |
[CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
|
396 |
unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
|
397 |
scheduler ([`SchedulerMixin`]):
|
398 |
+
A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
|
399 |
[`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
|
400 |
safety_checker ([`StableDiffusionSafetyChecker`]):
|
401 |
Classification module that estimates whether generated images could be considered offensive or harmful.
|
|
|
410 |
text_encoder: CLIPTextModel,
|
411 |
tokenizer: CLIPTokenizer,
|
412 |
unet: UNet2DConditionModel,
|
413 |
+
scheduler: SchedulerMixin,
|
414 |
safety_checker: StableDiffusionSafetyChecker,
|
415 |
feature_extractor: CLIPFeatureExtractor,
|
416 |
+
requires_safety_checker: bool = True,
|
417 |
):
|
418 |
+
super().__init__(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
419 |
vae=vae,
|
420 |
text_encoder=text_encoder,
|
421 |
tokenizer=tokenizer,
|
|
|
423 |
scheduler=scheduler,
|
424 |
safety_checker=safety_checker,
|
425 |
feature_extractor=feature_extractor,
|
426 |
+
requires_safety_checker=requires_safety_checker,
|
427 |
)
|
428 |
|
429 |
+
def _encode_prompt(
|
430 |
+
self,
|
431 |
+
prompt,
|
432 |
+
device,
|
433 |
+
num_images_per_prompt,
|
434 |
+
do_classifier_free_guidance,
|
435 |
+
negative_prompt,
|
436 |
+
max_embeddings_multiples,
|
437 |
+
):
|
438 |
r"""
|
439 |
+
Encodes the prompt into text encoder hidden states.
|
440 |
|
441 |
+
Args:
|
442 |
+
prompt (`str` or `list(int)`):
|
443 |
+
prompt to be encoded
|
444 |
+
device: (`torch.device`):
|
445 |
+
torch device
|
446 |
+
num_images_per_prompt (`int`):
|
447 |
+
number of images that should be generated per prompt
|
448 |
+
do_classifier_free_guidance (`bool`):
|
449 |
+
whether to use classifier free guidance or not
|
450 |
+
negative_prompt (`str` or `List[str]`):
|
451 |
+
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
452 |
+
if `guidance_scale` is less than `1`).
|
453 |
+
max_embeddings_multiples (`int`, *optional*, defaults to `3`):
|
454 |
+
The max multiple length of prompt embeddings compared to the max output length of text encoder.
|
455 |
"""
|
456 |
+
batch_size = len(prompt) if isinstance(prompt, list) else 1
|
457 |
|
458 |
+
if negative_prompt is None:
|
459 |
+
negative_prompt = [""] * batch_size
|
460 |
+
elif isinstance(negative_prompt, str):
|
461 |
+
negative_prompt = [negative_prompt] * batch_size
|
462 |
+
if batch_size != len(negative_prompt):
|
463 |
+
raise ValueError(
|
464 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
465 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
466 |
+
" the batch size of `prompt`."
|
467 |
+
)
|
468 |
|
469 |
+
text_embeddings, uncond_embeddings = get_weighted_text_embeddings(
|
470 |
+
pipe=self,
|
471 |
+
prompt=prompt,
|
472 |
+
uncond_prompt=negative_prompt if do_classifier_free_guidance else None,
|
473 |
+
max_embeddings_multiples=max_embeddings_multiples,
|
474 |
+
)
|
475 |
+
bs_embed, seq_len, _ = text_embeddings.shape
|
476 |
+
text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
|
477 |
+
text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
478 |
|
479 |
+
if do_classifier_free_guidance:
|
480 |
+
bs_embed, seq_len, _ = uncond_embeddings.shape
|
481 |
+
uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
|
482 |
+
uncond_embeddings = uncond_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
483 |
+
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
|
484 |
|
485 |
+
return text_embeddings
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
486 |
|
487 |
+
def check_inputs(self, prompt, height, width, strength, callback_steps):
|
488 |
+
if not isinstance(prompt, str) and not isinstance(prompt, list):
|
489 |
+
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
|
|
|
|
|
|
|
|
490 |
|
491 |
+
if strength < 0 or strength > 1:
|
492 |
+
raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
|
493 |
+
|
494 |
+
if height % 8 != 0 or width % 8 != 0:
|
495 |
+
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
496 |
+
|
497 |
+
if (callback_steps is None) or (
|
498 |
+
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
|
499 |
+
):
|
500 |
+
raise ValueError(
|
501 |
+
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
502 |
+
f" {type(callback_steps)}."
