diffusers/community-pipelines-mirror
Community Pipeline Examples For more information about community pipelines, please have a look at this issue. Community pipeline examples consist pipelines that have been added by the community. Please have a look at the following tables to get an overview of all community examples. Click on the Code Example to get a copy-and-paste ready code example that you can try out. If a community pipeline doesn't work as expected, please open an issue and ping the author on it. Please… See the full description on the dataset page: https://huggingface.co/datasets/diffusers/community-pipelines-mirror.
922k
1# Inspired by: https://github.com/Mikubill/sd-webui-controlnet/discussions/1236 and https://github.com/Mikubill/sd-webui-controlnet/discussions/12802from typing import Any, Callable, Dict, List, Optional, Tuple, Union3 4import numpy as np5import PIL.Image6import torch7 8from diffusers import StableDiffusionPipeline9from diffusers.models.attention import BasicTransformerBlock10from diffusers.models.unet_2d_blocks import CrossAttnDownBlock2D, CrossAttnUpBlock2D, DownBlock2D, UpBlock2D11from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput12from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import rescale_noise_cfg13from diffusers.utils import PIL_INTERPOLATION, logging, randn_tensor14 15 16logger = logging.get_logger(__name__) # pylint: disable=invalid-name17 18EXAMPLE_DOC_STRING = """19 Examples:20 ```py21 >>> import torch22 >>> from diffusers import UniPCMultistepScheduler23 >>> from diffusers.utils import load_image24 25 >>> input_image = load_image("https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png")26 27 >>> pipe = StableDiffusionReferencePipeline.from_pretrained(28 "runwayml/stable-diffusion-v1-5",29 safety_checker=None,30 torch_dtype=torch.float1631 ).to('cuda:0')32 33 >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe_controlnet.scheduler.config)34 35 >>> result_img = pipe(ref_image=input_image,36 prompt="1girl",37 num_inference_steps=20,38 reference_attn=True,39 reference_adain=True).images[0]40 41 >>> result_img.show()42 ```43"""44 45 46def torch_dfs(model: torch.nn.Module):47 result = [model]48 for child in model.children():49 result += torch_dfs(child)50 return result51 52 53class StableDiffusionReferencePipeline(StableDiffusionPipeline):54 def _default_height_width(self, height, width, image):55 # NOTE: It is possible that a list of images have different56 # dimensions for each image, so just checking the first image57 # is not _exactly_ correct, but it is simple.58 while isinstance(image, list):59 image = image[0]60 61 if height is None:62 if isinstance(image, PIL.Image.Image):63 height = image.height64 elif isinstance(image, torch.Tensor):65 height = image.shape[2]66 67 height = (height // 8) * 8 # round down to nearest multiple of 868 69 if width is None:70 if isinstance(image, PIL.Image.Image):71 width = image.width72 elif isinstance(image, torch.Tensor):73 width = image.shape[3]74 75 width = (width // 8) * 8 # round down to nearest multiple of 876 77 return height, width78 79 def prepare_image(80 self,81 image,82 width,83 height,84 batch_size,85 num_images_per_prompt,86 device,87 dtype,88 do_classifier_free_guidance=False,89 guess_mode=False,90 ):91 if not isinstance(image, torch.Tensor):92 if isinstance(image, PIL.Image.Image):93 image = [image]94 95 if isinstance(image[0], PIL.Image.Image):96 images = []97 98 for image_ in image:99 image_ = image_.convert("RGB")100 image_ = image_.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])101 image_ = np.array(image_)102 image_ = image_[None, :]103 images.append(image_)104 105 image = images106 107 image = np.concatenate(image, axis=0)108 image = np.array(image).astype(np.float32) / 255.0109 image = (image - 0.5) / 0.5110 image = image.transpose(0, 3, 1, 2)111 image = torch.from_numpy(image)112 elif isinstance(image[0], torch.Tensor):113 image = torch.cat(image, dim=0)114 115 image_batch_size = image.shape[0]116 117 if image_batch_size == 1:118 repeat_by = batch_size119 else:120 # image batch size is the same as prompt batch size121 repeat_by = num_images_per_prompt122 123 image = image.repeat_interleave(repeat_by, dim=0)124 125 image = image.to(device=device, dtype=dtype)126 127 if