Datasets:

ArXiv:
diffusers-benchmarking-bot commited on
Commit
6fb704d
·
verified ·
1 Parent(s): 36ac85b

Upload folder using huggingface_hub

Browse files
main/README.md CHANGED
@@ -4336,19 +4336,19 @@ The Abstract of the paper:
4336
 
4337
  **64x64**
4338
  :-------------------------:
4339
- | <img src="https://github.com/user-attachments/assets/9e7bb2cd-45a0-4bd1-adb8-23e283baed39" width="222" height="222" alt="bird_64"> |
4340
 
4341
  - `256×256, nesting_level=1`: 1.776 GiB. With `150` DDIM inference steps:
4342
 
4343
  **64x64** | **256x256**
4344
  :-------------------------:|:-------------------------:
4345
- | <img src="https://github.com/user-attachments/assets/6b724c2e-5e6a-4b63-9b65-c1182cbb67e0" width="222" height="222" alt="64x64"> | <img src="https://github.com/user-attachments/assets/7dbab2ad-bf40-4a73-ab04-f178347cb7d5" width="222" height="222" alt="256x256"> |
4346
 
4347
- - `1024×1024, nesting_level=2`: 1.792 GiB. As one can realize the cost of adding another layer is really negligible. With `250` DDIM inference steps:
4348
 
4349
  **64x64** | **256x256** | **1024x1024**
4350
  :-------------------------:|:-------------------------:|:-------------------------:
4351
- | <img src="https://github.com/user-attachments/assets/4a9454e4-e20a-4736-a196-270e2ae796c0" width="222" height="222" alt="64x64"> | <img src="https://github.com/user-attachments/assets/4a96555d-0fda-4303-82b1-a4d886f770b9" width="222" height="222" alt="256x256"> | <img src="https://github.com/user-attachments/assets/e0239b7a-ab73-4d45-8f3e-b4e6b4b50abe" width="222" height="222" alt="1024x1024"> |
4352
 
4353
  ```py
4354
  from diffusers import DiffusionPipeline
@@ -4362,8 +4362,7 @@ pipe = DiffusionPipeline.from_pretrained("tolgacangoz/matryoshka-diffusion-model
4362
 
4363
  prompt0 = "a blue jay stops on the top of a helmet of Japanese samurai, background with sakura tree"
4364
  prompt = f"breathtaking {prompt0}. award-winning, professional, highly detailed"
4365
- negative_prompt = "deformed, mutated, ugly, disfigured, blur, blurry, noise, noisy"
4366
- image = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=50).images
4367
  make_image_grid(image, rows=1, cols=len(image))
4368
 
4369
  # pipe.change_nesting_level(<int>) # 0, 1, or 2
 
4336
 
4337
  **64x64**
4338
  :-------------------------:
4339
+ | <img src="https://github.com/user-attachments/assets/032738eb-c6cd-4fd9-b4d7-a7317b4b6528" width="222" height="222" alt="bird_64_64"> |
4340
 
4341
  - `256×256, nesting_level=1`: 1.776 GiB. With `150` DDIM inference steps:
4342
 
4343
  **64x64** | **256x256**
4344
  :-------------------------:|:-------------------------:
4345
+ | <img src="https://github.com/user-attachments/assets/21b9ad8b-eea6-4603-80a2-31180f391589" width="222" height="222" alt="bird_256_64"> | <img src="https://github.com/user-attachments/assets/fc411682-8a36-422c-9488-395b77d4406e" width="222" height="222" alt="bird_256_256"> |
4346
 
4347
+ - `1024×1024, nesting_level=2`: 1.792 GiB. As one can realize the cost of adding another layer is really negligible in this context! With `250` DDIM inference steps:
4348
 
4349
  **64x64** | **256x256** | **1024x1024**
4350
  :-------------------------:|:-------------------------:|:-------------------------:
4351
+ | <img src="https://github.com/user-attachments/assets/febf4b98-3dee-4a8e-9946-fd42e1f232e6" width="222" height="222" alt="bird_1024_64"> | <img src="https://github.com/user-attachments/assets/c5f85b40-5d6d-4267-a92a-c89dff015b9b" width="222" height="222" alt="bird_1024_256"> | <img src="https://github.com/user-attachments/assets/ad66b913-4367-4cb9-889e-bc06f4d96148" width="222" height="222" alt="bird_1024_1024"> |
4352
 
4353
  ```py
4354
  from diffusers import DiffusionPipeline
 
4362
 
4363
  prompt0 = "a blue jay stops on the top of a helmet of Japanese samurai, background with sakura tree"
4364
  prompt = f"breathtaking {prompt0}. award-winning, professional, highly detailed"
4365
+ image = pipe(prompt, num_inference_steps=50).images
 
