diffusers 0.15.1__py3-none-any.whl → 0.16.1__py3-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (57) hide show
  1. diffusers/__init__.py +7 -2
  2. diffusers/configuration_utils.py +4 -0
  3. diffusers/loaders.py +262 -12
  4. diffusers/models/attention.py +31 -12
  5. diffusers/models/attention_processor.py +189 -0
  6. diffusers/models/controlnet.py +9 -2
  7. diffusers/models/embeddings.py +66 -0
  8. diffusers/models/modeling_pytorch_flax_utils.py +6 -0
  9. diffusers/models/modeling_utils.py +5 -2
  10. diffusers/models/transformer_2d.py +1 -1
  11. diffusers/models/unet_2d_condition.py +45 -6
  12. diffusers/models/vae.py +3 -0
  13. diffusers/pipelines/__init__.py +8 -0
  14. diffusers/pipelines/alt_diffusion/modeling_roberta_series.py +25 -10
  15. diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion.py +8 -0
  16. diffusers/pipelines/alt_diffusion/pipeline_alt_diffusion_img2img.py +8 -0
  17. diffusers/pipelines/audioldm/pipeline_audioldm.py +1 -1
  18. diffusers/pipelines/deepfloyd_if/__init__.py +54 -0
  19. diffusers/pipelines/deepfloyd_if/pipeline_if.py +854 -0
  20. diffusers/pipelines/deepfloyd_if/pipeline_if_img2img.py +979 -0
  21. diffusers/pipelines/deepfloyd_if/pipeline_if_img2img_superresolution.py +1097 -0
  22. diffusers/pipelines/deepfloyd_if/pipeline_if_inpainting.py +1098 -0
  23. diffusers/pipelines/deepfloyd_if/pipeline_if_inpainting_superresolution.py +1208 -0
  24. diffusers/pipelines/deepfloyd_if/pipeline_if_superresolution.py +947 -0
  25. diffusers/pipelines/deepfloyd_if/safety_checker.py +59 -0
  26. diffusers/pipelines/deepfloyd_if/timesteps.py +579 -0
  27. diffusers/pipelines/deepfloyd_if/watermark.py +46 -0
  28. diffusers/pipelines/pipeline_utils.py +54 -25
  29. diffusers/pipelines/stable_diffusion/convert_from_ckpt.py +37 -20
  30. diffusers/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_controlnet.py +1 -1
  31. diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py +12 -1
  32. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py +10 -2
  33. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_attend_and_excite.py +10 -8
  34. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_controlnet.py +59 -4
  35. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py +9 -2
  36. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py +10 -2
  37. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py +9 -2
  38. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint_legacy.py +22 -12
  39. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py +9 -2
  40. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_pix2pix_zero.py +34 -30
  41. diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py +93 -10
  42. diffusers/pipelines/versatile_diffusion/modeling_text_unet.py +45 -6
  43. diffusers/schedulers/scheduling_ddpm.py +63 -16
  44. diffusers/schedulers/scheduling_heun_discrete.py +51 -1
  45. diffusers/utils/__init__.py +4 -1
  46. diffusers/utils/dummy_torch_and_transformers_objects.py +80 -5
  47. diffusers/utils/dynamic_modules_utils.py +1 -1
  48. diffusers/utils/hub_utils.py +4 -1
  49. diffusers/utils/import_utils.py +41 -0
  50. diffusers/utils/pil_utils.py +24 -0
  51. diffusers/utils/testing_utils.py +10 -0
  52. {diffusers-0.15.1.dist-info → diffusers-0.16.1.dist-info}/METADATA +1 -1
  53. {diffusers-0.15.1.dist-info → diffusers-0.16.1.dist-info}/RECORD +57 -47
  54. {diffusers-0.15.1.dist-info → diffusers-0.16.1.dist-info}/LICENSE +0 -0
  55. {diffusers-0.15.1.dist-info → diffusers-0.16.1.dist-info}/WHEEL +0 -0
  56. {diffusers-0.15.1.dist-info → diffusers-0.16.1.dist-info}/entry_points.txt +0 -0
  57. {diffusers-0.15.1.dist-info → diffusers-0.16.1.dist-info}/top_level.txt +0 -0
@@ -0,0 +1,979 @@
1
+ import html
2
+ import inspect
3
+ import re
4
+ import urllib.parse as ul
5
+ from typing import Any, Callable, Dict, List, Optional, Union
6
+
7
+ import numpy as np
8
+ import PIL
9
+ import torch
10
+ from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
11
+
12
+ from ...models import UNet2DConditionModel
13
+ from ...schedulers import DDPMScheduler
14