|
503 |
+
)
|
504 |
+
|
505 |
+
def get_timesteps(self, num_inference_steps, strength, device, is_text2img):
|
506 |
+
if is_text2img:
|
507 |
+
return self.scheduler.timesteps.to(device), num_inference_steps
|
508 |
+
else:
|
509 |
+
# get the original timestep using init_timestep
|
510 |
+
offset = self.scheduler.config.get("steps_offset", 0)
|
511 |
+
init_timestep = int(num_inference_steps * strength) + offset
|
512 |
+
init_timestep = min(init_timestep, num_inference_steps)
|
513 |
+
|
514 |
+
t_start = max(num_inference_steps - init_timestep + offset, 0)
|
515 |
+
timesteps = self.scheduler.timesteps[t_start:].to(device)
|
516 |
+
return timesteps, num_inference_steps - t_start
|
517 |
+
|
518 |
+
def run_safety_checker(self, image, device, dtype):
|
519 |
+
if self.safety_checker is not None:
|
520 |
+
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
|
521 |
+
image, has_nsfw_concept = self.safety_checker(
|
522 |
+
images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
|
523 |
+
)
|
524 |
else:
|
525 |
+
has_nsfw_concept = None
|
526 |
+
return image, has_nsfw_concept
|
527 |
|
528 |
+
def decode_latents(self, latents):
|
529 |
+
latents = 1 / 0.18215 * latents
|
530 |
+
image = self.vae.decode(latents).sample
|
531 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
532 |
+
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
|
533 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
534 |
+
return image
|
535 |
|
536 |
+
def prepare_extra_step_kwargs(self, generator, eta):
|
537 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
538 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
539 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
540 |
+
# and should be between [0, 1]
|
541 |
+
|
542 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
543 |
+
extra_step_kwargs = {}
|
544 |
+
if accepts_eta:
|
545 |
+
extra_step_kwargs["eta"] = eta
|
546 |
+
|
547 |
+
# check if the scheduler accepts generator
|
548 |
+
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
549 |
+
if accepts_generator:
|
550 |
+
extra_step_kwargs["generator"] = generator
|
551 |
+
return extra_step_kwargs
|
552 |
+
|
553 |
+
def prepare_latents(self, image, timestep, batch_size, height, width, dtype, device, generator, latents=None):
|
554 |
+
if image is None:
|
555 |
+
shape = (
|
556 |
+
batch_size,
|
557 |
+
self.unet.in_channels,
|
558 |
+
height // self.vae_scale_factor,
|
559 |
+
width // self.vae_scale_factor,
|
560 |
+
)
|
561 |
+
|
562 |
+
if latents is None:
|
563 |
+
if device.type == "mps":
|
564 |
+
# randn does not work reproducibly on mps
|
565 |
+
latents = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
|
566 |
+
else:
|
567 |
+
latents = torch.randn(shape, generator=generator, device=device, dtype=dtype)
|
568 |
+
else:
|
569 |
+
if latents.shape != shape:
|
570 |
+
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
|
571 |
+
latents = latents.to(device)
|
572 |
+
|
573 |
+
# scale the initial noise by the standard deviation required by the scheduler
|
574 |
+
latents = latents * self.scheduler.init_noise_sigma
|
575 |
+
return latents, None, None
|
576 |
+
else:
|
577 |
+
init_latent_dist = self.vae.encode(image).latent_dist
|
578 |
+
init_latents = init_latent_dist.sample(generator=generator)
|
579 |
+
init_latents = 0.18215 * init_latents
|
580 |
+
init_latents = torch.cat([init_latents] * batch_size, dim=0)
|
581 |
+
init_latents_orig = init_latents
|
582 |
+
shape = init_latents.shape
|
583 |
+
|
584 |
+
# add noise to latents using the timesteps
|
585 |
+
if device.type == "mps":
|
586 |
+
noise = torch.randn(shape, generator=generator, device="cpu", dtype=dtype).to(device)
|
587 |
+
else:
|
588 |
+
noise = torch.randn(shape, generator=generator, device=device, dtype=dtype)
|
589 |
+
latents = self.scheduler.add_noise(init_latents, noise, timestep)
|
590 |
+
return latents, init_latents_orig, noise
|
591 |
|
592 |
@torch.no_grad()
|
593 |
def __call__(
|
594 |
self,
|
595 |
prompt: Union[str, List[str]],
|
596 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
597 |
+
image: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
598 |
mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
|
599 |
height: int = 512,
|
600 |
width: int = 512,
|
|
|
622 |
negative_prompt (`str` or `List[str]`, *optional*):
|
623 |
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
624 |
if `guidance_scale` is less than `1`).