do_classifier_free_guidance and not guess_mode:128 image = torch.cat([image] * 2)129 130 return image131 132 def prepare_ref_latents(self, refimage, batch_size, dtype, device, generator, do_classifier_free_guidance):133 refimage = refimage.to(device=device, dtype=dtype)134 135 # encode the mask image into latents space so we can concatenate it to the latents136 if isinstance(generator, list):137 ref_image_latents = [138 self.vae.encode(refimage[i : i + 1]).latent_dist.sample(generator=generator[i])139 for i in range(batch_size)140 ]141 ref_image_latents = torch.cat(ref_image_latents, dim=0)142 else:143 ref_image_latents = self.vae.encode(refimage).latent_dist.sample(generator=generator)144 ref_image_latents = self.vae.config.scaling_factor * ref_image_latents145 146 # duplicate mask and ref_image_latents for each generation per prompt, using mps friendly method147 if ref_image_latents.shape[0] < batch_size:148 if not batch_size % ref_image_latents.shape[0] == 0:149 raise ValueError(150 "The passed images and the required batch size don't match. Images are supposed to be duplicated"151 f" to a total batch size of {batch_size}, but {ref_image_latents.shape[0]} images were passed."152 " Make sure the number of images that you pass is divisible by the total requested batch size."153 )154 ref_image_latents = ref_image_latents.repeat(batch_size // ref_image_latents.shape[0], 1, 1, 1)155 156 # aligning device to prevent device errors when concating it with the latent model input157 ref_image_latents = ref_image_latents.to(device=device, dtype=dtype)158 return ref_image_latents159 160 @torch.no_grad()161 def __call__(162 self,163 prompt: Union[str, List[str]] = None,164 ref_image: Union[torch.FloatTensor, PIL.Image.Image] = None,165 height: Optional[int] = None,166 width: Optional[int] = None,167 num_inference_steps: int = 50,168 guidance_scale: float = 7.5,169 negative_prompt: Optional[Union[str, List[str]]] = None,170 num_images_per_prompt: Optional[int] = 1,171 eta: float = 0.0,172 generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,173 latents: Optional[torch.FloatTensor] = None,174 prompt_embeds: Optional[torch.FloatTensor] = None,175 negative_prompt_embeds: Optional[torch.FloatTensor] = None,176 output_type: Optional[str] = "pil",177 return_dict: bool = True,178 callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,179 callback_steps: int = 1,180 cross_attention_kwargs: Optional[Dict[str, Any]] = None,181 guidance_rescale: float = 0.0,182 attention_auto_machine_weight: float = 1.0,183 gn_auto_machine_weight: float = 1.0,184 style_fidelity: float = 0.5,185 reference_attn: bool = True,186 reference_adain: bool = True,187 ):188 r"""189 Function invoked when calling the pipeline for generation.190 191 Args:192 prompt (`str` or `List[str]`, *optional*):193 The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.194 instead.195 ref_image (`torch.FloatTensor`, `PIL.Image.Image`):196 The Reference Control input condition. Reference Control uses this input condition to generate guidance to Unet. If197 the type is specified as `Torch.FloatTensor`, it is passed to Reference Control as is. `PIL.Image.Image` can198 also be accepted as an image.199 height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):200 The height in pixels of the generated image.201 width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):202 The width in pixels of the generated image.203 num_inference_steps (`int`, *optional*, defaults to 50):204 The number of denoising steps. More denoising steps usually lead to a higher quality image at the205 expense of slower inference.206 guidance_scale (`float`, *optional*, defaults to 7.5):207 Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).208 `guidance_scale` is defined as `w` of equation 2. of [Imagen209 Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >210 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,211 usually at the expense of lower image quality.212 negative_prompt (`str` or `List[str]`, *optional*):213 The prompt or prompts not to guide the image generation. If not defined, one has to pass214 `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is215 less than `1`).216 num_images_per_prompt (`int`, *optional*, defaults to 1):217 The number of images to generate per prompt.218 