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 +1,1098 @@
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
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.resize
41
+ def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
42
+ w, h = images.size
43
+
44
+ coef = w / h
45
+
46
+ w, h = img_size, img_size
47
+
48
+ if coef >= 1:
49
+ w = int(round(img_size / 8 * coef) * 8)
50
+ else:
51
+ h = int(round(img_size / 8 / coef) * 8)
52
+
53
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
54
+
55
+ return images
56
+
57
+
58
+ EXAMPLE_DOC_STRING = """
59
+ Examples:
60
+ ```py
61
+ >>> from diffusers import IFInpaintingPipeline, IFInpaintingSuperResolutionPipeline, DiffusionPipeline
62
+ >>> from diffusers.utils import pt_to_pil
63
+ >>> import torch
64
+ >>> from PIL import Image
65
+ >>> import requests
66
+ >>> from io import BytesIO
67
+
68
+ >>> url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/if/person.png"
69
+ >>> response = requests.get(url)
70
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
71
+ >>> original_image = original_image
72
+
73
+ >>> url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/if/glasses_mask.png"
74
+ >>> response = requests.get(url)
75
+ >>> mask_image = Image.open(BytesIO(response.content))
76
+ >>> mask_image = mask_image
77
+
78
+ >>> pipe = IFInpaintingPipeline.from_pretrained(
79
+ ... "DeepFloyd/IF-I-XL-v1.0", variant="fp16", torch_dtype=torch.float16
80
+ ... )
81
+ >>> pipe.enable_model_cpu_offload()
82
+
83
+ >>> prompt = "blue sunglasses"
84
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
85
+
86
+ >>> image = pipe(
87
+ ... image=original_image,
88
+ ... mask_image=mask_image,
89
+ ... prompt_embeds=prompt_embeds,
90
+ ... negative_prompt_embeds=negative_embeds,
91
+ ... output_type="pt",
92
+ ... ).images
93
+
94
+ >>> # save intermediate image
95
+ >>> pil_image = pt_to_pil(image)
96
+ >>> pil_image[0].save("./if_stage_I.png")
97
+
98
+ >>> super_res_1_pipe = IFInpaintingSuperResolutionPipeline.from_pretrained(
99
+ ... "DeepFloyd/IF-II-L-v1.0", text_encoder=None, variant="fp16", torch_dtype=torch.float16
100
+ ... )
101
+ >>> super_res_1_pipe.enable_model_cpu_offload()
102
+
103
+ >>> image = super_res_1_pipe(
104
+ ... image=image,
105
+ ... mask_image=mask_image,
106
+ ... original_image=original_image,
107
+ ... prompt_embeds=prompt_embeds,
108
+ ... negative_prompt_embeds=negative_embeds,
109
+ ... ).images
110
+ >>> image[0].save("./if_stage_II.png")
111
+ ```
112
+ """
113
+
114
+
115
+ class IFInpaintingPipeline(DiffusionPipeline):
116
+ tokenizer: T5Tokenizer
117
+ text_encoder: T5EncoderModel
118
+
119
+ unet: UNet2DConditionModel
120
+ scheduler: DDPMScheduler
121
+
122
+ feature_extractor: Optional[CLIPImageProcessor]
123
+ safety_checker: Optional[IFSafetyChecker]
124
+
125
+ watermarker: Optional[IFWatermarker]
126
+
127
+ bad_punct_regex = re.compile(
128
+ r"[" + "#®•©™&@·º½¾¿¡§~" + "\)" + "\(" + "\]" + "\[" + "\}" + "\{" + "\|" + "\\" + "\/" + "\*" + r"]{1,}"
129
+ ) # noqa
130
+
131
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
132
+
133
+ def __init__(
134
+ self,
135
+ tokenizer: T5Tokenizer,
136
+ text_encoder: T5EncoderModel,
137
+ unet: UNet2DConditionModel,
138
+ scheduler: DDPMScheduler,
139
+ safety_checker: Optional[IFSafetyChecker],
140
+ feature_extractor: Optional[CLIPImageProcessor],
141
+ watermarker: Optional[IFWatermarker],
142
+ requires_safety_checker: bool = True,
143
+ ):
144
+ super().__init__()
145
+
146
+ if safety_checker is None and requires_safety_checker:
147
+ logger.warning(
148
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
149
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
150
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
151
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
152
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
153
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
154
+ )
155
+
156
+ if safety_checker is not None and feature_extractor is None:
157
+ raise ValueError(
158
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
159
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
160
+ )
161
+
162
+ self.register_modules(
163
+ tokenizer=tokenizer,
164
+ text_encoder=text_encoder,
165
+ unet=unet,
166
+ scheduler=scheduler,
167
+ safety_checker=safety_checker,
168
+ feature_extractor=feature_extractor,
169
+ watermarker=watermarker,
170
+ )
171
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
172
+
173
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.enable_sequential_cpu_offload
174
+ def enable_sequential_cpu_offload(self, gpu_id=0):
175
+ r"""
176
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
177
+ models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
178
+ when their specific submodule has its `forward` method called.