+ from ...utils import (
15
+ BACKENDS_MAPPING,
16
+ PIL_INTERPOLATION,
17
+ is_accelerate_available,
18
+ is_accelerate_version,
19
+ is_bs4_available,
20
+ is_ftfy_available,
21
+ logging,
22
+ randn_tensor,
23
+ replace_example_docstring,
24
+ )
25
+ from ..pipeline_utils import DiffusionPipeline
26
+ from . import IFPipelineOutput
27
+ from .safety_checker import IFSafetyChecker
28
+ from .watermark import IFWatermarker
29
+
30
+
31
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
32
+
33
+ if is_bs4_available():
34
+ from bs4 import BeautifulSoup
35
+
36
+ if is_ftfy_available():
37
+ import ftfy
38
+
39
+
40
+ def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
41
+ w, h = images.size
42
+
43
+ coef = w / h
44
+
45
+ w, h = img_size, img_size
46
+
47
+ if coef >= 1:
48
+ w = int(round(img_size / 8 * coef) * 8)
49
+ else:
50
+ h = int(round(img_size / 8 / coef) * 8)
51
+
52
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
53
+
54
+ return images
55
+
56
+
57
+ EXAMPLE_DOC_STRING = """
58
+ Examples:
59
+ ```py
60
+ >>> from diffusers import IFImg2ImgPipeline, IFImg2ImgSuperResolutionPipeline, DiffusionPipeline
61
+ >>> from diffusers.utils import pt_to_pil
62
+ >>> import torch
63
+ >>> from PIL import Image
64
+ >>> import requests
65
+ >>> from io import BytesIO
66
+
67
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
68
+ >>> response = requests.get(url)
69
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
70
+ >>> original_image = original_image.resize((768, 512))
71
+
72
+ >>> pipe = IFImg2ImgPipeline.from_pretrained(
73
+ ... "DeepFloyd/IF-I-XL-v1.0",
74
+ ... variant="fp16",
75
+ ... torch_dtype=torch.float16,
76
+ ... )
77
+ >>> pipe.enable_model_cpu_offload()
78
+
79
+ >>> prompt = "A fantasy landscape in style minecraft"
80
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
81
+
82
+ >>> image = pipe(
83
+ ... image=original_image,
84
+ ... prompt_embeds=prompt_embeds,
85
+ ... negative_prompt_embeds=negative_embeds,
86
+ ... output_type="pt",
87
+ ... ).images
88
+
89
+ >>> # save intermediate image
90
+ >>> pil_image = pt_to_pil(image)
91
+ >>> pil_image[0].save("./if_stage_I.png")
92
+
93
+ >>> super_res_1_pipe = IFImg2ImgSuperResolutionPipeline.from_pretrained(
94
+ ... "DeepFloyd/IF-II-L-v1.0",
95
+ ... text_encoder=None,
96
+ ... variant="fp16",
97
+ ... torch_dtype=torch.float16,
98
+ ... )
99
+ >>> super_res_1_pipe.enable_model_cpu_offload()
100
+
101
+ >>> image = super_res_1_pipe(
102
+ ... image=image,
103
+ ... original_image=original_image,
104
+ ... prompt_embeds=prompt_embeds,
105
+ ... negative_prompt_embeds=negative_embeds,
106
+ ... ).images
107
+ >>> image[0].save("./if_stage_II.png")
108
+ ```
109
+ """
110
+
111
+
112
+ class IFImg2ImgPipeline(DiffusionPipeline):
113
+ tokenizer: T5Tokenizer
114
+ text_encoder: T5EncoderModel
115
+
116
+ unet: UNet2DConditionModel
117
+ scheduler: DDPMScheduler
118
+
119
+ feature_extractor: Optional[CLIPImageProcessor]
120
+ safety_checker: Optional[IFSafetyChecker]
121
+
122
+ watermarker: Optional[IFWatermarker]
123
+
124
+ bad_punct_regex = re.compile(
125
+ r"[" + "#®•©™&@·º½¾¿¡§~" + "\)" + "\(" + "\]" + "\[" + "\}" + "\{" + "\|" + "\\" + "\/" + "\*" + r"]{1,}"
126
+ ) # noqa
127
+
128
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
129
+
130
+ def __init__(
131
+ self,
132
+ tokenizer: T5Tokenizer,
133
+ text_encoder: T5EncoderModel,
134
+ unet: UNet2DConditionModel,
135
+ scheduler: DDPMScheduler,
136
+ safety_checker: Optional[IFSafetyChecker],
137
+ feature_extractor: Optional[CLIPImageProcessor],
138
+ watermarker: Optional[IFWatermarker],
139
+ requires_safety_checker: bool = True,
140
+ ):
141
+ super().__init__()
142
+
143
+ if safety_checker is None and requires_safety_checker:
144
+ logger.warning(
145
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
146
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
147
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
148
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
149
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
150
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