|
625 |
+
image (`torch.FloatTensor` or `PIL.Image.Image`):
|
626 |
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
627 |
process.
|
628 |
mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
629 |
+
`Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
|
630 |
replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
|
631 |
PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
|
632 |
contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
|
|
|
644 |
1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
|
645 |
usually at the expense of lower image quality.
|
646 |
strength (`float`, *optional*, defaults to 0.8):
|
647 |
+
Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
|
648 |
+
`image` will be used as a starting point, adding more noise to it the larger the `strength`. The
|
649 |
number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
|
650 |
noise will be maximum and the denoising process will run for the full number of iterations specified in
|
651 |
+
`num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
|
652 |
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
653 |
The number of images to generate per prompt.
|
654 |
eta (`float`, *optional*, defaults to 0.0):
|
|
|
687 |
list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
|
688 |
(nsfw) content, according to the `safety_checker`.
|
689 |
"""
|
690 |
+
message = "Please use `image` instead of `init_image`."
|
691 |
+
init_image = deprecate("init_image", "0.12.0", message, take_from=kwargs)
|
692 |
+
image = init_image or image
|
693 |
|
694 |
+
# 0. Default height and width to unet
|
695 |
+
height = height or self.unet.config.sample_size * self.vae_scale_factor
|
696 |
+
width = width or self.unet.config.sample_size * self.vae_scale_factor
|
|
|
|
|
|
|
|
|
697 |
|
698 |
+
# 1. Check inputs. Raise error if not correct
|
699 |
+
self.check_inputs(prompt, height, width, strength, callback_steps)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
700 |
|
701 |
+
# 2. Define call parameters
|
702 |
+
batch_size = 1 if isinstance(prompt, str) else len(prompt)
|
703 |
+
device = self._execution_device
|
704 |
# here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
|
705 |
# of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
|
706 |
# corresponds to doing no classifier free guidance.
|
707 |
do_classifier_free_guidance = guidance_scale > 1.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
708 |
|
709 |
+
# 3. Encode input prompt
|
710 |
+
text_embeddings = self._encode_prompt(
|
711 |
+
prompt,
|
712 |
+
device,
|
713 |
+
num_images_per_prompt,
|
714 |
+
do_classifier_free_guidance,
|
715 |
+
negative_prompt,
|
716 |
+
max_embeddings_multiples,
|
717 |
)
|
718 |
+
dtype = text_embeddings.dtype
|
719 |
+
|
720 |
+
# 4. Preprocess image and mask
|
721 |
+
if isinstance(image, PIL.Image.Image):
|
722 |
+
image = preprocess_image(image)
|
723 |
+
if image is not None:
|
724 |
+
image = image.to(device=self.device, dtype=dtype)
|
725 |
+
if isinstance(mask_image, PIL.Image.Image):
|
726 |
+
mask_image = preprocess_mask(mask_image, self.vae_scale_factor)
|
727 |
+
if mask_image is not None:
|
728 |
+
mask = mask_image.to(device=self.device, dtype=dtype)
|
729 |
+
mask = torch.cat([mask] * batch_size * num_images_per_prompt)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
730 |
else:
|
731 |
+
mask = None
|
732 |
+
|
733 |
+
# 5. set timesteps
|
734 |
+
self.scheduler.set_timesteps(num_inference_steps, device=device)
|
735 |
+
timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device, image is None)
|
736 |
+
latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
|
737 |
+
|
738 |
+
# 6. Prepare latent variables
|
739 |
+
latents, init_latents_orig, noise = self.prepare_latents(
|
740 |
+
image,
|
741 |
+
latent_timestep,
|
742 |
+
batch_size * num_images_per_prompt,
|
743 |
+
height,
|
744 |
+
width,
|
745 |
+
dtype,
|
746 |
+
device,
|
747 |
+
generator,
|
748 |
+
latents,
|
749 |
+
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
750 |
|
751 |
+
