from typing import Any, Callable, Dict, List, Optional, Union, Tuple import PIL import torch import torch.nn as nn from safetensors import safe_open from huggingface_hub.utils import validate_hf_hub_args from diffusers import StableDiffusionXLPipeline from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput from diffusers.utils import _get_model_file from transformers import CLIPImageProcessor from transformers.models.clip.modeling_clip import CLIPVisionModelWithProjection from transformers.models.clip.configuration_clip import CLIPVisionConfig PipelineImageInput = Union[ PIL.Image.Image, torch.FloatTensor, List[PIL.Image.Image], List[torch.FloatTensor], ] VISION_CONFIG_DICT = { "hidden_size": 1024, "intermediate_size": 4096, "num_attention_heads": 16, "num_hidden_layers": 24, "patch_size": 14, "projection_dim": 768 } # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): """ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 """ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) # rescale the results from guidance (fixes overexposure) noise_pred_rescaled = noise_cfg * (std_text / std_cfg) # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg return noise_cfg class MLP(nn.Module): def __init__(self, in_dim, out_dim, hidden_dim, use_residual=True): super().__init__() if use_residual: assert in_dim == out_dim self.layernorm = nn.LayerNorm(in_dim) self.fc1 = nn.Linear(in_dim, hidden_dim) self.fc2 = nn.Linear(hidden_dim, out_dim) self.use_residual = use_residual self.act_fn = nn.GELU() def forward(self, x): residual = x x = self.layernorm(x) x = self.fc1(x) x = self.act_fn(x) x = self.fc2(x) if self.use_residual: x = x + residual return x class FuseModule(nn.Module): def __init__(self, embed_dim): super().__init__() self.mlp1 = MLP(embed_dim * 2, embed_dim, embed_dim, use_residual=False) self.mlp2 = MLP(embed_dim, embed_dim, embed_dim, use_residual=True) self.layer_norm = nn.LayerNorm(embed_dim) def fuse_fn(self, prompt_embeds, id_embeds): stacked_id_embeds = torch.cat([prompt_embeds, id_embeds], dim=-1) stacked_id_embeds = self.mlp1(stacked_id_embeds) + prompt_embeds stacked_id_embeds = self.mlp2(stacked_id_embeds) stacked_id_embeds = self.layer_norm(stacked_id_embeds) return stacked_id_embeds def forward( self, prompt_embeds, id_embeds, class_tokens_mask, ) -> torch.Tensor: # id_embeds shape: [b, max_num_inputs, 1, 2048] id_embeds = id_embeds.to(prompt_embeds.dtype) num_inputs = class_tokens_mask.sum().unsqueeze(0) batch_size, max_num_inputs = id_embeds.shape[:2] # seq_length: 77 seq_length = prompt_embeds.shape[1] # flat_id_embeds shape: [b*max_num_inputs, 1, 2048] flat_id_embeds = id_embeds.view( -1, id_embeds.shape[-2], id_embeds.shape[-1] ) # valid_id_mask [b*max_num_inputs] valid_id_mask = ( torch.arange(max_num_inputs, device=flat_id_embeds.device)[None, :] < num_inputs[:, None] ) valid_id_embeds = flat_id_embeds[valid_id_mask.flatten()] prompt_embeds = prompt_embeds.view(-1, prompt_embeds.shape[-1]) class_tokens_mask = class_tokens_mask.view(-1) valid_id_embeds = valid_id_embeds.view(-1, valid_id_embeds.shape[-1]) # slice out the image token embeddings image_token_embeds = prompt_embeds[class_tokens_mask] stacked_id_embeds = self.fuse_fn(image_token_embeds, valid_id_embeds) assert class_tokens_mask.sum() == stacked_id_embeds.shape[0], f"{class_tokens_mask.sum()} != {stacked_id_embeds.shape[0]}" prompt_embeds.masked_scatter_(class_tokens_mask[:, None], stacked_id_embeds.to(prompt_embeds.dtype)) updated_prompt_embeds = prompt_embeds.view(batch_size, seq_length, -1) return updated_prompt_embeds class PhotoMakerIDEncoder(CLIPVisionModelWithProjection): def __init__(self): super().