151
+ )
152
+
153
+ if safety_checker is not None and feature_extractor is None:
154
+ raise ValueError(
155
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
156
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
157
+ )
158
+
159
+ self.register_modules(
160
+ tokenizer=tokenizer,
161
+ text_encoder=text_encoder,
162
+ unet=unet,
163
+ scheduler=scheduler,
164
+ safety_checker=safety_checker,
165
+ feature_extractor=feature_extractor,
166
+ watermarker=watermarker,
167
+ )
168
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
169
+
170
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.enable_sequential_cpu_offload
171
+ def enable_sequential_cpu_offload(self, gpu_id=0):
172
+ r"""
173
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
174
+ models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
175
+ when their specific submodule has its `forward` method called.
176
+ """
177
+ if is_accelerate_available():
178
+ from accelerate import cpu_offload
179
+ else:
180
+ raise ImportError("Please install accelerate via `pip install accelerate`")
181
+
182
+ device = torch.device(f"cuda:{gpu_id}")
183
+
184
+ models = [
185
+ self.text_encoder,
186
+ self.unet,
187
+ ]
188
+ for cpu_offloaded_model in models:
189
+ if cpu_offloaded_model is not None:
190
+ cpu_offload(cpu_offloaded_model, device)
191
+
192
+ if self.safety_checker is not None:
193
+ cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
194
+
195
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.enable_model_cpu_offload
196
+ def enable_model_cpu_offload(self, gpu_id=0):
197
+ r"""
198
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
199
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
200
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
201
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
202
+ """
203
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
204
+ from accelerate import cpu_offload_with_hook
205
+ else:
206
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
207
+
208
+ device = torch.device(f"cuda:{gpu_id}")
209
+
210
+ if self.device.type != "cpu":
211
+ self.to("cpu", silence_dtype_warnings=True)
212
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
213
+
214
+ hook = None
215
+
216
+ if self.text_encoder is not None:
217
+ _, hook = cpu_offload_with_hook(self.text_encoder, device, prev_module_hook=hook)
218
+
219
+ # Accelerate will move the next model to the device _before_ calling the offload hook of the
220
+ # previous model. This will cause both models to be present on the device at the same time.
221
+ # IF uses T5 for its text encoder which is really large. We can manually call the offload
222
+ # hook for the text encoder to ensure it's moved to the cpu before the unet is moved to
223
+ # the GPU.
224
+ self.text_encoder_offload_hook = hook
225
+
226
+ _, hook = cpu_offload_with_hook(self.unet, device, prev_module_hook=hook)
227
+
228
+ # if the safety checker isn't called, `unet_offload_hook` will have to be called to manually offload the unet
229
+ self.unet_offload_hook = hook
230
+
231
+ if self.safety_checker is not None:
232
+ _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
233
+
234
+ # We'll offload the last model manually.
235
+ self.final_offload_hook = hook
236
+
237
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
238
+ def remove_all_hooks(self):
239
+ if is_accelerate_available():
240
+ from accelerate.hooks import remove_hook_from_module
241
+ else:
242
+ raise ImportError("Please install accelerate via `pip install accelerate`")
243
+
244
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
245
+ if model is not None:
246
+ remove_hook_from_module(model, recurse=True)
247
+
248
+ self.unet_offload_hook = None
249
+ self.text_encoder_offload_hook = None
250
+ self.final_offload_hook = None
251
+
252
+ @property
253
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
254
+ def _execution_device(self):
255
+ r"""
256
+ Returns the device on which the pipeline's models will be executed. After calling
257
+ `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
258
+ hooks.
