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 +1,1097 @@
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
+ import torch.nn.functional as F
11
+ from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
12
+
13
+ from ...models import UNet2DConditionModel
14
+ from ...schedulers import DDPMScheduler
15
+ from ...utils import (
16
+ BACKENDS_MAPPING,
17
+ PIL_INTERPOLATION,
18
+ is_accelerate_available,
19
+ is_accelerate_version,
20
+ is_bs4_available,
21
+ is_ftfy_available,
22
+ logging,
23
+ randn_tensor,
24
+ replace_example_docstring,
25
+ )
26
+ from ..pipeline_utils import DiffusionPipeline
27
+ from . import IFPipelineOutput
28
+ from .safety_checker import IFSafetyChecker
29
+ from .watermark import IFWatermarker
30
+
31
+
32
+ if is_bs4_available():
33
+ from bs4 import BeautifulSoup
34
+
35
+ if is_ftfy_available():
36
+ import ftfy
37
+
38
+
39
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
40
+
41
+
42
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.resize
43
+ def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
44
+ w, h = images.size
45
+
46
+ coef = w / h
47
+
48
+ w, h = img_size, img_size
49
+
50
+ if coef >= 1:
51
+ w = int(round(img_size / 8 * coef) * 8)
52
+ else:
53
+ h = int(round(img_size / 8 / coef) * 8)
54
+
55
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
56
+
57
+ return images
58
+
59
+
60
+ EXAMPLE_DOC_STRING = """
61
+ Examples:
62
+ ```py
63
+ >>> from diffusers import IFImg2ImgPipeline, IFImg2ImgSuperResolutionPipeline, DiffusionPipeline
64
+ >>> from diffusers.utils import pt_to_pil
65
+ >>> import torch
66
+ >>> from PIL import Image
67
+ >>> import requests
68
+ >>> from io import BytesIO
69
+
70
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
71
+ >>> response = requests.get(url)
72
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
73
+ >>> original_image = original_image.resize((768, 512))
74
+
75
+ >>> pipe = IFImg2ImgPipeline.from_pretrained(
76
+ ... "DeepFloyd/IF-I-XL-v1.0",
77
+ ... variant="fp16",
78
+ ... torch_dtype=torch.float16,
79
+ ... )
80
+ >>> pipe.enable_model_cpu_offload()
81
+
82
+ >>> prompt = "A fantasy landscape in style minecraft"
83
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
84
+
85
+ >>> image = pipe(
86
+ ... image=original_image,
87
+ ... prompt_embeds=prompt_embeds,
88
+ ... negative_prompt_embeds=negative_embeds,
89
+ ... output_type="pt",
90
+ ... ).images
91
+
92
+ >>> # save intermediate image
93
+ >>> pil_image = pt_to_pil(image)
94
+ >>> pil_image[0].save("./if_stage_I.png")
95
+
96
+ >>> super_res_1_pipe = IFImg2ImgSuperResolutionPipeline.from_pretrained(
97
+ ... "DeepFloyd/IF-II-L-v1.0",
98
+ ... text_encoder=None,
99
+ ... variant="fp16",
100
+ ... torch_dtype=torch.float16,
101
+ ... )
102
+ >>> super_res_1_pipe.enable_model_cpu_offload()
103
+
104
+ >>> image = super_res_1_pipe(
105
+ ... image=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 IFImg2ImgSuperResolutionPipeline(DiffusionPipeline):
116
+ tokenizer: T5Tokenizer
117
+ text_encoder: T5EncoderModel
118
+
119
+ unet: UNet2DConditionModel
120
+ scheduler: DDPMScheduler
121
+ image_noising_scheduler: DDPMScheduler
122
+
123
+ feature_extractor: Optional[CLIPImageProcessor]
124
+ safety_checker: Optional[IFSafetyChecker]
125
+
126
+ watermarker: Optional[IFWatermarker]
127
+
128
+ bad_punct_regex = re.compile(
129
+ r"[" + "#®•©™&@·º½¾¿¡§~" + "\)" + "\(" + "\]" + "\[" + "\}" + "\{" + "\|" + "\\" + "\/" + "\*" + r"]{1,}"
130
+ ) # noqa
131
+
132
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor"]
133
+
134
+ def __init__(
135
+ self,
136
+ tokenizer: T5Tokenizer,
137
+ text_encoder: T5EncoderModel,
138
+ unet: UNet2DConditionModel,
139
+ scheduler: DDPMScheduler,
140
+ image_noising_scheduler: DDPMScheduler,
141
+ safety_checker: Optional[IFSafetyChecker],
142
+ feature_extractor: Optional[CLIPImageProcessor],
143
+ watermarker: Optional[IFWatermarker],
144
+ requires_safety_checker: bool = True,
145
+ ):
146
+ super().__init__()
147
+
148
+ if safety_checker is None and requires_safety_checker:
149
+ logger.warning(
150
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
151
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
152
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
153
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
154
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
155
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
156
+ )
157
+
158
+ if safety_checker is not None and feature_extractor is None:
159
+ raise ValueError(
160
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
161
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
162
+ )
163
+
164
+ if unet.config.in_channels != 6:
165
+ logger.warn(
166
+ "It seems like you have loaded a checkpoint that shall not be used for super resolution from {unet.config._name_or_path} as it accepts {unet.config.in_channels} input channels instead of 6. Please make sure to pass a super resolution checkpoint as the `'unet'`: IFSuperResolutionPipeline.from_pretrained(unet=super_resolution_unet, ...)`."