179
+ """
180
+ if is_accelerate_available():
181
+ from accelerate import cpu_offload
182
+ else:
183
+ raise ImportError("Please install accelerate via `pip install accelerate`")
184
+
185
+ device = torch.device(f"cuda:{gpu_id}")
186
+
187
+ models = [
188
+ self.text_encoder,
189
+ self.unet,
190
+ ]
191
+ for cpu_offloaded_model in models:
192
+ if cpu_offloaded_model is not None:
193
+ cpu_offload(cpu_offloaded_model, device)
194
+
195
+ if self.safety_checker is not None:
196
+ cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
197
+
198
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.enable_model_cpu_offload
199
+ def enable_model_cpu_offload(self, gpu_id=0):
200
+ r"""
201
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
202
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
203
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
204
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
205
+ """
206
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
207
+ from accelerate import cpu_offload_with_hook
208
+ else:
209
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
210
+
211
+ device = torch.device(f"cuda:{gpu_id}")
212
+
213
+ if self.device.type != "cpu":
214
+ self.to("cpu", silence_dtype_warnings=True)
215
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
216
+
217
+ hook = None
218
+
219
+ if self.text_encoder is not None:
220
+ _, hook = cpu_offload_with_hook(self.text_encoder, device, prev_module_hook=hook)
221
+
222
+ # Accelerate will move the next model to the device _before_ calling the offload hook of the
223
+ # previous model. This will cause both models to be present on the device at the same time.
224
+ # IF uses T5 for its text encoder which is really large. We can manually call the offload
225
+ # hook for the text encoder to ensure it's moved to the cpu before the unet is moved to
226
+ # the GPU.
227
+ self.text_encoder_offload_hook = hook
228
+
229
+ _, hook = cpu_offload_with_hook(self.unet, device, prev_module_hook=hook)
230
+
231
+ # if the safety checker isn't called, `unet_offload_hook` will have to be called to manually offload the unet
232
+ self.unet_offload_hook = hook
233
+
234
+ if self.safety_checker is not None:
235
+ _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
236
+
237
+ # We'll offload the last model manually.
238
+ self.final_offload_hook = hook
239
+
240
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
241
+ def remove_all_hooks(self):
242
+ if is_accelerate_available():
243
+ from accelerate.hooks import remove_hook_from_module
244
+ else:
245
+ raise ImportError("Please install accelerate via `pip install accelerate`")
246
+
247
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
248
+ if model is not None:
249
+ remove_hook_from_module(model, recurse=True)
250
+
251
+ self.unet_offload_hook = None
252
+ self.text_encoder_offload_hook = None
253
+ self.final_offload_hook = None
254
+
255
+ @property
256
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
257
+ def _execution_device(self):
258
+ r"""
259
+ Returns the device on which the pipeline's models will be executed. After calling
260
+ `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
261
+ hooks.
262
+ """
263
+ if not hasattr(self.unet, "_hf_hook"):
264
+ return self.device
265
+ for module in self.unet.modules():
266
+ if (
267
+ hasattr(module, "_hf_hook")
268
+ and hasattr(module._hf_hook, "execution_device")
269
+ and module._hf_hook.execution_device is not None
270
+ ):
271
+ return torch.device(module._hf_hook.execution_device)
272
+ return self.device
273
+
274
+ @torch.no_grad()
275
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
276
+ def encode_prompt(
277
+ self,
278
+ prompt,
279
+ do_classifier_free_guidance=True,
280
+ num_images_per_prompt=1,
281
+ device=None,
282
+ negative_prompt=None,
283
+ prompt_embeds: Optional[torch.FloatTensor] = None,
284
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
285
+ clean_caption: bool = False,
286
+ ):
287
+ r"""
288
+ Encodes the prompt into text encoder hidden states.
