diff --git "a/pipeline.py" "b/pipeline.py" --- "a/pipeline.py" +++ "b/pipeline.py" @@ -1,52 +1,53 @@ -# source https://github.com/huggingface/diffusers/blob/main/examples/community/lpw_stable_diffusion_xl.py -## ---------------------------------------------------------- -# A SDXL pipeline can take unlimited weighted prompt -# -# Author: Andrew Zhu -# Github: https://github.com/xhinker -# Medium: https://medium.com/@xhinker -## ----------------------------------------------------------- - +# source https://github.com/huggingface/diffusers/blob/main/examples/community/lpw_stable_diffusion.py import inspect -import os -from typing import Any, Callable, Dict, List, Optional, Tuple, Union +import re +from typing import Any, Callable, Dict, List, Optional, Union +import numpy as np +import PIL.Image import torch -from PIL import Image -from transformers import ( - CLIPImageProcessor, - CLIPTextModel, - CLIPTextModelWithProjection, - CLIPTokenizer, - CLIPVisionModelWithProjection, -) - -from diffusers import DiffusionPipeline, StableDiffusionXLPipeline -from diffusers.image_processor import PipelineImageInput, VaeImageProcessor -from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin -from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel -from diffusers.models.attention_processor import ( - AttnProcessor2_0, - FusedAttnProcessor2_0, - LoRAAttnProcessor2_0, - LoRAXFormersAttnProcessor, - XFormersAttnProcessor, -) -from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer + +from diffusers import DiffusionPipeline +from diffusers.configuration_utils import FrozenDict +from diffusers.image_processor import VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderKL, UNet2DConditionModel +from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker from diffusers.schedulers import KarrasDiffusionSchedulers from diffusers.utils import ( + PIL_INTERPOLATION, deprecate, is_accelerate_available, is_accelerate_version, - is_invisible_watermark_available, logging, - replace_example_docstring, ) from diffusers.utils.torch_utils import randn_tensor -if is_invisible_watermark_available(): - from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker +# ------------------------------------------------------------------------------ + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +re_attention = re.compile( + r""" +\\\(| +\\\)| +\\\[| +\\]| +\\\\| +\\| +\(| +\[| +:([+-]?[.\d]+)\)| +\)| +]| +[^\\()\[\]:]+| +: +""", + re.X, +) def parse_prompt_attention(text): @@ -62,7 +63,6 @@ def parse_prompt_attention(text): \\] - literal character ']' \\ - literal character '\' anything else - just text - >>> parse_prompt_attention('normal text') [['normal text', 1.0]] >>> parse_prompt_attention('an (important) word') @@ -84,17 +84,6 @@ def parse_prompt_attention(text): ['sky', 1.4641000000000006], ['.', 1.1]] """ - import re - - re_attention = re.compile( - r""" - \\\(|\\\)|\\\[|\\]|\\\\|\\|\(|\[|:([+-]?[.\d]+)\)| - \)|]|[^\\()\[\]:]+|: - """, - re.X, - ) - - re_break = re.compile(r"\s*\bBREAK\b\s*", re.S) res = [] round_brackets = [] @@ -124,11 +113,7 @@ def parse_prompt_attention(text): elif text == "]" and len(square_brackets) > 0: multiply_range(square_brackets.pop(), square_bracket_multiplier) else: - parts = re.split(re_break, text) - for i, part in enumerate(parts): - if i > 0: - res.append(["BREAK", -1]) - res.append([part, 1.0]) + res.append([text, 1.0]) for pos in round_brackets: multiply_range(pos, round_bracket_multiplier) @@ -151,538 +136,464 @@ def parse_prompt_attention(text): return res -def get_prompts_tokens_with_weights(clip_tokenizer: CLIPTokenizer, prompt: str): - """ - Get prompt token ids and weights, this function works for both prompt and negative prompt - - Args: - pipe (CLIPTokenizer) - A CLIPTokenizer - prompt (str) - A prompt string with weights - - Returns: - text_tokens (list) - A list contains token ids - text_weight (list) - A list contains the correspodent weight of token ids - - Example: - import torch - from transformers import CLIPTokenizer - - clip_tokenizer = CLIPTokenizer.from_pretrained( - "stablediffusionapi/deliberate-v2" - , subfolder = "tokenizer" - , dtype = torch.float16 - ) +def