167
+ )
168
+
169
+ self.register_modules(
170
+ tokenizer=tokenizer,
171
+ text_encoder=text_encoder,
172
+ unet=unet,
173
+ scheduler=scheduler,
174
+ image_noising_scheduler=image_noising_scheduler,
175
+ safety_checker=safety_checker,
176
+ feature_extractor=feature_extractor,
177
+ watermarker=watermarker,
178
+ )
179
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
180
+
181
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.enable_sequential_cpu_offload
182
+ def enable_sequential_cpu_offload(self, gpu_id=0):
183
+ r"""
184
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
185
+ models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
186
+ when their specific submodule has its `forward` method called.
187
+ """
188
+ if is_accelerate_available():
189
+ from accelerate import cpu_offload
190
+ else:
191
+ raise ImportError("Please install accelerate via `pip install accelerate`")
192
+
193
+ device = torch.device(f"cuda:{gpu_id}")
194
+
195
+ models = [
196
+ self.text_encoder,
197
+ self.unet,
198
+ ]
199
+ for cpu_offloaded_model in models:
200
+ if cpu_offloaded_model is not None:
201
+ cpu_offload(cpu_offloaded_model, device)
202
+
203
+ if self.safety_checker is not None:
204
+ cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True)
205
+
206
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.enable_model_cpu_offload
207
+ def enable_model_cpu_offload(self, gpu_id=0):
208
+ r"""
209
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
210
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
211
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
212
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
213
+ """
214
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
215
+ from accelerate import cpu_offload_with_hook
216
+ else:
217
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
218
+
219
+ device = torch.device(f"cuda:{gpu_id}")
220
+
221
+ if self.device.type != "cpu":
222
+ self.to("cpu", silence_dtype_warnings=True)
223
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
224
+
225
+ hook = None
226
+
227
+ if self.text_encoder is not None:
228
+ _, hook = cpu_offload_with_hook(self.text_encoder, device, prev_module_hook=hook)
229
+
230
+ # Accelerate will move the next model to the device _before_ calling the offload hook of the
231
+ # previous model. This will cause both models to be present on the device at the same time.
232
+ # IF uses T5 for its text encoder which is really large. We can manually call the offload
233
+ # hook for the text encoder to ensure it's moved to the cpu before the unet is moved to
234
+ # the GPU.
235
+ self.text_encoder_offload_hook = hook
236
+
237
+ _, hook = cpu_offload_with_hook(self.unet, device, prev_module_hook=hook)
238
+
239
+ # if the safety checker isn't called, `unet_offload_hook` will have to be called to manually offload the unet
240
+ self.unet_offload_hook = hook
241
+
242
+ if self.safety_checker is not None:
243
+ _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
244
+
245
+ # We'll offload the last model manually.
246
+ self.final_offload_hook = hook
247
+
248
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
249
+ def remove_all_hooks(self):
250
+ if is_accelerate_available():
251
+ from accelerate.hooks import remove_hook_from_module
252
+ else:
253
+ raise ImportError("Please install accelerate via `pip install accelerate`")
254
+
255
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
256
+ if model is not None:
257
+ remove_hook_from_module(model, recurse=True)
258
+
259
+ self.unet_offload_hook = None
260
+ self.text_encoder_offload_hook = None
261
+ self.final_offload_hook = None
262
+
263
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
264
+ def _text_preprocessing(self, text, clean_caption=False):
265
+ if clean_caption and not is_bs4_available():
266
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
267
+ logger.warn("Setting `clean_caption` to False...")