# 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
752 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
753 |
+
|
754 |
+
# 8. Denoising loop
|
755 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
756 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
757 |
+
for i, t in enumerate(timesteps):
|
758 |
+
# expand the latents if we are doing classifier free guidance
|
759 |
+
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
760 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
761 |
+
|
762 |
+
# predict the noise residual
|
763 |
+
noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
|
764 |
+
|
765 |
+
# perform guidance
|
766 |
+
if do_classifier_free_guidance:
|
767 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
768 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
769 |
+
|
770 |
+
# compute the previous noisy sample x_t -> x_t-1
|
771 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
|
772 |
+
|
773 |
+
if mask is not None:
|
774 |
+
# masking
|
775 |
+
init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t]))
|
776 |
+
latents = (init_latents_proper * mask) + (latents * (1 - mask))
|
777 |
+
|
778 |
+
# call the callback, if provided
|
779 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
780 |
+
progress_bar.update()
|
781 |
+
if i % callback_steps == 0:
|
782 |
+
if callback is not None:
|
783 |
+
callback(i, t, latents)
|
784 |
+
if is_cancelled_callback is not None and is_cancelled_callback():
|
785 |
+
return None
|
786 |
+
|
787 |
+
# 9. Post-processing
|
788 |
+
image = self.decode_latents(latents)
|
789 |
+
|
790 |
+
# 10. Run safety checker
|
791 |
+
image, has_nsfw_concept = self.run_safety_checker(image, device, text_embeddings.dtype)
|
792 |
+
|
793 |
+
# 11. Convert to PIL
|
794 |
if output_type == "pil":
|
795 |
image = self.numpy_to_pil(image)
|
796 |
|
797 |
if not return_dict:
|
798 |
+
return image, has_nsfw_concept
|
799 |
|
800 |
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
801 |
|
|
|
815 |
output_type: Optional[str] = "pil",
|
816 |
return_dict: bool = True,
|
817 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
818 |
+
is_cancelled_callback: Optional[Callable[[], bool]] = None,
|
819 |
callback_steps: Optional[int] = 1,
|
820 |
**kwargs,
|
821 |
):
|
|
|
863 |
callback (`Callable`, *optional*):
|
864 |
A function that will be called every `callback_steps` steps during inference. The function will be
|
865 |
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
866 |
+
is_cancelled_callback (`Callable`, *optional*):
|
867 |
+
A function that will be called every `callback_steps` steps during inference. If the function returns
|
868 |
+
`True`, the inference will be cancelled.
|
869 |
callback_steps (`int`, *optional*, defaults to 1):
|
870 |
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
871 |
called at every step.
|
|
|
891 |
output_type=output_type,
|
892 |
return_dict=return_dict,
|
893 |
callback=callback,
|
894 |
+
is_cancelled_callback=is_cancelled_callback,
|
895 |
callback_steps=callback_steps,
|
896 |
**kwargs,
|
897 |
)
|
898 |
|
899 |
def img2img(
|
900 |
self,
|
901 |
+
image: Union[torch.FloatTensor, PIL.Image.Image],
|
902 |
prompt: Union[str, List[str]],
|
903 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
904 |
strength: float = 0.8,
|
|
|
911 |
output_type: Optional[str] = "pil",
|
912 |
return_dict: bool = True,
|
913 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
914 |
+
is_cancelled_callback: Optional[Callable[[], bool]] = None,
|
915 |
callback_steps: Optional[int] = 1,
|
916 |
**kwargs,
|
917 |
):
|
918 |
r"""
|
919 |
Function for image-to-image generation.
|
920 |
Args:
|
921 |
+
image (`torch.FloatTensor` or `PIL.Image.Image`):
|
922 |
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
923 |
process.
|
924 |
prompt (`str` or `List[str]`):
|
|
|
927 |
The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
|
928 |
if `guidance_scale` is less than `1`).
|
929 |
strength (`float`, *optional*, defaults to 0.8):
|
930 |
+
Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1.