eta (`float`, *optional*, defaults to 0.0):219 Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to220 [`schedulers.DDIMScheduler`], will be ignored for others.221 generator (`torch.Generator` or `List[torch.Generator]`, *optional*):222 One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)223 to make generation deterministic.224 latents (`torch.FloatTensor`, *optional*):225 Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image226 generation. Can be used to tweak the same generation with different prompts. If not provided, a latents227 tensor will ge generated by sampling using the supplied random `generator`.228 prompt_embeds (`torch.FloatTensor`, *optional*):229 Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not230 provided, text embeddings will be generated from `prompt` input argument.231 negative_prompt_embeds (`torch.FloatTensor`, *optional*):232 Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt233 weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input234 argument.235 output_type (`str`, *optional*, defaults to `"pil"`):236 The output format of the generate image. Choose between237 [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.238 return_dict (`bool`, *optional*, defaults to `True`):239 Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a240 plain tuple.241 callback (`Callable`, *optional*):242 A function that will be called every `callback_steps` steps during inference. The function will be243 called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.244 callback_steps (`int`, *optional*, defaults to 1):245 The frequency at which the `callback` function will be called. If not specified, the callback will be246 called at every step.247 cross_attention_kwargs (`dict`, *optional*):248 A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under249 `self.processor` in250 [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).251 guidance_rescale (`float`, *optional*, defaults to 0.7):252 Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are253 Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of254 [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).255 Guidance rescale factor should fix overexposure when using zero terminal SNR.256 attention_auto_machine_weight (`float`):257 Weight of using reference query for self attention's context.258 If attention_auto_machine_weight=1.0, use reference query for all self attention's context.259 gn_auto_machine_weight (`float`):260 Weight of using reference adain. If gn_auto_machine_weight=2.0, use all reference adain plugins.261 style_fidelity (`float`):262 style fidelity of ref_uncond_xt. If style_fidelity=1.0, control more important,263 elif style_fidelity=0.0, prompt more important, else balanced.264 reference_attn (`bool`):265 Whether to use reference query for self attention's context.266 reference_adain (`bool`):267 Whether to use reference adain.268 269 Examples:270 271 Returns:272 [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:273 [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.274 When returning a tuple, the first element is a list with the generated images, and the second element is a275 list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"276 (nsfw) content, according to the `safety_checker`.277 """278 assert reference_attn or reference_adain, "`reference_attn` or `reference_adain` must be True."279 280 # 0. Default height and width to unet281 height, width = self._default_height_width(height, width, ref_image)282 283 # 1. Check inputs. Raise error if not correct284 self.check_inputs(285 prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds286 )287 288 # 2. Define call parameters289 if prompt is not None and isinstance(prompt, str):290 batch_size = 1291 elif prompt is not None and isinstance(prompt, list):292 batch_size = len(prompt)293 else:294 batch_size = prompt_embeds.shape[0]295 296 device = self._execution_device297 # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)298 # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`299 # corresponds to doing no classifier free guidance.300 do_classifier_free_guidance = guidance_scale > 1.0301 302 # 3. Encode input prompt303 text_encoder_lora_scale = (304 cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None305 )306 prompt_embeds = self._encode_prompt(307 prompt,308 device,309 num_images_per_prompt,310 do_classifier_free_guidance,311 negative_prompt,312 prompt_embeds=prompt_embeds,313 negative_prompt_embeds=negative_prompt_embeds,314 lora_scale=text_encoder_lora_scale,315 )316 317 # 4. Preprocess reference image318 ref_image = self.prepare_image(319 image=ref_image,320 width=width,321 height=height,322 batch_size=batch_size * num_images_per_prompt,323 num_images_per_prompt=num_images_per_prompt,324 device=device,325 dtype=prompt_embeds.dtype,326 )327 328 # 5. Prepare timesteps329 self.scheduler.set_timesteps(num_inference_steps, device=device)330 timesteps = self.scheduler.timesteps331 332 # 6. Prepare latent variables333 num_channels_latents = self.unet.config.in_channels334 latents = self.prepare_latents(335 batch_size * num_images_per_prompt,336 num_channels_latents,337 height,338 width,339 prompt_embeds.dtype,340 device,341 generator,342 latents,343 )344 345 # 7. Prepare reference latent variables346 ref_image_latents = self.prepare_ref_latents(347 ref_image,348 batch_size * num_images_per_prompt,349 prompt_embeds.dtype,350 device,351 generator,352 do_classifier_free_guidance,353 )354 355 # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline356 extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)357 358 # 9. Modify self attention and group norm359 MODE = "write"360 uc_mask = (361 torch.Tensor([1] * batch_size * num_images_per_prompt + [0] * batch_size * num_images_per_prompt)362 .type_as(ref_image_latents)363 .bool()364 )365 366 def hacked_basic_transformer_inner_forward(367 self,368 hidden_states: torch.FloatTensor,369 attention_mask: Optional[torch.FloatTensor] = None,370 encoder_hidden_states: Optional[torch.FloatTensor] = None,371 encoder_attention_mask: Optional[torch.FloatTensor] = None,372 timestep: Optional[torch.LongTensor] = None,373 cross_attention_kwargs: Dict[str, Any] = None,374 class_labels: Optional[torch.LongTensor] = None,375 ):376 if self.use_ada_layer_norm:377 norm_hidden_states = self.norm1(hidden_states, timestep)378 elif self.use_ada_layer_norm_zero:379 norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(380 hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype381 )382 else:383 norm_hidden_states = self.norm1(hidden_states)384 385 # 1. Self-Attention386 cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}387 if self.only_cross_attention:388 attn_output = self.attn1(389 norm_hidden_states,390 encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,391 attention_mask=attention_mask,392 **cross_attention_kwargs,393 )394 else:395 if MODE == "write":396 self.bank.append(norm_hidden_states.detach().clone())397 attn_output = self.attn1(398 norm_hidden_states,399 encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,400 attention_mask=attention_mask,401 **cross_attention_kwargs,402 )403 if MODE == "read":404 if attention_auto_machine_weight > self.attn_weight:405 attn_output_uc = self.attn1(406 norm_hidden_states,407 encoder_hidden_states=torch.cat([norm_hidden_states] + self.bank, dim=1),408 # attention_mask=attention_mask,409 **cross_attention_kwargs,410 )411 attn_output_c = attn_output_uc.clone()412 if do_classifier_free_guidance and style_fidelity > 0:413 attn_output_c[uc_mask] = self.attn1(414 norm_hidden_states[uc_mask],415 encoder_hidden_states=norm_hidden_states[uc_mask],416 **cross_attention_kwargs,417 )418 attn_output = style_fidelity * attn_output_c + (1.0 - style_fidelity) * attn_output_uc419 self.bank.clear()420 else:421 attn_output = self.attn1(422 norm_hidden_states,423 encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,424 attention_mask=attention_mask,425 **cross_attention_kwargs,426 )427 if self.use_ada_layer_norm_zero:428 attn_output = gate_msa.unsqueeze(1) * attn_output429 hidden_states = attn_output + hidden_states430 431 if self.attn2 is not None:432 norm_hidden_states = (433 self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)434 )435 436 # 2. Cross-Attention437 attn_output = self.attn2(438 norm_hidden_states,439 encoder_hidden_states=encoder_hidden_states,440 attention_mask=encoder_attention_mask,441 **cross_attention_kwargs,442 )443 hidden_states = attn_output + hidden_states444 445 # 3. Feed-forward446 norm_hidden_states = self.norm3(hidden_states)447 448 if