268
+ clean_caption = False
269
+
270
+ if clean_caption and not is_ftfy_available():
271
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
272
+ logger.warn("Setting `clean_caption` to False...")
273
+ clean_caption = False
274
+
275
+ if not isinstance(text, (tuple, list)):
276
+ text = [text]
277
+
278
+ def process(text: str):
279
+ if clean_caption:
280
+ text = self._clean_caption(text)
281
+ text = self._clean_caption(text)
282
+ else:
283
+ text = text.lower().strip()
284
+ return text
285
+
286
+ return [process(t) for t in text]
287
+
288
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
289
+ def _clean_caption(self, caption):
290
+ caption = str(caption)
291
+ caption = ul.unquote_plus(caption)
292
+ caption = caption.strip().lower()
293
+ caption = re.sub("<person>", "person", caption)
294
+ # urls:
295
+ caption = re.sub(
296
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
297
+ "",
298
+ caption,
299
+ ) # regex for urls
300
+ caption = re.sub(
301
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
302
+ "",
303
+ caption,
304
+ ) # regex for urls
305
+ # html:
306
+ caption = BeautifulSoup(caption, features="html.parser").text
307
+
308
+ # @<nickname>
309
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
310
+
311
+ # 31C0—31EF CJK Strokes
312
+ # 31F0—31FF Katakana Phonetic Extensions
313
+ # 3200—32FF Enclosed CJK Letters and Months
314
+ # 3300—33FF CJK Compatibility
315
+ # 3400—4DBF CJK Unified Ideographs Extension A
316
+ # 4DC0—4DFF Yijing Hexagram Symbols
317
+ # 4E00—9FFF CJK Unified Ideographs
318
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
319
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
320
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
321
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
322
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
323
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
324
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
325
+ #######################################################
326
+
327
+ # все виды тире / all types of dash --> "-"
328
+ caption = re.sub(
329
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
330
+ "-",
331
+ caption,
332
+ )
333
+
334
+ # кавычки к одному стандарту
335
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
336
+ caption = re.sub(r"[‘’]", "'", caption)
337
+
338
+ # &quot;
339
+ caption = re.sub(r"&quot;?", "", caption)
340
+ # &amp
341
+ caption = re.sub(r"&amp", "", caption)
342
+
343
+ # ip adresses:
344
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
345
+
346
+ # article ids:
347
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
348
+
349
+ # \n
350
+ caption = re.sub(r"\\n", " ", caption)
351
+
352
+ # "#123"
353
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
354
+ # "#12345.."
355
+ caption = re.sub(r"#\d{5,}\b", "", caption)
356
+ # "123456.."
357
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
358
+ # filenames:
359
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
360
+
361
+ #
362
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
363
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
364
+
365
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
366
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
367
+
368
+ # this-is-my-cute-cat / this_is_my_cute_cat
369
+ regex2 = re.compile(r"(?:\-|\_)")
370
+ if len(re.findall(regex2, caption)) > 3:
371
+ caption = re.sub(regex2, " ", caption)
372
+
373
+ caption = ftfy.fix_text(caption)
374
+ caption = html.unescape(html.unescape(caption))
375
+
376
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
377
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
378
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
379
+
380
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
381
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
382
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
383
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
384
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
385
+
386
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
387
+
388
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
389
+
390
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
391
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
392
+ caption = re.sub(r"\s+", " ", caption)
393
+
394
+ caption.strip()
395
+
396
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
397
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
398
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
399
+ caption = re.sub(r"^\.\S+$", "", caption)
400
+
401
+ return caption.strip()
402
+
403
+ @property
404
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
405
+ def _execution_device(self):
406
+ r"""
407
+ Returns the device on which the pipeline's models will be executed. After calling
408
+ `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
409
+ hooks.