4366
  make_image_grid(image, rows=1, cols=len(image))
4367
 
4368
  # pipe.change_nesting_level(<int>) # 0, 1, or 2
main/README_community_scripts.md CHANGED
@@ -8,6 +8,7 @@ If a community script doesn't work as expected, please open an issue and ping th
8
  |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
9
  | Using IP-Adapter with negative noise | Using negative noise with IP-adapter to better control the generation (see the [original post](https://github.com/huggingface/diffusers/discussions/7167) on the forum for more details) | [IP-Adapter Negative Noise](#ip-adapter-negative-noise) | | [Álvaro Somoza](https://github.com/asomoza)|
10
  | asymmetric tiling |configure seamless image tiling independently for the X and Y axes | [Asymmetric Tiling](#asymmetric-tiling ) | | [alexisrolland](https://github.com/alexisrolland)|
 
11
 
12
 
13
  ## Example usages
@@ -229,4 +230,86 @@ seamless_tiling(pipeline=pipeline, x_axis=False, y_axis=False)
229
 
230
  torch.cuda.empty_cache()
231
  image.save('image.png')
232
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  |:--------------------------------------------------------------------------------------------------------------------------------------|:---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:------------------------------------------------------------------------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------:|
9
  | Using IP-Adapter with negative noise | Using negative noise with IP-adapter to better control the generation (see the [original post](https://github.com/huggingface/diffusers/discussions/7167) on the forum for more details) | [IP-Adapter Negative Noise](#ip-adapter-negative-noise) | | [Álvaro Somoza](https://github.com/asomoza)|
10
  | asymmetric tiling |configure seamless image tiling independently for the X and Y axes | [Asymmetric Tiling](#asymmetric-tiling ) | | [alexisrolland](https://github.com/alexisrolland)|
11
+ | Prompt scheduling callback |Allows changing prompts during a generation | [Prompt Scheduling](#prompt-scheduling ) | | [hlky](https://github.com/hlky)|
12
 
13
 
14
  ## Example usages
 
230
 
231
  torch.cuda.empty_cache()
232
  image.save('image.png')
233
+ ```
234
+
235
+ ### Prompt Scheduling callback
236
+
237
+ Prompt scheduling callback allows changing prompts during a generation, like [prompt editing in A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#prompt-editing)
238
+
239
+ ```python
240
+ from diffusers import StableDiffusionPipeline
241
+ from diffusers.callbacks import PipelineCallback, MultiPipelineCallbacks
242
+ from diffusers.configuration_utils import register_to_config
243
+ import torch
244
+ from typing import Any, Dict, Optional
245
+
246
+
247
+ pipeline: StableDiffusionPipeline = StableDiffusionPipeline.from_pretrained(
248
+ "stable-diffusion-v1-5/stable-diffusion-v1-5",
249
+ torch_dtype=torch.float16,
250
+ variant="fp16",
251
+ use_safetensors=True,
252
+ ).to("cuda")
253
+ pipeline.safety_checker = None
254
+ pipeline.requires_safety_checker = False
255
+
256
+
257
+ class SDPromptScheduleCallback(PipelineCallback):
258
+ @register_to_config
259
+ def __init__(
260
+ self,
261
+ prompt: str,
262
+ negative_prompt: Optional[str] = None,
263
+ num_images_per_prompt: int = 1,
264
+ cutoff_step_ratio=1.0,
265
+ cutoff_step_index=None,
266
+ ):
267
+ super().__init__(
268
+ cutoff_step_ratio=cutoff_step_ratio, cutoff_step_index=cutoff_step_index
269
+ )
270
+
271
+ tensor_inputs = ["prompt_embeds"]
272
+
273
+ def callback_fn(
274
+ self, pipeline, step_index, timestep, callback_kwargs
275
+ ) -> Dict[str, Any]:
276
+ cutoff_step_ratio = self.config.cutoff_step_ratio
277
+ cutoff_step_index = self.config.cutoff_step_index
278
+
279
+ # Use cutoff_step_index if it's not None, otherwise use cutoff_step_ratio
280
+ cutoff_step = (
281
+ cutoff_step_index
282
+ if cutoff_step_index is not None
283
+ else int(pipeline.num_timesteps * cutoff_step_ratio)
284
+ )
285
+
286
+ if step_index == cutoff_step:
287
+ prompt_embeds, negative_prompt_embeds = pipeline.encode_prompt(
288
+ prompt=self.config.prompt,
289
+ negative_prompt=self.config.negative_prompt,
290
+ device=pipeline._execution_device,
291
+ num_images_per_prompt=self.config.num_images_per_prompt,
292
+ do_classifier_free_guidance=pipeline.do_classifier_free_guidance,
293
+ )
294
+ if pipeline.do_classifier_free_guidance:
295
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
296
+ callback_kwargs[self.tensor_inputs[0]] = prompt_embeds
297
+ return callback_kwargs
298
+
299
+ callback = MultiPipelineCallbacks(
300
+ [
301
+ SDPromptScheduleCallback(
302
+ prompt="Official portrait of a smiling world war ii general, female, cheerful, happy, detailed face, 20th century, highly detailed, cinematic lighting, digital art painting by Greg Rutkowski",
303
+ negative_prompt="Deformed, ugly, bad anatomy",
304
+ cutoff_step_ratio=0.25,
305
+ )
306
+ ]
307
+ )
308
+
309
+ image = pipeline(
310
+ prompt="Official portrait of a smiling world war ii general, male, cheerful, happy, detailed face, 20th century, highly detailed, cinematic lighting, digital art painting by Greg Rutkowski",
311
+ negative_prompt="Deformed, ugly, bad anatomy",
312
+ callback_on_step_end=callback,
313
+ callback_on_step_end_tensor_inputs=["prompt_embeds"],
314
+ ).images[0]
315
+ ```
main/matryoshka.py CHANGED
@@ -107,15 +107,16 @@ EXAMPLE_DOC_STRING = """
107
 
108
  >>> # nesting_level=0 -> 64x64; nesting_level=1 -> 256x256 - 64x64; nesting_level=2 -> 1024x1024 - 256x256 - 64x64
109
  >>> pipe = DiffusionPipeline.from_pretrained("tolgacangoz/matryoshka-diffusion-models",
110
- >>> custom_pipeline="matryoshka").to("cuda")
 
 
111
 
112
  >>> prompt0 = "a blue jay stops on the top of a helmet of Japanese samurai, background with sakura tree"
113
  >>> prompt = f"breathtaking {prompt0}. award-winning, professional, highly detailed"
114
- >>> negative_prompt = "deformed, mutated, ugly, disfigured, blur, blurry, noise, noisy"
115
- >>> image = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=50).images
116
  >>> make_image_grid(image, rows=1, cols=len(image))
117
 
118
- >>> pipe.change_nesting_level(<int>) # 0, 1, or 2
119
  >>> # 50+, 100+, and 250+ num_inference_steps are recommended for nesting levels 0, 1, and 2 respectively.
120
  ```
121
  """
@@ -420,6 +421,7 @@ class MatryoshkaDDIMScheduler(SchedulerMixin, ConfigMixin):
420
  self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
421
 