self.use_ada_layer_norm_zero:449 norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]450 451 ff_output = self.ff(norm_hidden_states)452 453 if self.use_ada_layer_norm_zero:454 ff_output = gate_mlp.unsqueeze(1) * ff_output455 456 hidden_states = ff_output + hidden_states457 458 return hidden_states459 460 def hacked_mid_forward(self, *args, **kwargs):461 eps = 1e-6462 x = self.original_forward(*args, **kwargs)463 if MODE == "write":464 if gn_auto_machine_weight >= self.gn_weight:465 var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0)466 self.mean_bank.append(mean)467 self.var_bank.append(var)468 if MODE == "read":469 if len(self.mean_bank) > 0 and len(self.var_bank) > 0:470 var, mean = torch.var_mean(x, dim=(2, 3), keepdim=True, correction=0)471 std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5472 mean_acc = sum(self.mean_bank) / float(len(self.mean_bank))473 var_acc = sum(self.var_bank) / float(len(self.var_bank))474 std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5475 x_uc = (((x - mean) / std) * std_acc) + mean_acc476 x_c = x_uc.clone()477 if do_classifier_free_guidance and style_fidelity > 0:478 x_c[uc_mask] = x[uc_mask]479 x = style_fidelity * x_c + (1.0 - style_fidelity) * x_uc480 self.mean_bank = []481 self.var_bank = []482 return x483 484 def hack_CrossAttnDownBlock2D_forward(485 self,486 hidden_states: torch.FloatTensor,487 temb: Optional[torch.FloatTensor] = None,488 encoder_hidden_states: Optional[torch.FloatTensor] = None,489 attention_mask: Optional[torch.FloatTensor] = None,490 cross_attention_kwargs: Optional[Dict[str, Any]] = None,491 encoder_attention_mask: Optional[torch.FloatTensor] = None,492 ):493 eps = 1e-6494 495 # TODO(Patrick, William) - attention mask is not used496 output_states = ()497 498 for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):499 hidden_states = resnet(hidden_states, temb)500 hidden_states = attn(501 hidden_states,502 encoder_hidden_states=encoder_hidden_states,503 cross_attention_kwargs=cross_attention_kwargs,504 attention_mask=attention_mask,505 encoder_attention_mask=encoder_attention_mask,506 return_dict=False,507 )[0]508 if MODE == "write":509 if gn_auto_machine_weight >= self.gn_weight:510 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)511 self.mean_bank.append([mean])512 self.var_bank.append([var])513 if MODE == "read":514 if len(self.mean_bank) > 0 and len(self.var_bank) > 0:515 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)516 std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5517 mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))518 var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))519 std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5520 hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc521 hidden_states_c = hidden_states_uc.clone()522 if do_classifier_free_guidance and style_fidelity > 0:523 hidden_states_c[uc_mask] = hidden_states[uc_mask]524 hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc525 526 output_states = output_states + (hidden_states,)527 528 if MODE == "read":529 self.mean_bank = []530 self.var_bank = []531 532 if self.downsamplers is not None:533 for downsampler in self.downsamplers:534 hidden_states = downsampler(hidden_states)535 536 output_states = output_states + (hidden_states,)537 538 return hidden_states, output_states539 540 def hacked_DownBlock2D_forward(self, hidden_states, temb=None):541 eps = 1e-6542 543 output_states = ()544 545 for i, resnet in enumerate(self.resnets):546 hidden_states = resnet(hidden_states, temb)547 548 if MODE == "write":549 if gn_auto_machine_weight >= self.gn_weight:550 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)551 self.mean_bank.append([mean])552 self.var_bank.append([var])553 if MODE == "read":554 if len(self.mean_bank) > 0 and len(self.var_bank) > 0:555 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)556 std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5557 mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))558 var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))559 std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5560 hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc561 hidden_states_c = hidden_states_uc.clone()562 if do_classifier_free_guidance and style_fidelity > 0:563 