410
+ """
411
+ if not hasattr(self.unet, "_hf_hook"):
412
+ return self.device
413
+ for module in self.unet.modules():
414
+ if (
415
+ hasattr(module, "_hf_hook")
416
+ and hasattr(module._hf_hook, "execution_device")
417
+ and module._hf_hook.execution_device is not None
418
+ ):
419
+ return torch.device(module._hf_hook.execution_device)
420
+ return self.device
421
+
422
+ @torch.no_grad()
423
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
424
+ def encode_prompt(
425
+ self,
426
+ prompt,
427
+ do_classifier_free_guidance=True,
428
+ num_images_per_prompt=1,
429
+ device=None,
430
+ negative_prompt=None,
431
+ prompt_embeds: Optional[torch.FloatTensor] = None,
432
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
433
+ clean_caption: bool = False,
434
+ ):
435
+ r"""
436
+ Encodes the prompt into text encoder hidden states.
437
+
438
+ Args:
439
+ prompt (`str` or `List[str]`, *optional*):
440
+ prompt to be encoded
441
+ device: (`torch.device`, *optional*):
442
+ torch device to place the resulting embeddings on
443
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
444
+ number of images that should be generated per prompt
445
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
446
+ whether to use classifier free guidance or not
447
+ negative_prompt (`str` or `List[str]`, *optional*):
448
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
449
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
450
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
451
+ prompt_embeds (`torch.FloatTensor`, *optional*):
452
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
453
+ provided, text embeddings will be generated from `prompt` input argument.
454
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
455
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
456
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
457
+ argument.
458
+ """
459
+ if prompt is not None and negative_prompt is not None:
460
+ if type(prompt) is not type(negative_prompt):
461
+ raise TypeError(
462
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
463
+ f" {type(prompt)}."
464
+ )
465
+
466
+ if device is None:
467
+ device = self._execution_device
468
+
469
+ if prompt is not None and isinstance(prompt, str):
470
+ batch_size = 1
471
+ elif prompt is not None and isinstance(prompt, list):
472
+ batch_size = len(prompt)
473
+ else:
474
+ batch_size = prompt_embeds.shape[0]
475
+
476
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
477
+ max_length = 77
478
+
479
+ if prompt_embeds is None:
480
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
481
+ text_inputs = self.tokenizer(
482
+ prompt,
483
+ padding="max_length",
484
+ max_length=max_length,
485
+ truncation=True,
486
+ add_special_tokens=True,
487
+ return_tensors="pt",
488
+ )
489
+ text_input_ids = text_inputs.input_ids
490
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
491
+
492
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
493
+ text_input_ids, untruncated_ids
494
+ ):
495
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
496
+ logger.warning(
497
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
498
+ f" {max_length} tokens: {removed_text}"
499
+ )
500
+
501
+ attention_mask = text_inputs.attention_mask.to(device)
502
+
503
+ prompt_embeds = self.text_encoder(
504
+ text_input_ids.to(device),
505
+ attention_mask=attention_mask,
506
+ )
507
+ prompt_embeds = prompt_embeds[0]
508
+
509
+ if self.text_encoder is not None:
510
+ dtype = self.text_encoder.dtype
511
+ elif self.unet is not None:
512
+ dtype = self.unet.dtype
513
+ else:
514
+ dtype = None
515
+
516
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
517
+
518
+ bs_embed, seq_len, _ = prompt_embeds.shape
519
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
520
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
521
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
522
+
523
+ # get unconditional embeddings for classifier free guidance
524
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
525
+ uncond_tokens: List[str]
526
+ if negative_prompt is None:
527
+ uncond_tokens = [""] * batch_size
528
+ elif isinstance(negative_prompt, str):
529
+ uncond_tokens = [negative_prompt]
530
+ elif batch_size != len(negative_prompt):
531
+ raise ValueError(
532
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
533
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
534
+ " the batch size of `prompt`."
535
+ )
536
+ else:
537
+ uncond_tokens = negative_prompt
538
+
539
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
540
+ max_length = prompt_embeds.shape[1]
541
+ uncond_input = self.tokenizer(
542
+ uncond_tokens,
543
+ padding="max_length",
544
+ max_length=max_length,
545
+ truncation=True,
546
+ return_attention_mask=True,
547
+ add_special_tokens=True,
548
+ return_tensors="pt",
549
+ )
550
+ attention_mask = uncond_input.attention_mask.to(device)
551
+
552
+ negative_prompt_embeds = self.text_encoder(
553
+ uncond_input.input_ids.to(device),
554
+ attention_mask=attention_mask,
555
+ )
556
+ negative_prompt_embeds = negative_prompt_embeds[0]
557
+
558
+ if do_classifier_free_guidance:
559
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
560
+ seq_len = negative_prompt_embeds.shape[1]
561
+
562
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
563
+
564
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
565
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
566
+
567
+ # For classifier free guidance, we need to do two forward passes.