289
+
290
+ Args:
291
+ prompt (`str` or `List[str]`, *optional*):
292
+ prompt to be encoded
293
+ device: (`torch.device`, *optional*):
294
+ torch device to place the resulting embeddings on
295
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
296
+ number of images that should be generated per prompt
297
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
298
+ whether to use classifier free guidance or not
299
+ negative_prompt (`str` or `List[str]`, *optional*):
300
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
301
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
302
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
303
+ prompt_embeds (`torch.FloatTensor`, *optional*):
304
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
305
+ provided, text embeddings will be generated from `prompt` input argument.
306
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
307
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
308
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
309
+ argument.
310
+ """
311
+ if prompt is not None and negative_prompt is not None:
312
+ if type(prompt) is not type(negative_prompt):
313
+ raise TypeError(
314
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
315
+ f" {type(prompt)}."
316
+ )
317
+
318
+ if device is None:
319
+ device = self._execution_device
320
+
321
+ if prompt is not None and isinstance(prompt, str):
322
+ batch_size = 1
323
+ elif prompt is not None and isinstance(prompt, list):
324
+ batch_size = len(prompt)
325
+ else:
326
+ batch_size = prompt_embeds.shape[0]
327
+
328
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
329
+ max_length = 77
330
+
331
+ if prompt_embeds is None:
332
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
333
+ text_inputs = self.tokenizer(
334
+ prompt,
335
+ padding="max_length",
336
+ max_length=max_length,
337
+ truncation=True,
338
+ add_special_tokens=True,
339
+ return_tensors="pt",
340
+ )
341
+ text_input_ids = text_inputs.input_ids
342
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
343
+
344
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
345
+ text_input_ids, untruncated_ids
346
+ ):
347
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
348
+ logger.warning(
349
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
350
+ f" {max_length} tokens: {removed_text}"
351
+ )
352
+
353
+ attention_mask = text_inputs.attention_mask.to(device)
354
+
355
+ prompt_embeds = self.text_encoder(
356
+ text_input_ids.to(device),
357
+ attention_mask=attention_mask,
358
+ )
359
+ prompt_embeds = prompt_embeds[0]
360
+
361
+ if self.text_encoder is not None:
362
+ dtype = self.text_encoder.dtype
363
+ elif self.unet is not None:
364
+ dtype = self.unet.dtype
365
+ else:
366
+ dtype = None
367
+
368
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
369
+
370
+ bs_embed, seq_len, _ = prompt_embeds.shape
371
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
372
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
373
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
374
+
375
+ # get unconditional embeddings for classifier free guidance
376
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
377
+ uncond_tokens: List[str]
378
+ if negative_prompt is None:
379
+ uncond_tokens = [""] * batch_size
380
+ elif isinstance(negative_prompt, str):
381
+ uncond_tokens = [negative_prompt]
382
+ elif batch_size != len(negative_prompt):
383
+ raise ValueError(
384
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
385
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
386
+ " the batch size of `prompt`."
387
+ )
388
+ else:
389
+ uncond_tokens = negative_prompt
390
+
391
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
392
+ max_length = prompt_embeds.shape[1]
393
+ uncond_input = self.tokenizer(
394
+ uncond_tokens,
395
+ padding="max_length",
396
+ max_length=max_length,
397
+ truncation=True,
398
+ return_attention_mask=True,
399
+ add_special_tokens=True,
400
+ return_tensors="pt",
401
+ )
402
+ attention_mask = uncond_input.attention_mask.to(device)
403
+
404
+ negative_prompt_embeds = self.text_encoder(
405
+ uncond_input.input_ids.to(device),
406
+ attention_mask=attention_mask,
407
+ )
408
+ negative_prompt_embeds = negative_prompt_embeds[0]
409
+
410
+ if do_classifier_free_guidance:
411
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
412
+ seq_len = negative_prompt_embeds.shape[1]
413
+
414
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
415
+
416
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
417
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
418
+
419
+ # For classifier free guidance, we need to do two forward passes.