259
+ """
260
+ if not hasattr(self.unet, "_hf_hook"):
261
+ return self.device
262
+ for module in self.unet.modules():
263
+ if (
264
+ hasattr(module, "_hf_hook")
265
+ and hasattr(module._hf_hook, "execution_device")
266
+ and module._hf_hook.execution_device is not None
267
+ ):
268
+ return torch.device(module._hf_hook.execution_device)
269
+ return self.device
270
+
271
+ @torch.no_grad()
272
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
273
+ def encode_prompt(
274
+ self,
275
+ prompt,
276
+ do_classifier_free_guidance=True,
277
+ num_images_per_prompt=1,
278
+ device=None,
279
+ negative_prompt=None,
280
+ prompt_embeds: Optional[torch.FloatTensor] = None,
281
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
282
+ clean_caption: bool = False,
283
+ ):
284
+ r"""
285
+ Encodes the prompt into text encoder hidden states.
286
+
287
+ Args:
288
+ prompt (`str` or `List[str]`, *optional*):
289
+ prompt to be encoded
290
+ device: (`torch.device`, *optional*):
291
+ torch device to place the resulting embeddings on
292
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
293
+ number of images that should be generated per prompt
294
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
295
+ whether to use classifier free guidance or not
296
+ negative_prompt (`str` or `List[str]`, *optional*):
297
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
298
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
299
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
300
+ prompt_embeds (`torch.FloatTensor`, *optional*):
301
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
302
+ provided, text embeddings will be generated from `prompt` input argument.
303
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
304
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
305
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
306
+ argument.
307
+ """
308
+ if prompt is not None and negative_prompt is not None:
309
+ if type(prompt) is not type(negative_prompt):
310
+ raise TypeError(
311
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
312
+ f" {type(prompt)}."
313
+ )
314
+
315
+ if device is None:
316
+ device = self._execution_device
317
+
318
+ if prompt is not None and isinstance(prompt, str):
319
+ batch_size = 1
320
+ elif prompt is not None and isinstance(prompt, list):
321
+ batch_size = len(prompt)
322
+ else:
323
+ batch_size = prompt_embeds.shape[0]
324
+
325
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
326
+ max_length = 77
327
+
328
+ if prompt_embeds is None:
329
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
330
+ text_inputs = self.tokenizer(
331
+ prompt,
332
+ padding="max_length",
333
+ max_length=max_length,
334
+ truncation=True,
335
+ add_special_tokens=True,
336
+ return_tensors="pt",
337
+ )
338
+ text_input_ids = text_inputs.input_ids
339
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
340
+
341
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
342
+ text_input_ids, untruncated_ids
343
+ ):
344
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
345
+ logger.warning(
346
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
347
+ f" {max_length} tokens: {removed_text}"
348
+ )
349
+
350
+ attention_mask = text_inputs.attention_mask.to(device)
351
+
352
+ prompt_embeds = self.text_encoder(
353
+ text_input_ids.to(device),
354
+ attention_mask=attention_mask,
355
+ )
356
+ prompt_embeds = prompt_embeds[0]
357
+
358
+ if self.text_encoder is not None:
359
+ dtype = self.text_encoder.dtype
360
+ elif self.unet is not None:
361
+ dtype = self.unet.dtype
362
+ else:
363
+ dtype = None
364
+
365
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
366
+
367
+ bs_embed, seq_len, _ = prompt_embeds.shape
368
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
369
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
370
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
371
+
372
+ # get unconditional embeddings for classifier free guidance
373
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
374
+ uncond_tokens: List[str]
375
+ if negative_prompt is None:
376
+ uncond_tokens = [""] * batch_size
377
+ elif isinstance(negative_prompt, str):
378
+ uncond_tokens = [negative_prompt]
379
+ elif batch_size != len(negative_prompt):
380
+ raise ValueError(
381
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
382
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
383
+ " the batch size of `prompt`."