568
+ # Here we concatenate the unconditional and text embeddings into a single batch
569
+ # to avoid doing two forward passes
570
+ else:
571
+ negative_prompt_embeds = None
572
+
573
+ return prompt_embeds, negative_prompt_embeds
574
+
575
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
576
+ def run_safety_checker(self, image, device, dtype):
577
+ if self.safety_checker is not None:
578
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
579
+ image, nsfw_detected, watermark_detected = self.safety_checker(
580
+ images=image,
581
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
582
+ )
583
+ else:
584
+ nsfw_detected = None
585
+ watermark_detected = None
586
+
587
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
588
+ self.unet_offload_hook.offload()
589
+
590
+ return image, nsfw_detected, watermark_detected
591
+
592
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
593
+ def prepare_extra_step_kwargs(self, generator, eta):
594
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
595
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
596
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
597
+ # and should be between [0, 1]
598
+
599
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
600
+ extra_step_kwargs = {}
601
+ if accepts_eta:
602
+ extra_step_kwargs["eta"] = eta
603
+
604
+ # check if the scheduler accepts generator
605
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
606
+ if accepts_generator:
607
+ extra_step_kwargs["generator"] = generator
608
+ return extra_step_kwargs
609
+
610
+ def check_inputs(
611
+ self,
612
+ prompt,
613
+ image,
614
+ original_image,
615
+ batch_size,
616
+ callback_steps,
617
+ negative_prompt=None,
618
+ prompt_embeds=None,
619
+ negative_prompt_embeds=None,
620
+ ):
621
+ if (callback_steps is None) or (
622
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
623
+ ):
624
+ raise ValueError(
625
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
626
+ f" {type(callback_steps)}."
627
+ )
628
+
629
+ if prompt is not None and prompt_embeds is not None:
630
+ raise ValueError(
631
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
632
+ " only forward one of the two."
633
+ )
634
+ elif prompt is None and prompt_embeds is None:
635
+ raise ValueError(
636
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
637
+ )
638
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
639
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
640
+
641
+ if negative_prompt is not None and negative_prompt_embeds is not None:
642
+ raise ValueError(
643
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
644
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
645
+ )
646
+
647
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
648
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
649
+ raise ValueError(
650
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
651
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
652
+ f" {negative_prompt_embeds.shape}."
653
+ )
654
+
655
+ # image
656
+
657
+ if isinstance(image, list):
658
+ check_image_type = image[0]
659
+ else:
660
+ check_image_type = image
661
+
662
+ if (
663
+ not isinstance(check_image_type, torch.Tensor)
664
+ and not isinstance(check_image_type, PIL.Image.Image)
665
+ and not isinstance(check_image_type, np.ndarray)
666
+ ):
667
+ raise ValueError(
668
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
669
+ f" {type(check_image_type)}"
670
+ )
671
+
672
+ if isinstance(image, list):
673
+ image_batch_size = len(image)
674
+ elif isinstance(image, torch.Tensor):
675
+ image_batch_size = image.shape[0]
676
+ elif isinstance(image, PIL.Image.Image):
677
+ image_batch_size = 1
678
+ elif isinstance(image, np.ndarray):
679
+ image_batch_size = image.shape[0]
680
+ else:
681
+ assert False
682
+
683
+ if batch_size != image_batch_size:
684
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
685
+
686
+ # original_image
687
+
688
+ if isinstance(original_image, list):
689
+ check_image_type = original_image[0]
690
+ else:
691
+ check_image_type = original_image
692
+
693
+ if (
694
+ not isinstance(check_image_type, torch.Tensor)
695
+ and not isinstance(check_image_type, PIL.Image.Image)
696
+ and not isinstance(check_image_type, np.ndarray)
697
+ ):
698
+ raise ValueError(
699
+ "`original_image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
700
+ f" {type(check_image_type)}"
701
+ )
702
+
703
+ if isinstance(original_image, list):
704
+ image_batch_size = len(original_image)
705
+ elif isinstance(original_image, torch.Tensor):
706
+ image_batch_size = original_image.shape[0]
707
+ elif isinstance(original_image, PIL.Image.Image):
708
+ image_batch_size = 1
709
+ elif isinstance(original_image, np.ndarray):
710
+ image_batch_size = original_image.shape[0]
711
+ else:
712
+ assert False
713
+
714
+ if batch_size != image_batch_size:
715
+ raise ValueError(
716