|
931 |
+
`image` will be used as a starting point, adding more noise to it the larger the `strength`. The
|
932 |
number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added
|
933 |
noise will be maximum and the denoising process will run for the full number of iterations specified in
|
934 |
+
`num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
|
935 |
num_inference_steps (`int`, *optional*, defaults to 50):
|
936 |
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
937 |
expense of slower inference. This parameter will be modulated by `strength`.
|
|
|
960 |
callback (`Callable`, *optional*):
|
961 |
A function that will be called every `callback_steps` steps during inference. The function will be
|
962 |
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
963 |
+
is_cancelled_callback (`Callable`, *optional*):
|
964 |
+
A function that will be called every `callback_steps` steps during inference. If the function returns
|
965 |
+
`True`, the inference will be cancelled.
|
966 |
callback_steps (`int`, *optional*, defaults to 1):
|
967 |
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
968 |
called at every step.
|
|
|
976 |
return self.__call__(
|
977 |
prompt=prompt,
|
978 |
negative_prompt=negative_prompt,
|
979 |
+
image=image,
|
980 |
num_inference_steps=num_inference_steps,
|
981 |
guidance_scale=guidance_scale,
|
982 |
strength=strength,
|
|
|
987 |
output_type=output_type,
|
988 |
return_dict=return_dict,
|
989 |
callback=callback,
|
990 |
+
is_cancelled_callback=is_cancelled_callback,
|
991 |
callback_steps=callback_steps,
|
992 |
**kwargs,
|
993 |
)
|
994 |
|
995 |
def inpaint(
|
996 |
self,
|
997 |
+
image: Union[torch.FloatTensor, PIL.Image.Image],
|
998 |
mask_image: Union[torch.FloatTensor, PIL.Image.Image],
|
999 |
prompt: Union[str, List[str]],
|
1000 |
negative_prompt: Optional[Union[str, List[str]]] = None,
|
|
|
1008 |
output_type: Optional[str] = "pil",
|
1009 |
return_dict: bool = True,
|
1010 |
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
|
1011 |
+
is_cancelled_callback: Optional[Callable[[], bool]] = None,
|
1012 |
callback_steps: Optional[int] = 1,
|
1013 |
**kwargs,
|
1014 |
):
|
1015 |
r"""
|
1016 |
Function for inpaint.
|
1017 |
Args:
|
1018 |
+
image (`torch.FloatTensor` or `PIL.Image.Image`):
|
1019 |
`Image`, or tensor representing an image batch, that will be used as the starting point for the
|
1020 |
process. This is the image whose masked region will be inpainted.
|
1021 |
mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
|
1022 |
+
`Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
|
1023 |
replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
|
1024 |
PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
|
1025 |
contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.
|
|
|
1031 |
strength (`float`, *optional*, defaults to 0.8):
|
1032 |
Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength`
|
1033 |
is 1, the denoising process will be run on the masked area for the full number of iterations specified
|
1034 |
+
in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more
|
1035 |
noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur.
|
1036 |
num_inference_steps (`int`, *optional*, defaults to 50):
|
1037 |
The reference number of denoising steps. More denoising steps usually lead to a higher quality image at
|
|
|
1061 |
callback (`Callable`, *optional*):
|
1062 |
A function that will be called every `callback_steps` steps during inference. The function will be
|
1063 |
called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
|
1064 |
+
is_cancelled_callback (`Callable`, *optional*):
|
1065 |
+
A function that will be called every `callback_steps` steps during inference. If the function returns
|
1066 |
+
`True`, the inference will be cancelled.
|
1067 |
callback_steps (`int`, *optional*, defaults to 1):
|
1068 |
The frequency at which the `callback` function will be called. If not specified, the callback will be
|
1069 |
called at every step.
|
|
|
1077 |
return self.__call__(
|
1078 |
prompt=prompt,
|
1079 |
negative_prompt=negative_prompt,
|
1080 |
+
image=image,
|
1081 |
mask_image=mask_image,
|
1082 |
num_inference_steps=num_inference_steps,
|
1083 |
guidance_scale=guidance_scale,
|
|
|
1089 |
output_type=output_type,
|
1090 |
return_dict=return_dict,
|
1091 |
callback=callback,
|
1092 |
+
is_cancelled_callback=is_cancelled_callback,
|
1093 |
callback_steps=callback_steps,
|
1094 |
**kwargs,
|
1095 |
)
|