get_prompts_with_weights(pipe: DiffusionPipeline, prompt: List[str], max_length: int): + r""" + Tokenize a list of prompts and return its tokens with weights of each token. - token_id_list, token_weight_list = get_prompts_tokens_with_weights( - clip_tokenizer = clip_tokenizer - ,prompt = "a (red:1.5) cat"*70 - ) + No padding, starting or ending token is included. """ - texts_and_weights = parse_prompt_attention(prompt) - text_tokens, text_weights = [], [] - for word, weight in texts_and_weights: - # tokenize and discard the starting and the ending token - token = clip_tokenizer(word, truncation=False).input_ids[1:-1] # so that tokenize whatever length prompt - # the returned token is a 1d list: [320, 1125, 539, 320] - - # merge the new tokens to the all tokens holder: text_tokens - text_tokens = [*text_tokens, *token] - - # each token chunk will come with one weight, like ['red cat', 2.0] - # need to expand weight for each token. - chunk_weights = [weight] * len(token) - - # append the weight back to the weight holder: text_weights - text_weights = [*text_weights, *chunk_weights] - return text_tokens, text_weights + tokens = [] + weights = [] + truncated = False + for text in prompt: + texts_and_weights = parse_prompt_attention(text) + text_token = [] + text_weight = [] + for word, weight in texts_and_weights: + # tokenize and discard the starting and the ending token + token = pipe.tokenizer(word).input_ids[1:-1] + text_token += token + # copy the weight by length of token + text_weight += [weight] * len(token) + # stop if the text is too long (longer than truncation limit) + if len(text_token) > max_length: + truncated = True + break + # truncate + if len(text_token) > max_length: + truncated = True + text_token = text_token[:max_length] + text_weight = text_weight[:max_length] + tokens.append(text_token) + weights.append(text_weight) + if truncated: + logger.warning("Prompt was truncated. Try to shorten the prompt or increase max_embeddings_multiples") + return tokens, weights + + +def pad_tokens_and_weights(tokens, weights, max_length, bos, eos, pad, no_boseos_middle=True, chunk_length=77): + r""" + Pad the tokens (with starting and ending tokens) and weights (with 1.0) to max_length. + """ + max_embeddings_multiples = (max_length - 2) // (chunk_length - 2) + weights_length = max_length if no_boseos_middle else max_embeddings_multiples * chunk_length + for i in range(len(tokens)): + tokens[i] = [bos] + tokens[i] + [pad] * (max_length - 1 - len(tokens[i]) - 1) + [eos] + if no_boseos_middle: + weights[i] = [1.0] + weights[i] + [1.0] * (max_length - 1 - len(weights[i])) + else: + w = [] + if len(weights[i]) == 0: + w = [1.0] * weights_length + else: + for j in range(max_embeddings_multiples): + w.append(1.0) # weight for starting token in this chunk + w += weights[i][j * (chunk_length - 2) : min(len(weights[i]), (j + 1) * (chunk_length - 2))] + w.append(1.0) # weight for ending token in this chunk + w += [1.0] * (weights_length - len(w)) + weights[i] = w[:] + return tokens, weights -def group_tokens_and_weights(token_ids: list, weights: list, pad_last_block=False): - """ - Produce tokens and weights in groups and pad the missing tokens - Args: - token_ids (list) - The token ids from tokenizer - weights (list) - The weights list from function get_prompts_tokens_with_weights - pad_last_block (bool) - Control if fill the last token list to 75 tokens with eos - Returns: - new_token_ids (2d list) - new_weights (2d list) - - Example: - token_groups,weight_groups = group_tokens_and_weights( - token_ids = token_id_list - , weights = token_weight_list - ) - """ - bos, eos = 49406, 49407 - - # this will be a 2d list - new_token_ids = [] - new_weights = [] - while len(token_ids) >= 75: - # get the first 75 tokens - head_75_tokens = [token_ids.pop(0) for _ in range(75)] - head_75_weights = [weights.pop(0) for _ in range(75)] - - # extract token ids and weights - temp_77_token_ids = [bos] + head_75_tokens + [eos] - temp_77_weights = [1.0] + head_75_weights + [1.0] - - # add 77 token and weights chunk to the holder list - new_token_ids.append(temp_77_token_ids) - new_weights.append(temp_77_weights) - - # padding the left - if len(token_ids) > 