hidden_states_c[uc_mask] = hidden_states[uc_mask]564 hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc565 566 output_states = output_states + (hidden_states,)567 568 if MODE == "read":569 self.mean_bank = []570 self.var_bank = []571 572 if self.downsamplers is not None:573 for downsampler in self.downsamplers:574 hidden_states = downsampler(hidden_states)575 576 output_states = output_states + (hidden_states,)577 578 return hidden_states, output_states579 580 def hacked_CrossAttnUpBlock2D_forward(581 self,582 hidden_states: torch.FloatTensor,583 res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],584 temb: Optional[torch.FloatTensor] = None,585 encoder_hidden_states: Optional[torch.FloatTensor] = None,586 cross_attention_kwargs: Optional[Dict[str, Any]] = None,587 upsample_size: Optional[int] = None,588 attention_mask: Optional[torch.FloatTensor] = None,589 encoder_attention_mask: Optional[torch.FloatTensor] = None,590 ):591 eps = 1e-6592 # TODO(Patrick, William) - attention mask is not used593 for i, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)):594 # pop res hidden states595 res_hidden_states = res_hidden_states_tuple[-1]596 res_hidden_states_tuple = res_hidden_states_tuple[:-1]597 hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)598 hidden_states = resnet(hidden_states, temb)599 hidden_states = attn(600 hidden_states,601 encoder_hidden_states=encoder_hidden_states,602 cross_attention_kwargs=cross_attention_kwargs,603 attention_mask=attention_mask,604 encoder_attention_mask=encoder_attention_mask,605 return_dict=False,606 )[0]607 608 if MODE == "write":609 if gn_auto_machine_weight >= self.gn_weight:610 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)611 self.mean_bank.append([mean])612 self.var_bank.append([var])613 if MODE == "read":614 if len(self.mean_bank) > 0 and len(self.var_bank) > 0:615 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)616 std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5617 mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))618 var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))619 std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5620 hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc621 hidden_states_c = hidden_states_uc.clone()622 if do_classifier_free_guidance and style_fidelity > 0:623 hidden_states_c[uc_mask] = hidden_states[uc_mask]624 hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc625 626 if MODE == "read":627 self.mean_bank = []628 self.var_bank = []629 630 if self.upsamplers is not None:631 for upsampler in self.upsamplers:632 hidden_states = upsampler(hidden_states, upsample_size)633 634 return hidden_states635 636 def hacked_UpBlock2D_forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):637 eps = 1e-6638 for i, resnet in enumerate(self.resnets):639 # pop res hidden states640 res_hidden_states = res_hidden_states_tuple[-1]641 res_hidden_states_tuple = res_hidden_states_tuple[:-1]642 hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)643 hidden_states = resnet(hidden_states, temb)644 645 if MODE == "write":646 if gn_auto_machine_weight >= self.gn_weight:647 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)648 self.mean_bank.append([mean])649 self.var_bank.append([var])650 if MODE == "read":651 if len(self.mean_bank) > 0 and len(self.var_bank) > 0:652 var, mean = torch.var_mean(hidden_states, dim=(2, 3), keepdim=True, correction=0)653 std = torch.maximum(var, torch.zeros_like(var) + eps) ** 0.5654 mean_acc = sum(self.mean_bank[i]) / float(len(self.mean_bank[i]))655 var_acc = sum(self.var_bank[i]) / float(len(self.var_bank[i]))656 std_acc = torch.maximum(var_acc, torch.zeros_like(var_acc) + eps) ** 0.5657 hidden_states_uc = (((hidden_states - mean) / std) * std_acc) + mean_acc658 hidden_states_c = hidden_states_uc.clone()659 if do_classifier_free_guidance and style_fidelity > 0:660 hidden_states_c[uc_mask] = hidden_states[uc_mask]661 hidden_states = style_fidelity * hidden_states_c + (1.0 - style_fidelity) * hidden_states_uc662 663 if MODE == "read":664 self.mean_bank = []665 self.var_bank = []666 667 if self.upsamplers is not None:668 for upsampler in self.upsamplers:669 hidden_states = upsampler(hidden_states, upsample_size)670 671 return hidden_states672 673 if reference_attn:674 attn_modules = [module for module in torch_dfs(self.unet) if isinstance(module, BasicTransformerBlock)]675 attn_modules = sorted(attn_modules, key=lambda x: -x.norm1.normalized_shape[0])676 677 for i, module in enumerate(attn_modules):678 module._original_inner_forward = module.forward679 module.forward = hacked_basic_transformer_inner_forward.