422
  self.scales = None
 
423
 
424
  def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor:
425
  """
@@ -532,6 +534,7 @@ class MatryoshkaDDIMScheduler(SchedulerMixin, ConfigMixin):
532
 
533
  def get_schedule_shifted(self, alpha_prod, scale_factor=None):
534
  if (scale_factor is not None) and (scale_factor > 1): # rescale noise schedule
 
535
  snr = alpha_prod / (1 - alpha_prod)
536
  scaled_snr = snr / scale_factor
537
  alpha_prod = 1 / (1 + 1 / scaled_snr)
@@ -639,17 +642,14 @@ class MatryoshkaDDIMScheduler(SchedulerMixin, ConfigMixin):
639
  # 4. Clip or threshold "predicted x_0"
640
  if self.config.thresholding:
641
  if len(model_output) > 1:
642
- pred_original_sample = [
643
- self._threshold_sample(p_o_s * scale) / scale
644
- for p_o_s, scale in zip(pred_original_sample, self.scales)
645
- ]
646
  else:
647
  pred_original_sample = self._threshold_sample(pred_original_sample)
648
  elif self.config.clip_sample:
649
  if len(model_output) > 1:
650
  pred_original_sample = [
651
- (p_o_s * scale).clamp(-self.config.clip_sample_range, self.config.clip_sample_range) / scale
652
- for p_o_s, scale in zip(pred_original_sample, self.scales)
653
  ]
654
  else:
655
  pred_original_sample = pred_original_sample.clamp(
@@ -3816,6 +3816,8 @@ class MatryoshkaPipeline(
3816
 
3817
  if hasattr(unet, "nest_ratio"):
3818
  scheduler.scales = unet.nest_ratio + [1]
 
 
3819
 
3820
  self.register_modules(
3821
  text_encoder=text_encoder,
@@ -3842,12 +3844,14 @@ class MatryoshkaPipeline(
3842
  ).to(self.device)
3843
  self.config.nesting_level = 1
3844
  self.scheduler.scales = self.unet.nest_ratio + [1]
 