420
+ # Here we concatenate the unconditional and text embeddings into a single batch
421
+ # to avoid doing two forward passes
422
+ else:
423
+ negative_prompt_embeds = None
424
+
425
+ return prompt_embeds, negative_prompt_embeds
426
+
427
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
428
+ def run_safety_checker(self, image, device, dtype):
429
+ if self.safety_checker is not None:
430
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
431
+ image, nsfw_detected, watermark_detected = self.safety_checker(
432
+ images=image,
433
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
434
+ )
435
+ else:
436
+ nsfw_detected = None
437
+ watermark_detected = None
438
+
439
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
440
+ self.unet_offload_hook.offload()
441
+
442
+ return image, nsfw_detected, watermark_detected
443
+
444
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
445
+ def prepare_extra_step_kwargs(self, generator, eta):
446
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
447
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
448
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
449
+ # and should be between [0, 1]
450
+
451
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
452
+ extra_step_kwargs = {}
453
+ if accepts_eta:
454
+ extra_step_kwargs["eta"] = eta
455
+
456
+ # check if the scheduler accepts generator
457
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
458
+ if accepts_generator:
459
+ extra_step_kwargs["generator"] = generator
460
+ return extra_step_kwargs
461
+
462
+ def check_inputs(
463
+ self,
464
+ prompt,
465
+ image,
466
+ mask_image,
467
+ batch_size,
468
+ callback_steps,
469
+ negative_prompt=None,
470
+ prompt_embeds=None,
471
+ negative_prompt_embeds=None,
472
+ ):
473
+ if (callback_steps is None) or (
474
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
475
+ ):
476
+ raise ValueError(
477
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
478
+ f" {type(callback_steps)}."
479
+ )
480
+
481
+ if prompt is not None and prompt_embeds is not None:
482
+ raise ValueError(
483
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
484
+ " only forward one of the two."
485
+ )
486
+ elif prompt is None and prompt_embeds is None:
487
+ raise ValueError(
488
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
489
+ )
490
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
491
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
492
+
493
+ if negative_prompt is not None and negative_prompt_embeds is not None:
494
+ raise ValueError(
495
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
496
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
497
+ )
498
+
499
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
500
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
501
+ raise ValueError(
502
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
503
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
504
+ f" {negative_prompt_embeds.shape}."
505
+ )
506
+
507
+ # image
508
+
509
+ if isinstance(image, list):
510
+ check_image_type = image[0]
511
+ else:
512
+ check_image_type = image
513
+
514
+ if (
515
+ not isinstance(check_image_type, torch.Tensor)
516
+ and not isinstance(check_image_type, PIL.Image.Image)
517
+ and not isinstance(check_image_type, np.ndarray)
518
+ ):
519
+ raise ValueError(
520
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
521
+ f" {type(check_image_type)}"
522
+ )
523
+
524
+ if isinstance(image, list):
525
+ image_batch_size = len(image)
526
+ elif isinstance(image, torch.Tensor):
527
+ image_batch_size = image.shape[0]
528
+ elif isinstance(image, PIL.Image.Image):
529
+ image_batch_size = 1
530
+ elif isinstance(image, np.ndarray):
531
+ image_batch_size = image.shape[0]
532
+ else:
533
+ assert False
534
+
535
+ if batch_size != image_batch_size:
536
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
537
+
538
+ # mask_image
539
+
540
+ if isinstance(mask_image, list):
541
+ check_image_type = mask_image[0]
542
+ else:
543
+ check_image_type = mask_image
544
+
545
+ if (
546
+ not isinstance(check_image_type, torch.Tensor)
547
+ and not isinstance(check_image_type, PIL.Image.Image)
548
+ and not isinstance(check_image_type, np.ndarray)
549
+ ):
550
+ raise ValueError(
551
+ "`mask_image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
552
+ f" {type(check_image_type)}"
553
+ )
554
+
555
+ if isinstance(mask_image, list):
556
+ image_batch_size = len(mask_image)
557
+ elif isinstance(mask_image, torch.Tensor):
558
+ image_batch_size = mask_image.shape[0]
559
+ elif isinstance(mask_image, PIL.Image.Image):
560
+ image_batch_size = 1
561
+ elif isinstance(mask_image, np.ndarray):
562
+ image_batch_size = mask_image.shape[0]
563
+ else:
564
+ assert False
565
+
566
+ if image_batch_size != 1 and batch_size != image_batch_size:
567
+ raise ValueError(
568
+ f"mask_image batch size: {image_batch_size} must be `1` or the same as prompt batch size {batch_size}"
569
+ )
570
+
571
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
572
+ def _text_preprocessing(self, text, clean_caption=False):
573
+ if clean_caption and not is_bs4_available():
574
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
575
+ logger.warn("Setting `clean_caption` to False...")