+ f"original_image batch size: {image_batch_size} must be same as prompt batch size {batch_size}"
717
+ )
718
+
719
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.preprocess_image with preprocess_image -> preprocess_original_image
720
+ def preprocess_original_image(self, image: PIL.Image.Image) -> torch.Tensor:
721
+ if not isinstance(image, list):
722
+ image = [image]
723
+
724
+ def numpy_to_pt(images):
725
+ if images.ndim == 3:
726
+ images = images[..., None]
727
+
728
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
729
+ return images
730
+
731
+ if isinstance(image[0], PIL.Image.Image):
732
+ new_image = []
733
+
734
+ for image_ in image:
735
+ image_ = image_.convert("RGB")
736
+ image_ = resize(image_, self.unet.sample_size)
737
+ image_ = np.array(image_)
738
+ image_ = image_.astype(np.float32)
739
+ image_ = image_ / 127.5 - 1
740
+ new_image.append(image_)
741
+
742
+ image = new_image
743
+
744
+ image = np.stack(image, axis=0) # to np
745
+ image = numpy_to_pt(image) # to pt
746
+
747
+ elif isinstance(image[0], np.ndarray):
748
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
749
+ image = numpy_to_pt(image)
750
+
751
+ elif isinstance(image[0], torch.Tensor):
752
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
753
+
754
+ return image
755
+
756
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_superresolution.IFSuperResolutionPipeline.preprocess_image
757
+ def preprocess_image(self, image: PIL.Image.Image, num_images_per_prompt, device) -> torch.Tensor:
758
+ if not isinstance(image, torch.Tensor) and not isinstance(image, list):
759
+ image = [image]
760
+
761
+ if isinstance(image[0], PIL.Image.Image):
762
+ image = [np.array(i).astype(np.float32) / 255.0 for i in image]
763
+
764
+ image = np.stack(image, axis=0) # to np
765
+ torch.from_numpy(image.transpose(0, 3, 1, 2))
766
+ elif isinstance(image[0], np.ndarray):
767
+ image = np.stack(image, axis=0) # to np
768
+ if image.ndim == 5:
769
+ image = image[0]
770
+
771
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
772
+ elif isinstance(image, list) and isinstance(image[0], torch.Tensor):
773
+ dims = image[0].ndim
774
+
775
+ if dims == 3:
776
+ image = torch.stack(image, dim=0)
777
+ elif dims == 4:
778
+ image = torch.concat(image, dim=0)
779
+ else:
780
+ raise ValueError(f"Image must have 3 or 4 dimensions, instead got {dims}")
781
+
782
+ image = image.to(device=device, dtype=self.unet.dtype)
783
+
784
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
785
+
786
+ return image
787
+
788
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.get_timesteps
789
+ def get_timesteps(self, num_inference_steps, strength):
790
+ # get the original timestep using init_timestep
791
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
792
+
793
+ t_start = max(num_inference_steps - init_timestep, 0)
794
+ timesteps = self.scheduler.timesteps[t_start:]
795
+
796
+ return timesteps, num_inference_steps - t_start
797
+
798
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.prepare_intermediate_images
799
+ def prepare_intermediate_images(
800
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None
801
+ ):
802
+ _, channels, height, width = image.shape
803
+
804
+ batch_size = batch_size * num_images_per_prompt
805
+
806
+ shape = (batch_size, channels, height, width)
807
+
808
+ if isinstance(generator, list) and len(generator) != batch_size:
809
+ raise ValueError(
810
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
811
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
812
+ )
813
+
814
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
815
+
816
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
817
+ image = self.scheduler.add_noise(image, noise, timestep)
818
+
819
+ return image
820
+
821
+ @torch.no_grad()
822
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
823
+ def __call__(
824
+ self,
825
+ image: Union[PIL.Image.Image, np.ndarray, torch.FloatTensor],
826
+ original_image: Union[
827
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
828
+ ] = None,
829
+ strength: float = 0.8,
830
+ prompt: Union[str, List[str]] = None,
831
+ num_inference_steps: int = 50,
832
+ timesteps: List[int] = None,
833
+ guidance_scale: float = 4.0,
834
+ negative_prompt: Optional[Union[str, List[str]]] = None,
835
+ num_images_per_prompt: Optional[int] = 1,
836
+ eta: float = 0.0,
837
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
838
+ prompt_embeds: Optional[torch.FloatTensor] = None,
839
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
840
+ output_type: Optional[str] = "pil",
841
+ return_dict: bool = True,
842
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
843
+ callback_steps: int = 1,
844
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
845
+ noise_level: int = 250,
846
+ clean_caption: bool = True,
847
+ ):
848
+ """
849
+ Function invoked when calling the pipeline for generation.