__init__(CLIPVisionConfig(**VISION_CONFIG_DICT)) self.visual_projection_2 = nn.Linear(1024, 1280, bias=False) self.fuse_module = FuseModule(2048) def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask): b, num_inputs, c, h, w = id_pixel_values.shape id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w) shared_id_embeds = self.vision_model(id_pixel_values)[1] id_embeds = self.visual_projection(shared_id_embeds) id_embeds_2 = self.visual_projection_2(shared_id_embeds) id_embeds = id_embeds.view(b, num_inputs, 1, -1) id_embeds_2 = id_embeds_2.view(b, num_inputs, 1, -1) id_embeds = torch.cat((id_embeds, id_embeds_2), dim=-1) updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask) return updated_prompt_embeds class PhotoMakerStableDiffusionXLPipeline(StableDiffusionXLPipeline): @validate_hf_hub_args def load_photomaker_adapter( self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], weight_name: str, subfolder: str = '', trigger_word: str = 'img', **kwargs, ): # Load the main state dict first. cache_dir = kwargs.pop("cache_dir", None) force_download = kwargs.pop("force_download", False) resume_download = kwargs.pop("resume_download", False) proxies = kwargs.pop("proxies", None) local_files_only = kwargs.pop("local_files_only", None) token = kwargs.pop("token", None) revision = kwargs.pop("revision", None) user_agent = { "file_type": "attn_procs_weights", "framework": "pytorch", } if not isinstance(pretrained_model_name_or_path_or_dict, dict): model_file = _get_model_file( pretrained_model_name_or_path_or_dict, weights_name=weight_name, cache_dir=cache_dir, force_download=force_download, resume_download=resume_download, proxies=proxies, local_files_only=local_files_only, token=token, revision=revision, subfolder=subfolder, user_agent=user_agent, ) if weight_name.endswith(".safetensors"): state_dict = {"id_encoder": {}, "lora_weights": {}} with safe_open(model_file, framework="pt", device="cpu") as f: for key in f.keys(): if key.startswith("id_encoder."): state_dict["id_encoder"][key.replace("id_encoder.", "")] = f.get_tensor(key) elif key.startswith("lora_weights."): state_dict["lora_weights"][key.replace("lora_weights.", "")] = f.get_tensor(key) else: state_dict = torch.load(model_file, map_location="cpu") else: state_dict = pretrained_model_name_or_path_or_dict keys = list(state_dict.keys()) if keys != ["id_encoder", "lora_weights"]: raise ValueError("Required keys are (`id_encoder` and `lora_weights`) missing from the state dict.") self.trigger_word = trigger_word # load finetuned CLIP image encoder and fuse module here if it has not been registered to the pipeline yet id_encoder = PhotoMakerIDEncoder() id_encoder.load_state_dict(state_dict["id_encoder"], strict=True) id_encoder = id_encoder.to(self.device, dtype=self.unet.dtype) self.id_encoder = id_encoder self.id_image_processor = CLIPImageProcessor() # load lora into models self.load_lora_weights(state_dict["lora_weights"], adapter_name="photomaker") # Add trigger word token if self.tokenizer is not None: self.tokenizer.add_tokens([self.trigger_word], special_tokens=True) self.tokenizer_2.add_tokens([self.trigger_word], special_tokens=True) def encode_prompt_with_trigger_word( self, prompt: str, prompt_2: Optional[str] = None, num_id_images: int = 1, device: Optional[torch.device] = None, prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.FloatTensor] = None, class_tokens_mask: Optional[torch.LongTensor] = None, ): device = device or self._execution_device """ if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] """ # Find the token id of the trigger word image_token_id = self.tokenizer_2.convert_tokens_to_ids(self.trigger_word) # Define tokenizers and text encoders tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] text_encoders = ( [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] ) if prompt_embeds is None: prompt_2 = prompt_2 or prompt prompt_embeds_list = [] prompts = [prompt, prompt_2] for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): input_ids = tokenizer.encode(prompt) # TODO: batch encode clean_index = 0 clean_input_ids = [] class_token_index = [] # Find out the corrresponding class word token based on the newly added trigger word token for _i, token_id in enumerate(input_ids): if token_id == image_token_id: class_token_index.append(clean_index - 1) else: clean_input_ids.append(token_id) clean_index += 1 if len(class_token_index) != 1: raise ValueError( f"PhotoMaker currently does not support multiple trigger words in a single prompt.