3845
  elif nesting_level == 2:
3846
  self.unet = NestedUNet2DConditionModel.from_pretrained(
3847
  "tolgacangoz/matryoshka-diffusion-models", subfolder="unet/nesting_level_2"
3848
  ).to(self.device)
3849
  self.config.nesting_level = 2
3850
  self.scheduler.scales = self.unet.nest_ratio + [1]
 
3851
  else:
3852
  raise ValueError("Currently, nesting levels 0, 1, and 2 are supported.")
3853
 
@@ -4627,8 +4631,8 @@ class MatryoshkaPipeline(
4627
  image = latents
4628
 
4629
  if self.scheduler.scales is not None:
4630
- for i, (img, scale) in enumerate(zip(image, self.scheduler.scales)):
4631
- image[i] = self.image_processor.postprocess(img * scale, output_type=output_type)[0]
4632
  else:
4633
  image = self.image_processor.postprocess(image, output_type=output_type)
4634
 
 
107
 
108
  >>> # nesting_level=0 -> 64x64; nesting_level=1 -> 256x256 - 64x64; nesting_level=2 -> 1024x1024 - 256x256 - 64x64
109
  >>> pipe = DiffusionPipeline.from_pretrained("tolgacangoz/matryoshka-diffusion-models",
110
+ ... nesting_level=0,
111
+ ... trust_remote_code=False, # One needs to give permission for this code to run
112
+ ... ).to("cuda")
113
 
114
  >>> prompt0 = "a blue jay stops on the top of a helmet of Japanese samurai, background with sakura tree"
115
  >>> prompt = f"breathtaking {prompt0}. award-winning, professional, highly detailed"
116
+ >>> image = pipe(prompt, num_inference_steps=50).images
 
117
  >>> make_image_grid(image, rows=1, cols=len(image))
118
 
119
+ >>> # pipe.change_nesting_level(<int>) # 0, 1, or 2
120
  >>> # 50+, 100+, and 250+ num_inference_steps are recommended for nesting levels 0, 1, and 2 respectively.
121
  ```
122
  """
 
421
  self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
422
 
423
  self.scales = None
424
+ self.schedule_shifted_power = 1.0
425
 
426
  def scale_model_input(self, sample: torch.Tensor, timestep: Optional[int] = None) -> torch.Tensor:
427
  """
 
534
 
535
  def get_schedule_shifted(self, alpha_prod, scale_factor=None):
536
  if (scale_factor is not None) and (scale_factor > 1): # rescale noise schedule
537
+ scale_factor = scale_factor**self.schedule_shifted_power
538
  snr = alpha_prod / (1 - alpha_prod)
539
  scaled_snr = snr / scale_factor
540
  alpha_prod = 1 / (1 + 1 / scaled_snr)
 
642
  # 4. Clip or threshold "predicted x_0"
643
  if self.config.thresholding:
644
  if len(model_output) > 1:
645
+ pred_original_sample = [self._threshold_sample(p_o_s) for p_o_s in pred_original_sample]
 
 
 
646
  else:
647
  pred_original_sample = self._threshold_sample(pred_original_sample)
648
  elif self.config.clip_sample:
649
  if len(model_output) > 1:
650
  pred_original_sample = [
651
+ p_o_s.clamp(-self.config.clip_sample_range, self.config.clip_sample_range)
652
+ for p_o_s in pred_original_sample
653
  ]
654
  else:
655
  pred_original_sample = pred_original_sample.clamp(
 
3816
 
3817
  if hasattr(unet, "nest_ratio"):
3818
  scheduler.scales = unet.nest_ratio + [1]
3819
+ if nesting_level == 2:
3820
+ scheduler.schedule_shifted_power = 2.0
3821
 
3822
  self.register_modules(
3823
  text_encoder=text_encoder,
 
3844
  ).to(self.device)
3845
  self.config.nesting_level = 1
3846
  self.scheduler.scales = self.unet.nest_ratio + [1]
3847
+ self.scheduler.schedule_shifted_power = 1.0
3848
  elif nesting_level == 2:
3849
  self.unet = NestedUNet2DConditionModel.from_pretrained(
3850
  "tolgacangoz/matryoshka-diffusion-models", subfolder="unet/nesting_level_2"
3851
  ).to(self.device)
3852
  self.config.nesting_level = 2
3853
  self.scheduler.scales = self.unet.nest_ratio + [1]
3854
+ self.scheduler.schedule_shifted_power = 2.0
3855
  else:
3856
  raise ValueError("Currently, nesting levels 0, 1, and 2 are supported.")
3857
 
 
4631
  image = latents
4632
 
4633
  if self.scheduler.scales is not None:
4634
+ for i, img in enumerate(image):
4635
+ image[i] = self.image_processor.postprocess(img, output_type=output_type)[0]
4636
  else:
4637
  image = self.image_processor.postprocess(image, output_type=output_type)
4638