384
+ )
385
+ else:
386
+ uncond_tokens = negative_prompt
387
+
388
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
389
+ max_length = prompt_embeds.shape[1]
390
+ uncond_input = self.tokenizer(
391
+ uncond_tokens,
392
+ padding="max_length",
393
+ max_length=max_length,
394
+ truncation=True,
395
+ return_attention_mask=True,
396
+ add_special_tokens=True,
397
+ return_tensors="pt",
398
+ )
399
+ attention_mask = uncond_input.attention_mask.to(device)
400
+
401
+ negative_prompt_embeds = self.text_encoder(
402
+ uncond_input.input_ids.to(device),
403
+ attention_mask=attention_mask,
404
+ )
405
+ negative_prompt_embeds = negative_prompt_embeds[0]
406
+
407
+ if do_classifier_free_guidance:
408
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
409
+ seq_len = negative_prompt_embeds.shape[1]
410
+
411
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
412
+
413
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
414
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
415
+
416
+ # For classifier free guidance, we need to do two forward passes.
417
+ # Here we concatenate the unconditional and text embeddings into a single batch
418
+ # to avoid doing two forward passes
419
+ else:
420
+ negative_prompt_embeds = None
421
+
422
+ return prompt_embeds, negative_prompt_embeds
423
+
424
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
425
+ def run_safety_checker(self, image, device, dtype):
426
+ if self.safety_checker is not None:
427
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
428
+ image, nsfw_detected, watermark_detected = self.safety_checker(
429
+ images=image,
430
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
431
+ )
432
+ else:
433
+ nsfw_detected = None
434
+ watermark_detected = None
435
+
436
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
437
+ self.unet_offload_hook.offload()
438
+
439
+ return image, nsfw_detected, watermark_detected
440
+
441
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
442
+ def prepare_extra_step_kwargs(self, generator, eta):
443
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
444
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
445
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
446
+ # and should be between [0, 1]
447
+
448
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
449
+ extra_step_kwargs = {}
450
+ if accepts_eta:
451
+ extra_step_kwargs["eta"] = eta
452
+
453
+ # check if the scheduler accepts generator
454
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
455
+ if accepts_generator:
456
+ extra_step_kwargs["generator"] = generator
457
+ return extra_step_kwargs
458
+
459
+ def check_inputs(
460
+ self,
461
+ prompt,
462
+ image,
463
+ batch_size,
464
+ callback_steps,
465
+ negative_prompt=None,
466
+ prompt_embeds=None,
467
+ negative_prompt_embeds=None,
468
+ ):
469
+ if (callback_steps is None) or (
470
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
471
+ ):
472
+ raise ValueError(
473
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
474
+ f" {type(callback_steps)}."
475
+ )
476
+
477
+ if prompt is not None and prompt_embeds is not None:
478
+ raise ValueError(
479
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
480
+ " only forward one of the two."
481
+ )
482
+ elif prompt is None and prompt_embeds is None:
483
+ raise ValueError(
484
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
485
+ )
486
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
487
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
488
+
489
+ if negative_prompt is not None and negative_prompt_embeds is not None:
490
+ raise ValueError(
491
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
492
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
493
+ )
494
+
495
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
496
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
497
+ raise ValueError(
498
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
499
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
500
+ f" {negative_prompt_embeds.shape}."
501
+ )
502
+
503
+ if isinstance(image, list):
504
+ check_image_type = image[0]
505
+ else:
506
+ check_image_type = image
507
+
508
+ if (
509
+ not isinstance(check_image_type, torch.Tensor)
510
+ and not isinstance(check_image_type, PIL.Image.Image)
511
+ and not isinstance(check_image_type, np.ndarray)
512
+ ):
513
+ raise ValueError(
514
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
515
+ f" {type(check_image_type)}"
516
+ )
517
+
518
+ if isinstance(image, list):
519
+ image_batch_size = len(image)
520
+ elif isinstance(image, torch.Tensor):
521
+ image_batch_size = image.shape[0]
522
+ elif isinstance(image, PIL.Image.Image):
523
+ image_batch_size = 1
524
+ elif isinstance(image, np.ndarray):
525
+ image_batch_size = image.shape[0]
526
+ else:
527
+ assert False
528
+
529
+ if batch_size != image_batch_size:
530
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
531
+
532
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
533
+ def _text_preprocessing(self, text, clean_caption=False):
534
+ if clean_caption and not is_bs4_available():
535
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
536
+ logger.warn("Setting `clean_caption` to False...")
537
+ clean_caption = False
538
+
539
+ if clean_caption and not is_ftfy_available():
540
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
541
+ logger.warn("Setting `clean_caption` to False...")