\ Trigger word: {self.trigger_word}, Prompt: {prompt}." ) class_token_index = class_token_index[0] # Expand the class word token and corresponding mask class_token = clean_input_ids[class_token_index] clean_input_ids = clean_input_ids[:class_token_index] + [class_token] * num_id_images + \ clean_input_ids[class_token_index+1:] # Truncation or padding max_len = tokenizer.model_max_length if len(clean_input_ids) > max_len: clean_input_ids = clean_input_ids[:max_len] else: clean_input_ids = clean_input_ids + [tokenizer.pad_token_id] * ( max_len - len(clean_input_ids) ) class_tokens_mask = [True if class_token_index <= i < class_token_index+num_id_images else False \ for i in range(len(clean_input_ids))] clean_input_ids = torch.tensor(clean_input_ids, dtype=torch.long).unsqueeze(0) class_tokens_mask = torch.tensor(class_tokens_mask, dtype=torch.bool).unsqueeze(0) prompt_embeds = text_encoder( clean_input_ids.to(device), output_hidden_states=True, ) # We are only ALWAYS interested in the pooled output of the final text encoder pooled_prompt_embeds = prompt_embeds[0] prompt_embeds = prompt_embeds.hidden_states[-2] prompt_embeds_list.append(prompt_embeds) prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) class_tokens_mask = class_tokens_mask.to(device=device) # TODO: ignoring two-prompt case return prompt_embeds, pooled_prompt_embeds, class_tokens_mask @torch.no_grad() def __call__( self, prompt: Union[str, List[str]] = None, prompt_2: Optional[Union[str, List[str]]] = None, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 50, denoising_end: Optional[float] = None, guidance_scale: float = 5.0, negative_prompt: Optional[Union[str, List[str]]] = None, negative_prompt_2: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.FloatTensor] = None, negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, cross_attention_kwargs: Optional[Dict[str, Any]] = None, guidance_rescale: float = 0.0, original_size: Optional[Tuple[int, int]] = None, crops_coords_top_left: Tuple[int, int] = (0, 0), target_size: Optional[Tuple[int, int]] = None, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, # Added parameters (for PhotoMaker) input_id_images: PipelineImageInput = None, start_merge_step: int = 0, # TODO: change to `style_strength_ratio` in the future class_tokens_mask: Optional[torch.LongTensor] = None, prompt_embeds_text_only: Optional[torch.FloatTensor] = None, pooled_prompt_embeds_text_only: Optional[torch.FloatTensor] = None, ): # 0. Default height and width to unet height = height or self.unet.config.sample_size * self.vae_scale_factor width = width or self.unet.config.sample_size * self.vae_scale_factor original_size = original_size or (height, width) target_size = target_size or (height, width) # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, prompt_2, height, width, callback_steps, negative_prompt, negative_prompt_2, prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, ) # if prompt_embeds is not None and class_tokens_mask is None: raise ValueError( "If `prompt_embeds` are provided, `class_tokens_mask` also have to be passed. Make sure to generate `class_tokens_mask` from the same tokenizer that was used to generate `prompt_embeds`." ) # check the input id images if input_id_images is None: raise ValueError( "Provide `input_id_images`. Cannot leave `input_id_images` undefined for PhotoMaker pipeline." ) if not isinstance(input_id_images, list): input_id_images = [input_id_images] # 2. Define call parameters if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] device = self._execution_device # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` # corresponds to doing no classifier free guidance. do_classifier_free_guidance = guidance_scale > 1.0 assert do_classifier_free_guidance # 3. Encode input prompt num_id_images = len(input_id_images) ( prompt_embeds, pooled_prompt_embeds, class_tokens_mask, ) = self.encode_prompt_with_trigger_word( prompt=prompt, prompt_2=prompt_2, device=device, num_id_images=num_id_images, prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, class_tokens_mask=class_tokens_mask, ) # 4. Encode input prompt without the trigger word for delayed conditioning prompt_text_only = prompt.replace(" "+self.trigger_word, "") # sensitive to white space ( prompt_embeds_text_only, negative_prompt_embeds, pooled_prompt_embeds_text_only, # TODO: replace the pooled_prompt_embeds with text only prompt negative_pooled_prompt_embeds, ) = self.encode_prompt( prompt=prompt_text_only, prompt_2=prompt_2, device=device, num_images_per_prompt=num_images_per_prompt, do_classifier_free_guidance=do_classifier_free_guidance, negative_prompt=negative_prompt, negative_prompt_2=negative_prompt_2, prompt_embeds=prompt_embeds_text_only, negative_prompt_embeds=negative_prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds_text_only, negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, ) # 5. Prepare the input ID images dtype = next(self.id_encoder.parameters()).dtype if not isinstance(input_id_images[0], torch.Tensor): id_pixel_values = self.id_image_processor(input_id_images, return_tensors="pt").pixel_values id_pixel_values = id_pixel_values.unsqueeze(0).to(device=device, dtype=dtype) # TODO: multiple prompts # 6. Get the update text embedding with the stacked ID embedding prompt_embeds = self.id_encoder(id_pixel_values, prompt_embeds, class_tokens_mask) bs_embed, seq_len, _ = prompt_embeds.shape # duplicate text embeddings for each generation per prompt, using mps friendly method prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( bs_embed * num_images_per_prompt, -1 ) # 7. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps # 8. Prepare latent variables num_channels_latents = self.unet.config.in_channels latents = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, latents, ) # 9. Prepare extra step kwargs. extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) # 10. Prepare added time ids & embeddings if self.text_encoder_2 is None: text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) else: text_encoder_projection_dim = self.text_encoder_2.config.projection_dim add_time_ids = self._get_add_time_ids( original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype, text_encoder_projection_dim=text_encoder_projection_dim, ) add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) # 11. Denoising loop num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): latent_model_input = ( torch.cat([latents] * 2) if do_classifier_free_guidance else latents ) latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) if i <= start_merge_step: current_prompt_embeds = torch.cat( [negative_prompt_embeds, prompt_embeds_text_only], dim=0 ) add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds_text_only], dim=0) else: current_prompt_embeds = torch.cat( [negative_prompt_embeds, prompt_embeds], dim=0 ) add_text_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds], dim=0) # predict the noise residual added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} noise_pred = self.unet( latent_model_input, t, encoder_hidden_states=current_prompt_embeds, cross_attention_kwargs=cross_attention_kwargs, added_cond_kwargs=added_cond_kwargs, return_dict=False, )[0] # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) if do_classifier_free_guidance and guidance_rescale > 0.0: # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] # call the callback, if provided if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): progress_bar.update() if callback is not None and i % callback_steps == 0: callback(i, t, latents) # make sure the VAE is in float32 mode, as it overflows in float16 if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: self.upcast_vae() latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) if output_type != "latent": image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] else: image = latents return StableDiffusionXLPipelineOutput(images=image) # apply watermark if available # if self.watermark is not None: # image = self.watermark.apply_watermark(image) image = self.image_processor.postprocess(image, output_type=output_type) # Offload last model to CPU if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: self.final_offload_hook.offload() if not return_dict: return (image,) return StableDiffusionXLPipelineOutput(images=image)