0: - padding_len = 75 - len(token_ids) if pad_last_block else 0 - - temp_77_token_ids = [bos] + token_ids + [eos] * padding_len + [eos] - new_token_ids.append(temp_77_token_ids) - - temp_77_weights = [1.0] + weights + [1.0] * padding_len + [1.0] - new_weights.append(temp_77_weights) - - return new_token_ids, new_weights - - -def get_weighted_text_embeddings_sdxl( - pipe: StableDiffusionXLPipeline, - prompt: str = "", - prompt_2: str = None, - neg_prompt: str = "", - neg_prompt_2: str = None, - num_images_per_prompt: int = 1, - device: Optional[torch.device] = None, - clip_skip: Optional[int] = None, +def get_unweighted_text_embeddings( + pipe: DiffusionPipeline, + text_input: torch.Tensor, + chunk_length: int, + no_boseos_middle: Optional[bool] = True, ): """ - This function can process long prompt with weights, no length limitation - for Stable Diffusion XL - - Args: - pipe (StableDiffusionPipeline) - prompt (str) - prompt_2 (str) - neg_prompt (str) - neg_prompt_2 (str) - num_images_per_prompt (int) - device (torch.device) - clip_skip (int) - Returns: - prompt_embeds (torch.Tensor) - neg_prompt_embeds (torch.Tensor) + When the length of tokens is a multiple of the capacity of the text encoder, + it should be split into chunks and sent to the text encoder individually. """ - device = device or pipe._execution_device - - if prompt_2: - prompt = f"{prompt} {prompt_2}" - - if neg_prompt_2: - neg_prompt = f"{neg_prompt} {neg_prompt_2}" - - prompt_t1 = prompt_t2 = prompt - neg_prompt_t1 = neg_prompt_t2 = neg_prompt - - if isinstance(pipe, TextualInversionLoaderMixin): - prompt_t1 = pipe.maybe_convert_prompt(prompt_t1, pipe.tokenizer) - neg_prompt_t1 = pipe.maybe_convert_prompt(neg_prompt_t1, pipe.tokenizer) - prompt_t2 = pipe.maybe_convert_prompt(prompt_t2, pipe.tokenizer_2) - neg_prompt_t2 = pipe.maybe_convert_prompt(neg_prompt_t2, pipe.tokenizer_2) - - eos = pipe.tokenizer.eos_token_id - - # tokenizer 1 - prompt_tokens, prompt_weights = get_prompts_tokens_with_weights(pipe.tokenizer, prompt_t1) - neg_prompt_tokens, neg_prompt_weights = get_prompts_tokens_with_weights(pipe.tokenizer, neg_prompt_t1) - - # tokenizer 2 - prompt_tokens_2, prompt_weights_2 = get_prompts_tokens_with_weights(pipe.tokenizer_2, prompt_t2) - neg_prompt_tokens_2, neg_prompt_weights_2 = get_prompts_tokens_with_weights(pipe.tokenizer_2, neg_prompt_t2) - - # padding the shorter one for prompt set 1 - prompt_token_len = len(prompt_tokens) - neg_prompt_token_len = len(neg_prompt_tokens) + max_embeddings_multiples = (text_input.shape[1] - 2) // (chunk_length - 2) + if max_embeddings_multiples > 1: + text_embeddings = [] + for i in range(max_embeddings_multiples): + # extract the i-th chunk + text_input_chunk = text_input[:, i * (chunk_length - 2) : (i + 1) * (chunk_length - 2) + 2].clone() + + # cover the head and the tail by the starting and the ending tokens + text_input_chunk[:, 0] = text_input[0, 0] + text_input_chunk[:, -1] = text_input[0, -1] + text_embedding = pipe.text_encoder(text_input_chunk)[0] + + if no_boseos_middle: + if i == 0: + # discard the ending token + text_embedding = text_embedding[:, :-1] + elif i == max_embeddings_multiples - 1: + # discard the starting token + text_embedding = text_embedding[:, 1:] + else: + # discard both starting and ending tokens + text_embedding = text_embedding[:, 1:-1] - if prompt_token_len > neg_prompt_token_len: - # padding the neg_prompt with eos token - neg_prompt_tokens = neg_prompt_tokens + [eos] * abs(prompt_token_len - neg_prompt_token_len) - neg_prompt_weights = neg_prompt_weights + [1.0] * abs(prompt_token_len - neg_prompt_token_len) - else: - # padding the prompt - prompt_tokens = prompt_tokens + [eos] * abs(prompt_token_len - neg_prompt_token_len) - prompt_weights = prompt_weights + [1.0] * abs(prompt_token_len - neg_prompt_token_len) - - # padding the shorter one for token set 2 - prompt_token_len_2 = len(prompt_tokens_2) - neg_prompt_token_len_2 = len(neg_prompt_tokens_2) - - if prompt_token_len_2 > neg_prompt_token_len_2: - # padding the neg_prompt with eos token - neg_prompt_tokens_2 = neg_prompt_tokens_2 + [eos] * abs(prompt_token_len_2 - neg_prompt_token_len_2) - neg_prompt_weights_2 = neg_prompt_weights_2 + [1.0] * abs(prompt_token_len_2 - neg_prompt_token_len_2) + text_embeddings.append(text_embedding) + text_embeddings = torch.concat(text_embeddings, axis=1) else: - # padding the prompt - prompt_tokens_2 = prompt_tokens_2 + [eos] * abs(prompt_token_len_2 - neg_prompt_token_len_2) - prompt_weights_2 = prompt_weights + [1.0] * abs(prompt_token_len_2 - neg_prompt_token_len_2) - - embeds = [] - neg_embeds = [] + text_embeddings = pipe.text_encoder(text_input)[0] + return text_embeddings + + +def get_weighted_text_embeddings( + pipe: DiffusionPipeline, + prompt: Union[str, List[str]], + uncond_prompt: Optional[Union[str, List[str]]] = None, + max_embeddings_multiples: Optional[int] = 3, + no_boseos_middle: Optional[bool] = False, + skip_parsing: Optional[bool] = False, + skip_weighting: Optional[bool] = False, +): + r""" + Prompts can be assigned with local weights using brackets. For example, + prompt 'A (very beautiful) masterpiece' highlights the words 'very beautiful', + and the embedding tokens corresponding to the words get multiplied by a constant, 1.1. - prompt_token_groups, prompt_weight_groups = group_tokens_and_weights(prompt_tokens.copy(), prompt_weights.copy()) + Also, to regularize of the embedding, the weighted embedding would be scaled to preserve the original mean. - neg_prompt_token_groups, neg_prompt_weight_groups = group_tokens_and_weights( - neg_prompt_tokens.copy(), neg_prompt_weights.copy() - ) + Args: + pipe (`DiffusionPipeline`): + Pipe to provide access to the tokenizer and the text encoder. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + uncond_prompt (`str` or `List[str]`): + The unconditional prompt or prompts for guide the image generation. If unconditional prompt + is provided, the embeddings of prompt and uncond_prompt are concatenated. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + no_boseos_middle (`bool`, *optional*, defaults to `False`): + If the length of text token is multiples of the capacity of text encoder, whether reserve the starting and + ending token in each of the chunk in the middle. + skip_parsing (`bool`, *optional*, defaults to `False`): + Skip the parsing of brackets. + skip_weighting (`bool`, *optional*, defaults to `False`): + Skip the weighting. When the parsing is skipped, it is forced True. + """ + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 + if isinstance(prompt, str): + prompt = [prompt] + + if not skip_parsing: + prompt_tokens, prompt_weights = get_prompts_with_weights(pipe, prompt, max_length - 2) + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens, uncond_weights = get_prompts_with_weights(pipe, uncond_prompt, max_length - 2) + else: + prompt_tokens = [ + token[1:-1] for token in pipe.tokenizer(prompt, max_length=max_length, truncation=True).input_ids + ] + prompt_weights = [[1.0] * len(token) for token in prompt_tokens] + if uncond_prompt is not None: + if isinstance(uncond_prompt, str): + uncond_prompt = [uncond_prompt] + uncond_tokens = [ + token[1:-1] + for token in pipe.tokenizer(uncond_prompt, max_length=max_length, truncation=True).input_ids + ] + uncond_weights = [[1.0] * len(token) for token in uncond_tokens] - prompt_token_groups_2, prompt_weight_groups_2 = group_tokens_and_weights( - prompt_tokens_2.copy(), prompt_weights_2.copy() - ) + # round up the longest length of tokens to a multiple of (model_max_length - 2) + max_length = max([len(token) for token in prompt_tokens]) + if uncond_prompt is not None: + max_length = max(max_length, max([len(token) for token in uncond_tokens])) - neg_prompt_token_groups_2, neg_prompt_weight_groups_2 = group_tokens_and_weights( - neg_prompt_tokens_2.copy(), neg_prompt_weights_2.copy() + max_embeddings_multiples = min( + max_embeddings_multiples, + (max_length - 1) // (pipe.tokenizer.model_max_length - 2) + 1, ) + max_embeddings_multiples = max(1, max_embeddings_multiples) + max_length = (pipe.tokenizer.model_max_length - 2) * max_embeddings_multiples + 2 - # get prompt embeddings one by one is not working. - for i in range(len(prompt_token_groups)): - # get positive prompt embeddings with weights - token_tensor = torch.tensor([prompt_token_groups[i]], dtype=torch.long, device=device) - weight_tensor = torch.tensor(prompt_weight_groups[i], dtype=torch.float16, device=device) - - token_tensor_2 = torch.tensor([prompt_token_groups_2[i]], dtype=torch.long, device=device) - - # use first text encoder - prompt_embeds_1 = pipe.text_encoder(token_tensor.to(device), output_hidden_states=True) - - # use second text encoder - prompt_embeds_2 = pipe.text_encoder_2(token_tensor_2.to(device), output_hidden_states=True) - pooled_prompt_embeds = prompt_embeds_2[0] - - if clip_skip is None: - prompt_embeds_1_hidden_states = prompt_embeds_1.hidden_states[-2] - prompt_embeds_2_hidden_states = prompt_embeds_2.hidden_states[-2] - else: - # "2" because SDXL always indexes from the penultimate layer. - prompt_embeds_1_hidden_states = prompt_embeds_1.hidden_states[-(clip_skip + 2)] - prompt_embeds_2_hidden_states = prompt_embeds_2.hidden_states[-(clip_skip + 2)] - - prompt_embeds_list = [prompt_embeds_1_hidden_states, prompt_embeds_2_hidden_states] - token_embedding = torch.concat(prompt_embeds_list, dim=-1).squeeze(0) - - for j in range(len(weight_tensor)): - if weight_tensor[j] != 1.0: - token_embedding[j] = ( - token_embedding[-1] + (token_embedding[j] - token_embedding[-1]) * weight_tensor[j] - ) - - token_embedding = token_embedding.unsqueeze(0) - embeds.append(token_embedding) - - # get negative prompt embeddings with weights - neg_token_tensor = torch.tensor([neg_prompt_token_groups[i]], dtype=torch.long, device=device) - neg_token_tensor_2 = torch.tensor([neg_prompt_token_groups_2[i]], dtype=torch.long, device=device) - neg_weight_tensor = torch.tensor(neg_prompt_weight_groups[i], dtype=torch.float16, device=device) - - # use first text encoder - neg_prompt_embeds_1 = pipe.text_encoder(neg_token_tensor.to(device), output_hidden_states=True) - neg_prompt_embeds_1_hidden_states = neg_prompt_embeds_1.hidden_states[-2] - - # use second text encoder - neg_prompt_embeds_2 = pipe.text_encoder_2(neg_token_tensor_2.to(device), output_hidden_states=True) - neg_prompt_embeds_2_hidden_states = neg_prompt_embeds_2.hidden_states[-2] - negative_pooled_prompt_embeds = neg_prompt_embeds_2[0] - - neg_prompt_embeds_list = [neg_prompt_embeds_1_hidden_states, neg_prompt_embeds_2_hidden_states] - neg_token_embedding = torch.concat(neg_prompt_embeds_list, dim=-1).squeeze(0) - - for z in range(len(neg_weight_tensor)): - if neg_weight_tensor[z] != 1.0: - neg_token_embedding[z] = ( - neg_token_embedding[-1] + (neg_token_embedding[z] - neg_token_embedding[-1]) * neg_weight_tensor[z] - ) - - neg_token_embedding = neg_token_embedding.unsqueeze(0) - neg_embeds.append(neg_token_embedding) - - prompt_embeds = torch.cat(embeds, dim=1) - negative_prompt_embeds = torch.cat(neg_embeds, dim=1) - - 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) - - seq_len = negative_prompt_embeds.shape[1] - negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) - negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) - - pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1).view( - bs_embed * num_images_per_prompt, -1 + # pad the length of tokens and weights + bos = pipe.tokenizer.bos_token_id + eos = pipe.tokenizer.eos_token_id + pad = getattr(pipe.tokenizer, "pad_token_id", eos) + prompt_tokens, prompt_weights = pad_tokens_and_weights( + prompt_tokens, + prompt_weights, + max_length, + bos, + eos, + pad, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, ) - negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt, 1).view( - bs_embed * num_images_per_prompt, -1 + prompt_tokens = torch.tensor(prompt_tokens, dtype=torch.long, device=pipe.device) + if uncond_prompt is not None: + uncond_tokens, uncond_weights = pad_tokens_and_weights( + uncond_tokens, + uncond_weights, + max_length, + bos, + eos, + pad, + no_boseos_middle=no_boseos_middle, + chunk_length=pipe.tokenizer.model_max_length, + ) + uncond_tokens = torch.tensor(uncond_tokens, dtype=torch.long, device=pipe.device) + + # get the embeddings + text_embeddings = get_unweighted_text_embeddings( + pipe, + prompt_tokens, + pipe.tokenizer.model_max_length, + no_boseos_middle=no_boseos_middle, ) - - return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds - - -# ------------------------------------------------------------------------------------------------------------------------------- -# reuse the backbone code from StableDiffusionXLPipeline -# ------------------------------------------------------------------------------------------------------------------------------- - -logger = logging.get_logger(__name__) # pylint: disable=invalid-name - -EXAMPLE_DOC_STRING = """ - Examples: - ```py - from diffusers import DiffusionPipeline - import torch - - pipe = DiffusionPipeline.from_pretrained( - "stabilityai/stable-diffusion-xl-base-1.0" - , torch_dtype = torch.float16 - , use_safetensors = True - , variant = "fp16" - , custom_pipeline = "lpw_stable_diffusion_xl", + prompt_weights = torch.tensor(prompt_weights, dtype=text_embeddings.dtype, device=text_embeddings.device) + if uncond_prompt is not None: + uncond_embeddings = get_unweighted_text_embeddings( + pipe, + uncond_tokens, + pipe.tokenizer.model_max_length, + no_boseos_middle=no_boseos_middle, ) + uncond_weights = torch.tensor(uncond_weights, dtype=uncond_embeddings.dtype, device=uncond_embeddings.device) + + # assign weights to the prompts and normalize in the sense of mean + # TODO: should we normalize by chunk or in a whole (current implementation)? + if (not skip_parsing) and (not skip_weighting): + previous_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= prompt_weights.unsqueeze(-1) + current_mean = text_embeddings.float().mean(axis=[-2, -1]).to(text_embeddings.dtype) + text_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + if uncond_prompt is not None: + previous_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= uncond_weights.unsqueeze(-1) + current_mean = uncond_embeddings.float().mean(axis=[-2, -1]).to(uncond_embeddings.dtype) + uncond_embeddings *= (previous_mean / current_mean).unsqueeze(-1).unsqueeze(-1) + + if uncond_prompt is not None: + return text_embeddings, uncond_embeddings + return text_embeddings, None + + +def preprocess_image(image, batch_size): + w, h = image.size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]) + image = np.array(image).astype(np.float32) / 255.0 + image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size) + image = torch.from_numpy(image) + return 2.0 * image - 1.0 + + +def preprocess_mask(mask, batch_size, scale_factor=8): + if not isinstance(mask, torch.FloatTensor): + mask = mask.convert("L") + w, h = mask.size + w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8 + mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"]) + mask = np.array(mask).astype(np.float32) / 255.0 + mask = np.tile(mask, (4, 1, 1)) + mask = np.vstack([mask[None]] * batch_size) + mask = 1 - mask # repaint white, keep black + mask = torch.from_numpy(mask) + return mask - prompt = "a white cat running on the grass"*20 - prompt2 = "play a football"*20 - prompt = f"{prompt},{prompt2}" - neg_prompt = "blur, low quality" - - pipe.to("cuda") - images = pipe( - prompt = prompt - , negative_prompt = neg_prompt - ).images[0] - - pipe.to("cpu") - torch.cuda.empty_cache() - images - ``` -""" - - -# 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 - - -# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents -def retrieve_latents( - encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample" -): - if hasattr(encoder_output, "latent_dist") and sample_mode == "sample": - return encoder_output.latent_dist.sample(generator) - elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax": - return encoder_output.latent_dist.mode() - elif hasattr(encoder_output, "latents"): - return encoder_output.latents else: - raise AttributeError("Could not access latents of provided encoder_output") - - -# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps -def retrieve_timesteps( - scheduler, - num_inference_steps: Optional[int] = None, - device: Optional[Union[str, torch.device]] = None, - timesteps: Optional[List[int]] = None, - **kwargs, -): - """ - Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles - custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. - - Args: - scheduler (`SchedulerMixin`): - The scheduler to get timesteps from. - num_inference_steps (`int`): - The number of diffusion