576
+ clean_caption = False
577
+
578
+ if clean_caption and not is_ftfy_available():
579
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
580
+ logger.warn("Setting `clean_caption` to False...")
581
+ clean_caption = False
582
+
583
+ if not isinstance(text, (tuple, list)):
584
+ text = [text]
585
+
586
+ def process(text: str):
587
+ if clean_caption:
588
+ text = self._clean_caption(text)
589
+ text = self._clean_caption(text)
590
+ else:
591
+ text = text.lower().strip()
592
+ return text
593
+
594
+ return [process(t) for t in text]
595
+
596
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
597
+ def _clean_caption(self, caption):
598
+ caption = str(caption)
599
+ caption = ul.unquote_plus(caption)
600
+ caption = caption.strip().lower()
601
+ caption = re.sub("<person>", "person", caption)
602
+ # urls:
603
+ caption = re.sub(
604
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
605
+ "",
606
+ caption,
607
+ ) # regex for urls
608
+ caption = re.sub(
609
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
610
+ "",
611
+ caption,
612
+ ) # regex for urls
613
+ # html:
614
+ caption = BeautifulSoup(caption, features="html.parser").text
615
+
616
+ # @<nickname>
617
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
618
+
619
+ # 31C0—31EF CJK Strokes
620
+ # 31F0—31FF Katakana Phonetic Extensions
621
+ # 3200—32FF Enclosed CJK Letters and Months
622
+ # 3300—33FF CJK Compatibility
623
+ # 3400—4DBF CJK Unified Ideographs Extension A
624
+ # 4DC0—4DFF Yijing Hexagram Symbols
625
+ # 4E00—9FFF CJK Unified Ideographs
626
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
627
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
628
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
629
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
630
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
631
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
632
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
633
+ #######################################################
634
+
635
+ # все виды тире / all types of dash --> "-"
636
+ caption = re.sub(
637
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
638
+ "-",
639
+ caption,
640
+ )
641
+
642
+ # кавычки к одному стандарту
643
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
644
+ caption = re.sub(r"[‘’]", "'", caption)
645
+
646
+ # &quot;
647
+ caption = re.sub(r"&quot;?", "", caption)
648
+ # &amp
649
+ caption = re.sub(r"&amp", "", caption)
650
+
651
+ # ip adresses:
652
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
653
+
654
+ # article ids:
655
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
656
+
657
+ # \n
658
+ caption = re.sub(r"\\n", " ", caption)
659
+
660
+ # "#123"
661
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
662
+ # "#12345.."
663
+ caption = re.sub(r"#\d{5,}\b", "", caption)
664
+ # "123456.."
665
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
666
+ # filenames:
667
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
668
+
669
+ #
670
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
671
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
672
+
673
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
674
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
675
+
676
+ # this-is-my-cute-cat / this_is_my_cute_cat
677
+ regex2 = re.compile(r"(?:\-|\_)")
678
+ if len(re.findall(regex2, caption)) > 3:
679
+ caption = re.sub(regex2, " ", caption)
680
+
681
+ caption = ftfy.fix_text(caption)
682
+ caption = html.unescape(html.unescape(caption))
683
+
684
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
685
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
686
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
687
+
688
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
689
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
690
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
691
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
692
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
693
+
694
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
695
+
696
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
697
+
698
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
699
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
700
+ caption = re.sub(r"\s+", " ", caption)
701
+
702
+ caption.strip()
703
+
704
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
705
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
706
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
707
+ caption = re.sub(r"^\.\S+$", "", caption)
708
+
709
+ return caption.strip()
710
+
711
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.preprocess_image
712
+ def preprocess_image(self, image: PIL.Image.Image) -> torch.Tensor:
713
+ if not isinstance(image, list):
714
+ image = [image]
715
+
716
+ def numpy_to_pt(images):
717
+ if images.ndim == 3:
718
+ images = images[..., None]
719
+
720
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
721
+ return images
722
+
723