542
+ clean_caption = False
543
+
544
+ if not isinstance(text, (tuple, list)):
545
+ text = [text]
546
+
547
+ def process(text: str):
548
+ if clean_caption:
549
+ text = self._clean_caption(text)
550
+ text = self._clean_caption(text)
551
+ else:
552
+ text = text.lower().strip()
553
+ return text
554
+
555
+ return [process(t) for t in text]
556
+
557
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
558
+ def _clean_caption(self, caption):
559
+ caption = str(caption)
560
+ caption = ul.unquote_plus(caption)
561
+ caption = caption.strip().lower()
562
+ caption = re.sub("<person>", "person", caption)
563
+ # urls:
564
+ caption = re.sub(
565
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
566
+ "",
567
+ caption,
568
+ ) # regex for urls
569
+ caption = re.sub(
570
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
571
+ "",
572
+ caption,
573
+ ) # regex for urls
574
+ # html:
575
+ caption = BeautifulSoup(caption, features="html.parser").text
576
+
577
+ # @<nickname>
578
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
579
+
580
+ # 31C0—31EF CJK Strokes
581
+ # 31F0—31FF Katakana Phonetic Extensions
582
+ # 3200—32FF Enclosed CJK Letters and Months
583
+ # 3300—33FF CJK Compatibility
584
+ # 3400—4DBF CJK Unified Ideographs Extension A
585
+ # 4DC0—4DFF Yijing Hexagram Symbols
586
+ # 4E00—9FFF CJK Unified Ideographs
587
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
588
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
589
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
590
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
591
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
592
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
593
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
594
+ #######################################################
595
+
596
+ # все виды тире / all types of dash --> "-"
597
+ caption = re.sub(
598
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
599
+ "-",
600
+ caption,
601
+ )
602
+
603
+ # кавычки к одному стандарту
604
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
605
+ caption = re.sub(r"[‘’]", "'", caption)
606
+
607
+ # &quot;
608
+ caption = re.sub(r"&quot;?", "", caption)
609
+ # &amp
610
+ caption = re.sub(r"&amp", "", caption)
611
+
612
+ # ip adresses:
613
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
614
+
615
+ # article ids:
616
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
617
+
618
+ # \n
619
+ caption = re.sub(r"\\n", " ", caption)
620
+
621
+ # "#123"
622
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
623
+ # "#12345.."
624
+ caption = re.sub(r"#\d{5,}\b", "", caption)
625
+ # "123456.."
626
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
627
+ # filenames:
628
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
629
+
630
+ #
631
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
632
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
633
+
634
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
635
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
636
+
637
+ # this-is-my-cute-cat / this_is_my_cute_cat
638
+ regex2 = re.compile(r"(?:\-|\_)")
639
+ if len(re.findall(regex2, caption)) > 3:
640
+ caption = re.sub(regex2, " ", caption)
641
+
642
+ caption = ftfy.fix_text(caption)
643
+ caption = html.unescape(html.unescape(caption))
644
+
645
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
646
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
647
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
648
+
649
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
650
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
651
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
652
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
653
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
654
+
655
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
656
+
657
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
658
+
659
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
660
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
661
+ caption = re.sub(r"\s+", " ", caption)
662
+
663
+ caption.strip()
664
+
665
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
666
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
667
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
668
+ caption = re.sub(r"^\.\S+$", "", caption)
669
+
670
+ return caption.strip()
671
+
672
+ def preprocess_image(self, image: PIL.Image.Image) -> torch.Tensor:
673
+ if not isinstance(image, list):
674
+ image = [image]
675
+
676
+ def numpy_to_pt(images):
677
+ if images.ndim == 3:
678
+ images = images[..., None]
679
+
680
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
681
+ return images
682
+
683
+ if isinstance(image[0], PIL.Image.Image):
684
+ new_image = []
685
+
686
+ for image_ in image:
687
+ image_ = image_.convert("RGB")
688
+ image_ = resize(image_, self.unet.sample_size)
689
+ image_ = np.array(image_)
690
+ image_ = image_.astype(np.float32)
691
+ image_ = image_ / 127.5 - 1
692
+ new_image.append(image_)
693
+
694
+ image = new_image
695
+
696
+ image = np.stack(image, axis=0) # to np
697
+ image = numpy_to_pt(image) # to pt
698
+
699
+ elif isinstance(image[0], np.ndarray):
700
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
701
+ image = numpy_to_pt(image)
702
+
703
+ elif isinstance(image[0], torch.Tensor):
704
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
705
+
706
+ return image
707
+
708
+ def get_timesteps(self, num_inference_steps, strength):
709
+ # get the original timestep using init_timestep
710
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