__get__(module, BasicTransformerBlock)680 module.bank = []681 module.attn_weight = float(i) / float(len(attn_modules))682 683 if reference_adain:684 gn_modules = [self.unet.mid_block]685 self.unet.mid_block.gn_weight = 0686 687 down_blocks = self.unet.down_blocks688 for w, module in enumerate(down_blocks):689 module.gn_weight = 1.0 - float(w) / float(len(down_blocks))690 gn_modules.append(module)691 692 up_blocks = self.unet.up_blocks693 for w, module in enumerate(up_blocks):694 module.gn_weight = float(w) / float(len(up_blocks))695 gn_modules.append(module)696 697 for i, module in enumerate(gn_modules):698 if getattr(module, "original_forward", None) is None:699 module.original_forward = module.forward700 if i == 0:701 # mid_block702 module.forward = hacked_mid_forward.__get__(module, torch.nn.Module)703 elif isinstance(module, CrossAttnDownBlock2D):704 module.forward = hack_CrossAttnDownBlock2D_forward.__get__(module, CrossAttnDownBlock2D)705 elif isinstance(module, DownBlock2D):706 module.forward = hacked_DownBlock2D_forward.__get__(module, DownBlock2D)707 elif isinstance(module, CrossAttnUpBlock2D):708 module.forward = hacked_CrossAttnUpBlock2D_forward.__get__(module, CrossAttnUpBlock2D)709 elif isinstance(module, UpBlock2D):710 module.forward = hacked_UpBlock2D_forward.__get__(module, UpBlock2D)711 module.mean_bank = []712 module.var_bank = []713 module.gn_weight *= 2714 715 # 10. Denoising loop716 num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order717 with self.progress_bar(total=num_inference_steps) as progress_bar:718 for i, t in enumerate(timesteps):719 # expand the latents if we are doing classifier free guidance720 latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents721 latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)722 723 # ref only part724 noise = randn_tensor(725 ref_image_latents.shape, generator=generator, device=device, dtype=ref_image_latents.dtype726 )727 ref_xt = self.scheduler.add_noise(728 ref_image_latents,729 noise,730 t.reshape(731 1,732 ),733 )734 ref_xt = torch.cat([ref_xt] * 2) if do_classifier_free_guidance else ref_xt735 ref_xt = self.scheduler.scale_model_input(ref_xt, t)736 737 MODE = "write"738 self.unet(739 ref_xt,740 t,741 encoder_hidden_states=prompt_embeds,742 cross_attention_kwargs=cross_attention_kwargs,743 return_dict=False,744 )745 746 # predict the noise residual747 MODE = "read"748 noise_pred = self.unet(749 latent_model_input,750 t,751 encoder_hidden_states=prompt_embeds,752 cross_attention_kwargs=cross_attention_kwargs,753 return_dict=False,754 )[0]755 756 # perform guidance757 if do_classifier_free_guidance:758 noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)759 noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)760 761 if do_classifier_free_guidance and guidance_rescale > 0.0:762 # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf763 noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)764 765 # compute the previous noisy sample x_t -> x_t-1766 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]767 768 # call the callback, if provided769 if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):770 progress_bar.update()771 if callback is not None and i % callback_steps == 0:772 callback(i, t, latents)773 774 if not output_type == "latent":775 image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]776 image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)777 else:778 image = latents779 has_nsfw_concept = None780 781 if has_nsfw_concept is None:782 do_denormalize = [True] * image.shape[0]783 else:784 do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]785 786 image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)787 788 # Offload last model to CPU789 if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:790 self.final_offload_hook.offload()791 792 if not return_dict:793 return (image, has_nsfw_concept)794 795 return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)796 