+ if isinstance(image[0], PIL.Image.Image):
724
+ new_image = []
725
+
726
+ for image_ in image:
727
+ image_ = image_.convert("RGB")
728
+ image_ = resize(image_, self.unet.sample_size)
729
+ image_ = np.array(image_)
730
+ image_ = image_.astype(np.float32)
731
+ image_ = image_ / 127.5 - 1
732
+ new_image.append(image_)
733
+
734
+ image = new_image
735
+
736
+ image = np.stack(image, axis=0) # to np
737
+ image = numpy_to_pt(image) # to pt
738
+
739
+ elif isinstance(image[0], np.ndarray):
740
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
741
+ image = numpy_to_pt(image)
742
+
743
+ elif isinstance(image[0], torch.Tensor):
744
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
745
+
746
+ return image
747
+
748
+ def preprocess_mask_image(self, mask_image) -> torch.Tensor:
749
+ if not isinstance(mask_image, list):
750
+ mask_image = [mask_image]
751
+
752
+ if isinstance(mask_image[0], torch.Tensor):
753
+ mask_image = torch.cat(mask_image, axis=0) if mask_image[0].ndim == 4 else torch.stack(mask_image, axis=0)
754
+
755
+ if mask_image.ndim == 2:
756
+ # Batch and add channel dim for single mask
757
+ mask_image = mask_image.unsqueeze(0).unsqueeze(0)
758
+ elif mask_image.ndim == 3 and mask_image.shape[0] == 1:
759
+ # Single mask, the 0'th dimension is considered to be
760
+ # the existing batch size of 1
761
+ mask_image = mask_image.unsqueeze(0)
762
+ elif mask_image.ndim == 3 and mask_image.shape[0] != 1:
763
+ # Batch of mask, the 0'th dimension is considered to be
764
+ # the batching dimension
765
+ mask_image = mask_image.unsqueeze(1)
766
+
767
+ mask_image[mask_image < 0.5] = 0
768
+ mask_image[mask_image >= 0.5] = 1
769
+
770
+ elif isinstance(mask_image[0], PIL.Image.Image):
771
+ new_mask_image = []
772
+
773
+ for mask_image_ in mask_image:
774
+ mask_image_ = mask_image_.convert("L")
775
+ mask_image_ = resize(mask_image_, self.unet.sample_size)
776
+ mask_image_ = np.array(mask_image_)
777
+ mask_image_ = mask_image_[None, None, :]
778
+ new_mask_image.append(mask_image_)
779
+
780
+ mask_image = new_mask_image
781
+
782
+ mask_image = np.concatenate(mask_image, axis=0)
783
+ mask_image = mask_image.astype(np.float32) / 255.0
784
+ mask_image[mask_image < 0.5] = 0
785
+ mask_image[mask_image >= 0.5] = 1
786
+ mask_image = torch.from_numpy(mask_image)
787
+
788
+ elif isinstance(mask_image[0], np.ndarray):
789
+ mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0)
790
+
791
+ mask_image[mask_image < 0.5] = 0
792
+ mask_image[mask_image >= 0.5] = 1
793
+ mask_image = torch.from_numpy(mask_image)
794
+
795
+ return mask_image
796
+
797
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.get_timesteps
798
+ def get_timesteps(self, num_inference_steps, strength):
799
+ # get the original timestep using init_timestep
800
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
801
+
802
+ t_start = max(num_inference_steps - init_timestep, 0)
803
+ timesteps = self.scheduler.timesteps[t_start:]
804
+
805
+ return timesteps, num_inference_steps - t_start
806
+
807
+ def prepare_intermediate_images(
808
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, mask_image, generator=None
809
+ ):
810
+ image_batch_size, channels, height, width = image.shape
811
+
812
+ batch_size = batch_size * num_images_per_prompt
813
+
814
+ shape = (batch_size, channels, height, width)
815
+
816
+ if isinstance(generator, list) and len(generator) != batch_size:
817
+ raise ValueError(
818
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
819
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
820
+ )
821
+
822
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
823
+
824
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
825
+ noised_image = self.scheduler.add_noise(image, noise, timestep)
826
+
827
+ image = (1 - mask_image) * image + mask_image * noised_image
828
+
829
+ return image
830
+
831
+ @torch.no_grad()
832
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
833
+ def __call__(
834
+ self,
835
+ prompt: Union[str, List[str]] = None,
836
+ image: Union[
837
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
838
+ ] = None,
839
+ mask_image: Union[
840
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
841
+ ] = None,
842
+ strength: float = 1.0,
843
+ num_inference_steps: int = 50,
844
+ timesteps: List[int] = None,
845
+ guidance_scale: float = 7.0,
846
+ negative_prompt: Optional[Union[str, List[str]]] = None,
847
+ num_images_per_prompt: Optional[int] = 1,
848
+ eta: float = 0.0,
849
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
850
+ prompt_embeds: Optional[torch.FloatTensor] = None,
851
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
852
+ output_type: Optional[str] = "pil",
853
+ return_dict: bool = True,
854
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
855
+ callback_steps: int = 1,
856
+ clean_caption: bool = True,
857
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
858
+ ):
859
+ """
860
+ Function invoked when calling the pipeline for generation.