711
+
712
+ t_start = max(num_inference_steps - init_timestep, 0)
713
+ timesteps = self.scheduler.timesteps[t_start:]
714
+
715
+ return timesteps, num_inference_steps - t_start
716
+
717
+ def prepare_intermediate_images(
718
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None
719
+ ):
720
+ _, channels, height, width = image.shape
721
+
722
+ batch_size = batch_size * num_images_per_prompt
723
+
724
+ shape = (batch_size, channels, height, width)
725
+
726
+ if isinstance(generator, list) and len(generator) != batch_size:
727
+ raise ValueError(
728
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
729
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
730
+ )
731
+
732
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
733
+
734
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
735
+ image = self.scheduler.add_noise(image, noise, timestep)
736
+
737
+ return image
738
+
739
+ @torch.no_grad()
740
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
741
+ def __call__(
742
+ self,
743
+ prompt: Union[str, List[str]] = None,
744
+ image: Union[
745
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
746
+ ] = None,
747
+ strength: float = 0.7,
748
+ num_inference_steps: int = 80,
749
+ timesteps: List[int] = None,
750
+ guidance_scale: float = 10.0,
751
+ negative_prompt: Optional[Union[str, List[str]]] = None,
752
+ num_images_per_prompt: Optional[int] = 1,
753
+ eta: float = 0.0,
754
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
755
+ prompt_embeds: Optional[torch.FloatTensor] = None,
756
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
757
+ output_type: Optional[str] = "pil",
758
+ return_dict: bool = True,
759
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
760
+ callback_steps: int = 1,
761
+ clean_caption: bool = True,
762
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
763
+ ):
764
+ """
765
+ Function invoked when calling the pipeline for generation.
766
+
767
+ Args:
768
+ prompt (`str` or `List[str]`, *optional*):
769
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
770
+ instead.
771
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
772
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
773
+ process.
774
+ strength (`float`, *optional*, defaults to 0.8):
775
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
776
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
777
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
778
+ be maximum and the denoising process will run for the full number of iterations specified in
779
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
780
+ num_inference_steps (`int`, *optional*, defaults to 50):
781
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
782
+ expense of slower inference.
783
+ timesteps (`List[int]`, *optional*):
784
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
785
+ timesteps are used. Must be in descending order.
786
+ guidance_scale (`float`, *optional*, defaults to 7.5):
787
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
788
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
789
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
790
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
791
+ usually at the expense of lower image quality.
792
+ negative_prompt (`str` or `List[str]`, *optional*):
793
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
794
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
795
+ less than `1`).
796
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
797
+ The number of images to generate per prompt.
798
+ eta (`float`, *optional*, defaults to 0.0):
799
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
800
+ [`schedulers.DDIMScheduler`], will be ignored for others.
801
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
802
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
803
+ to make generation deterministic.
804
+ prompt_embeds (`torch.FloatTensor`, *optional*):
805
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
806
+ provided, text embeddings will be generated from `prompt` input argument.
807
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
808
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
809
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
810
+ argument.
811
+ output_type (`str`, *optional*, defaults to `"pil"`):
812
+ The output format of the generate image. Choose between
813
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
814
+ return_dict (`bool`, *optional*, defaults to `True`):
815
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
816
+ callback (`Callable`, *optional*):
817
+ A function that will be called every `callback_steps` steps during inference. The function will be
818
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
819
+ callback_steps (`int`, *optional*, defaults to 1):
820
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
821
+ called at every step.
822
+ clean_caption (`bool`, *optional*, defaults to `True`):
823
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
824
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
825
+ prompt.
826
+ cross_attention_kwargs (`dict`, *optional*):
827
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
828
+ `self.processor` in
829
+ [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
830
+
831
+ Examples:
832
+
833
+ Returns:
834
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
835
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
836
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
837
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
838
+ or watermarked content, according to the `safety_checker`.