850
+
851
+ Args:
852
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
853
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
854
+ process.
855
+ original_image (`torch.FloatTensor` or `PIL.Image.Image`):
856
+ The original image that `image` was varied from.
857
+ strength (`float`, *optional*, defaults to 0.8):
858
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
859
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
860
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
861
+ be maximum and the denoising process will run for the full number of iterations specified in
862
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
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
+ num_inference_steps (`int`, *optional*, defaults to 50):
867
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
868
+ expense of slower inference.
869
+ timesteps (`List[int]`, *optional*):
870
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
871
+ timesteps are used. Must be in descending order.
872
+ guidance_scale (`float`, *optional*, defaults to 7.5):
873
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
874
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
875
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
876
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
877
+ usually at the expense of lower image quality.
878
+ negative_prompt (`str` or `List[str]`, *optional*):
879
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
880
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
881
+ less than `1`).
882
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
883
+ The number of images to generate per prompt.
884
+ eta (`float`, *optional*, defaults to 0.0):
885
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
886
+ [`schedulers.DDIMScheduler`], will be ignored for others.
887
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
888
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
889
+ to make generation deterministic.
890
+ prompt_embeds (`torch.FloatTensor`, *optional*):
891
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
892
+ provided, text embeddings will be generated from `prompt` input argument.
893
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
894
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
895
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
896
+ argument.
897
+ output_type (`str`, *optional*, defaults to `"pil"`):
898
+ The output format of the generate image. Choose between
899
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
900
+ return_dict (`bool`, *optional*, defaults to `True`):
901
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
902
+ callback (`Callable`, *optional*):
903
+ A function that will be called every `callback_steps` steps during inference. The function will be
904
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
905
+ callback_steps (`int`, *optional*, defaults to 1):
906
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
907
+ called at every step.
908
+ cross_attention_kwargs (`dict`, *optional*):
909
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
910
+ `self.processor` in
911
+ [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
912
+ noise_level (`int`, *optional*, defaults to 250):
913
+ The amount of noise to add to the upscaled image. Must be in the range `[0, 1000)`
914
+ clean_caption (`bool`, *optional*, defaults to `True`):
915
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
916
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
917
+ prompt.
918
+
919
+ Examples:
920
+
921
+ Returns:
922
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
923
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
924
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
925
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
926
+ or watermarked content, according to the `safety_checker`.
927
+ """
928
+ # 1. Check inputs. Raise error if not correct
929
+ if prompt is not None and isinstance(prompt, str):
930
+ batch_size = 1
931
+ elif prompt is not None and isinstance(prompt, list):
932
+ batch_size = len(prompt)
933
+ else:
934
+ batch_size = prompt_embeds.shape[0]
935
+
936
+ self.check_inputs(
937
+ prompt,
938
+ image,
939
+ original_image,
940
+ batch_size,
941
+ callback_steps,
942
+ negative_prompt,
943
+ prompt_embeds,
944
+ negative_prompt_embeds,
945
+ )
946
+
947
+ # 2. Define call parameters
948
+
949
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
950
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
951
+ # corresponds to doing no classifier free guidance.