861
+
862
+ Args:
863
+ prompt (`str` or `List[str]`, *optional*):
864
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
865
+ instead.
866
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
867
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
868
+ process.
869
+ mask_image (`PIL.Image.Image`):
870
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
871
+ repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
872
+ to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
873
+ instead of 3, so the expected shape would be `(B, H, W, 1)`.
874
+ strength (`float`, *optional*, defaults to 0.8):
875
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
876
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
877
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
878
+ be maximum and the denoising process will run for the full number of iterations specified in
879
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
880
+ num_inference_steps (`int`, *optional*, defaults to 50):
881
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
882
+ expense of slower inference.
883
+ timesteps (`List[int]`, *optional*):
884
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
885
+ timesteps are used. Must be in descending order.
886
+ guidance_scale (`float`, *optional*, defaults to 7.5):
887
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
888
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
889
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
890
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
891
+ usually at the expense of lower image quality.
892
+ negative_prompt (`str` or `List[str]`, *optional*):
893
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
894
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
895
+ less than `1`).
896
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
897
+ The number of images to generate per prompt.
898
+ eta (`float`, *optional*, defaults to 0.0):
899
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
900
+ [`schedulers.DDIMScheduler`], will be ignored for others.
901
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
902
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
903
+ to make generation deterministic.
904
+ prompt_embeds (`torch.FloatTensor`, *optional*):
905
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
906
+ provided, text embeddings will be generated from `prompt` input argument.
907
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
908
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
909
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
910
+ argument.
911
+ output_type (`str`, *optional*, defaults to `"pil"`):
912
+ The output format of the generate image. Choose between
913
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
914
+ return_dict (`bool`, *optional*, defaults to `True`):
915
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
916
+ callback (`Callable`, *optional*):
917
+ A function that will be called every `callback_steps` steps during inference. The function will be
918
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
919
+ callback_steps (`int`, *optional*, defaults to 1):
920
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
921
+ called at every step.
922
+ clean_caption (`bool`, *optional*, defaults to `True`):
923
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
924
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
925
+ prompt.
926
+ cross_attention_kwargs (`dict`, *optional*):
927
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
928
+ `self.processor` in
929
+ [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
930
+
931
+ Examples:
932
+
933
+ Returns:
934
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
935
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
936
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
937
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
938
+ or watermarked content, according to the `safety_checker`.
939
+ """
940
+ # 1. Check inputs. Raise error if not correct
941
+ if prompt is not None and isinstance(prompt, str):
942
+ batch_size = 1
943
+ elif prompt is not None and isinstance(prompt, list):
944
+ batch_size = len(prompt)
945
+ else:
946
+ batch_size = prompt_embeds.shape[0]
947
+
948
+ self.check_inputs(
949
+ prompt,
950
+ image,
951
+ mask_image,
952
+ batch_size,
953
+ callback_steps,
954
+ negative_prompt,
955
+ prompt_embeds,
956
+ negative_prompt_embeds,
957
+ )
958
+
959
+ # 2. Define call parameters
960
+ device = self._execution_device
961
+
962
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
963
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
964
+ # corresponds to doing no classifier free guidance.