839
+ """
840
+ # 1. Check inputs. Raise error if not correct
841
+ if prompt is not None and isinstance(prompt, str):
842
+ batch_size = 1
843
+ elif prompt is not None and isinstance(prompt, list):
844
+ batch_size = len(prompt)
845
+ else:
846
+ batch_size = prompt_embeds.shape[0]
847
+
848
+ self.check_inputs(
849
+ prompt, image, batch_size, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
850
+ )
851
+
852
+ # 2. Define call parameters
853
+ device = self._execution_device
854
+
855
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
856
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
857
+ # corresponds to doing no classifier free guidance.
858
+ do_classifier_free_guidance = guidance_scale > 1.0
859
+
860
+ # 3. Encode input prompt
861
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
862
+ prompt,
863
+ do_classifier_free_guidance,
864
+ num_images_per_prompt=num_images_per_prompt,
865
+ device=device,
866
+ negative_prompt=negative_prompt,
867
+ prompt_embeds=prompt_embeds,
868
+ negative_prompt_embeds=negative_prompt_embeds,
869
+ clean_caption=clean_caption,
870
+ )
871
+
872
+ if do_classifier_free_guidance:
873
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
874
+
875
+ dtype = prompt_embeds.dtype
876
+
877
+ # 4. Prepare timesteps
878
+ if timesteps is not None:
879
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
880
+ timesteps = self.scheduler.timesteps
881
+ num_inference_steps = len(timesteps)
882
+ else:
883
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
884
+ timesteps = self.scheduler.timesteps
885
+
886
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
887
+
888
+ # 5. Prepare intermediate images
889
+ image = self.preprocess_image(image)
890
+ image = image.to(device=device, dtype=dtype)
891
+
892
+ noise_timestep = timesteps[0:1]
893
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
894
+
895
+ intermediate_images = self.prepare_intermediate_images(
896
+ image, noise_timestep, batch_size, num_images_per_prompt, dtype, device, generator
897
+ )
898
+
899
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
900
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
901
+
902
+ # HACK: see comment in `enable_model_cpu_offload`
903
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
904
+ self.text_encoder_offload_hook.offload()
905
+
906
+ # 7. Denoising loop
907
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
908
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
909
+ for i, t in enumerate(timesteps):
910
+ model_input = (
911
+ torch.cat([intermediate_images] * 2) if do_classifier_free_guidance else intermediate_images
912
+ )
913
+ model_input = self.scheduler.scale_model_input(model_input, t)
914
+
915
+ # predict the noise residual
916
+ noise_pred = self.unet(
917
+ model_input,
918
+ t,
919
+ encoder_hidden_states=prompt_embeds,
920
+ cross_attention_kwargs=cross_attention_kwargs,
921
+ ).sample
922
+
923
+ # perform guidance
924
+ if do_classifier_free_guidance:
925
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
926
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1], dim=1)
927
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1], dim=1)
928
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
929
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
930
+
931
+ # compute the previous noisy sample x_t -> x_t-1
932
+ intermediate_images = self.scheduler.step(
933
+ noise_pred, t, intermediate_images, **extra_step_kwargs
934
+ ).prev_sample
935
+
936
+ # call the callback, if provided
937
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
938
+ progress_bar.update()
939
+ if callback is not None and i % callback_steps == 0:
940
+ callback(i, t, intermediate_images)
941
+
942
+ image = intermediate_images
943
+
944
+ if output_type == "pil":
945
+ # 8. Post-processing
946
+ image = (image / 2 + 0.5).clamp(0, 1)
947
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
948
+
949
+ # 9. Run safety checker
950
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
951
+
952
+ # 10. Convert to PIL
953
+ image = self.numpy_to_pil(image)
954
+
955
+ # 11. Apply watermark
956
+ if self.watermarker is not None:
957
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
958
+ elif output_type == "pt":
959
+ nsfw_detected = None
960
+ watermark_detected = None
961
+
962
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
963
+ self.unet_offload_hook.offload()
964
+ else:
965
+ # 8. Post-processing
966
+ image = (image / 2 + 0.5).clamp(0, 1)
967
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
968
+
969
+ # 9. Run safety checker
970
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
971
+
972
+ # Offload last model to CPU
973
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
974
+ self.final_offload_hook.offload()
975
+
976
+ if not return_dict:
977
+ return (image, nsfw_detected, watermark_detected)
978
+
979
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)