952
+ do_classifier_free_guidance = guidance_scale > 1.0
953
+
954
+ device = self._execution_device
955
+
956
+ # 3. Encode input prompt
957
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
958
+ prompt,
959
+ do_classifier_free_guidance,
960
+ num_images_per_prompt=num_images_per_prompt,
961
+ device=device,
962
+ negative_prompt=negative_prompt,
963
+ prompt_embeds=prompt_embeds,
964
+ negative_prompt_embeds=negative_prompt_embeds,
965
+ clean_caption=clean_caption,
966
+ )
967
+
968
+ if do_classifier_free_guidance:
969
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
970
+
971
+ dtype = prompt_embeds.dtype
972
+
973
+ # 4. Prepare timesteps
974
+ if timesteps is not None:
975
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
976
+ timesteps = self.scheduler.timesteps
977
+ num_inference_steps = len(timesteps)
978
+ else:
979
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
980
+ timesteps = self.scheduler.timesteps
981
+
982
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
983
+
984
+ # 5. prepare original image
985
+ original_image = self.preprocess_original_image(original_image)
986
+ original_image = original_image.to(device=device, dtype=dtype)
987
+
988
+ # 6. Prepare intermediate images
989
+ noise_timestep = timesteps[0:1]
990
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
991
+
992
+ intermediate_images = self.prepare_intermediate_images(
993
+ original_image,
994
+ noise_timestep,
995
+ batch_size,
996
+ num_images_per_prompt,
997
+ dtype,
998
+ device,
999
+ generator,
1000
+ )
1001
+
1002
+ # 7. Prepare upscaled image and noise level
1003
+ _, _, height, width = original_image.shape
1004
+
1005
+ image = self.preprocess_image(image, num_images_per_prompt, device)
1006
+
1007
+ upscaled = F.interpolate(image, (height, width), mode="bilinear", align_corners=True)
1008
+
1009
+ noise_level = torch.tensor([noise_level] * upscaled.shape[0], device=upscaled.device)
1010
+ noise = randn_tensor(upscaled.shape, generator=generator, device=upscaled.device, dtype=upscaled.dtype)
1011
+ upscaled = self.image_noising_scheduler.add_noise(upscaled, noise, timesteps=noise_level)
1012
+
1013
+ if do_classifier_free_guidance:
1014
+ noise_level = torch.cat([noise_level] * 2)
1015
+
1016
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
1017
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
1018
+
1019
+ # HACK: see comment in `enable_model_cpu_offload`
1020
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
1021
+ self.text_encoder_offload_hook.offload()
1022
+
1023
+ # 9. Denoising loop
1024
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
1025
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
1026
+ for i, t in enumerate(timesteps):
1027
+ model_input = torch.cat([intermediate_images, upscaled], dim=1)
1028
+
1029
+ model_input = torch.cat([model_input] * 2) if do_classifier_free_guidance else model_input
1030
+ model_input = self.scheduler.scale_model_input(model_input, t)
1031
+
1032
+ # predict the noise residual
1033
+ noise_pred = self.unet(
1034
+ model_input,
1035
+ t,
1036
+ encoder_hidden_states=prompt_embeds,
1037
+ class_labels=noise_level,
1038
+ cross_attention_kwargs=cross_attention_kwargs,
1039
+ ).sample
1040
+
1041
+ # perform guidance
1042
+ if do_classifier_free_guidance:
1043
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
1044
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1] // 2, dim=1)
1045
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1] // 2, dim=1)
1046
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
1047
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
1048
+
1049
+ # compute the previous noisy sample x_t -> x_t-1
1050
+ intermediate_images = self.scheduler.step(
1051
+ noise_pred, t, intermediate_images, **extra_step_kwargs
1052
+ ).prev_sample
1053
+
1054
+ # call the callback, if provided
1055
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
1056
+ progress_bar.update()
1057
+ if callback is not None and i % callback_steps == 0:
1058
+ callback(i, t, intermediate_images)
1059
+
1060
+ image = intermediate_images
1061
+
1062
+ if output_type == "pil":
1063
+ # 10. Post-processing
1064
+ image = (image / 2 + 0.5).clamp(0, 1)
1065
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
1066
+
1067
+ # 11. Run safety checker
1068
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
1069
+
1070
+ # 12. Convert to PIL
1071
+ image = self.numpy_to_pil(image)
1072
+
1073
+ # 13. Apply watermark
1074
+ if self.watermarker is not None:
1075
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
1076
+ elif output_type == "pt":
1077
+ nsfw_detected = None
1078
+ watermark_detected = None
1079
+
1080
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
1081
+ self.unet_offload_hook.offload()
1082
+ else:
1083
+ # 10. Post-processing
1084
+ image = (image / 2 + 0.5).clamp(0, 1)
1085
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
1086
+
1087
+ # 11. Run safety checker
1088
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
1089
+
1090
+ # Offload last model to CPU
1091
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
1092
+ self.final_offload_hook.offload()
1093
+
1094
+ if not return_dict:
1095
+ return (image, nsfw_detected, watermark_detected)
1096
+
1097
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)