965
+ do_classifier_free_guidance = guidance_scale > 1.0
966
+
967
+ # 3. Encode input prompt
968
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
969
+ prompt,
970
+ do_classifier_free_guidance,
971
+ num_images_per_prompt=num_images_per_prompt,
972
+ device=device,
973
+ negative_prompt=negative_prompt,
974
+ prompt_embeds=prompt_embeds,
975
+ negative_prompt_embeds=negative_prompt_embeds,
976
+ clean_caption=clean_caption,
977
+ )
978
+
979
+ if do_classifier_free_guidance:
980
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
981
+
982
+ dtype = prompt_embeds.dtype
983
+
984
+ # 4. Prepare timesteps
985
+ if timesteps is not None:
986
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
987
+ timesteps = self.scheduler.timesteps
988
+ num_inference_steps = len(timesteps)
989
+ else:
990
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
991
+ timesteps = self.scheduler.timesteps
992
+
993
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
994
+
995
+ # 5. Prepare intermediate images
996
+ image = self.preprocess_image(image)
997
+ image = image.to(device=device, dtype=dtype)
998
+
999
+ mask_image = self.preprocess_mask_image(mask_image)
1000
+ mask_image = mask_image.to(device=device, dtype=dtype)
1001
+
1002
+ if mask_image.shape[0] == 1:
1003
+ mask_image = mask_image.repeat_interleave(batch_size * num_images_per_prompt, dim=0)
1004
+ else:
1005
+ mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0)
1006
+
1007
+ noise_timestep = timesteps[0:1]
1008
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
1009
+
1010
+ intermediate_images = self.prepare_intermediate_images(
1011
+ image, noise_timestep, batch_size, num_images_per_prompt, dtype, device, mask_image, generator
1012
+ )
1013
+
1014
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
1015
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
1016
+
1017
+ # HACK: see comment in `enable_model_cpu_offload`
1018
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
1019
+ self.text_encoder_offload_hook.offload()
1020
+
1021
+ # 7. Denoising loop
1022
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
1023
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1024
+ for i, t in enumerate(timesteps):
1025
+ model_input = (
1026
+ torch.cat([intermediate_images] * 2) if do_classifier_free_guidance else intermediate_images
1027
+ )
1028
+ model_input = self.scheduler.scale_model_input(model_input, t)
1029
+
1030
+ # predict the noise residual
1031
+ noise_pred = self.unet(
1032
+ model_input,
1033
+ t,
1034
+ encoder_hidden_states=prompt_embeds,
1035
+ cross_attention_kwargs=cross_attention_kwargs,
1036
+ ).sample
1037
+
1038
+ # perform guidance
1039
+ if do_classifier_free_guidance:
1040
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
1041
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1], dim=1)
1042
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1], dim=1)
1043
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1044
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
1045
+
1046
+ # compute the previous noisy sample x_t -> x_t-1
1047
+ prev_intermediate_images = intermediate_images
1048
+
1049
+ intermediate_images = self.scheduler.step(
1050
+ noise_pred, t, intermediate_images, **extra_step_kwargs
1051
+ ).prev_sample
1052
+
1053
+ intermediate_images = (1 - mask_image) * prev_intermediate_images + mask_image * intermediate_images
1054
+
1055
+ # call the callback, if provided
1056
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1057
+ progress_bar.update()
1058
+ if callback is not None and i % callback_steps == 0:
1059
+ callback(i, t, intermediate_images)
1060
+
1061
+ image = intermediate_images
1062
+
1063
+ if output_type == "pil":
1064
+ # 8. Post-processing
1065
+ image = (image / 2 + 0.5).clamp(0, 1)
1066
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
1067
+
1068
+ # 9. Run safety checker
1069
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
1070
+
1071
+ # 10. Convert to PIL
1072
+ image = self.numpy_to_pil(image)
1073
+
1074
+ # 11. Apply watermark
1075
+ if self.watermarker is not None:
1076
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
1077
+ elif output_type == "pt":
1078
+ nsfw_detected = None
1079
+ watermark_detected = None
1080
+
1081
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
1082
+ self.unet_offload_hook.offload()
1083
+ else:
1084
+ # 8. Post-processing
1085
+ image = (image / 2 + 0.5).clamp(0, 1)
1086
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
1087
+
1088
+ # 9. Run safety checker
1089
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
1090
+
1091
+ # Offload last model to CPU
1092
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
1093
+ self.final_offload_hook.offload()
1094
+
1095
+ if not return_dict:
1096
+ return (image, nsfw_detected, watermark_detected)
1097
+
1098
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)