steps used when generating samples with a pre-trained model. If used, - `timesteps` must be `None`. - device (`str` or `torch.device`, *optional*): - The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. - timesteps (`List[int]`, *optional*): - Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default - timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` - must be `None`. - - Returns: - `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the - second element is the number of inference steps. - """ - if timesteps is not None: - accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) - if not accepts_timesteps: + valid_mask_channel_sizes = [1, 3] + # if mask channel is fourth tensor dimension, permute dimensions to pytorch standard (B, C, H, W) + if mask.shape[3] in valid_mask_channel_sizes: + mask = mask.permute(0, 3, 1, 2) + elif mask.shape[1] not in valid_mask_channel_sizes: raise ValueError( - f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" - f" timestep schedules. Please check whether you are using the correct scheduler." + f"Mask channel dimension of size in {valid_mask_channel_sizes} should be second or fourth dimension," + f" but received mask of shape {tuple(mask.shape)}" ) - scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) - timesteps = scheduler.timesteps - num_inference_steps = len(timesteps) - else: - scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) - timesteps = scheduler.timesteps - return timesteps, num_inference_steps + # (potentially) reduce mask channel dimension from 3 to 1 for broadcasting to latent shape + mask = mask.mean(dim=1, keepdim=True) + h, w = mask.shape[-2:] + h, w = (x - x % 8 for x in (h, w)) # resize to integer multiple of 8 + mask = torch.nn.functional.interpolate(mask, (h // scale_factor, w // scale_factor)) + return mask -class SDXLLongPromptWeightingPipeline( - DiffusionPipeline, FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin +class StableDiffusionLongPromptWeightingPipeline( + DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin ): r""" - Pipeline for text-to-image generation using Stable Diffusion XL. + Pipeline for text-to-image generation using Stable Diffusion without tokens length limit, and support parsing + weighting in prompt. - This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods - implemented for all pipelines (downloading, saving, running on a particular device, etc.). - - The pipeline also inherits the following loading methods: - - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files - - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters - - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights - - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights - - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) Args: vae ([`AutoencoderKL`]): Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. text_encoder ([`CLIPTextModel`]): - Frozen text-encoder. Stable Diffusion XL uses the text portion of + Frozen text-encoder. Stable Diffusion uses the text portion of [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. - text_encoder_2 ([` CLIPTextModelWithProjection`]): - Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of - [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), - specifically the - [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) - variant. tokenizer (`CLIPTokenizer`): Tokenizer of class [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). - tokenizer_2 (`CLIPTokenizer`): - Second Tokenizer of class - [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). - unet ([`UNet2DConditionModel`]): - Conditional U-Net architecture to denoise the encoded image latents. + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. scheduler ([`SchedulerMixin`]): A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. - feature_extractor ([`~transformers.CLIPImageProcessor`]): - A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details. + feature_extractor ([`CLIPImageProcessor`]): + Model that extracts features from generated images to be used as inputs for the `safety_checker`. """ - model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" - _optional_components = [ - "tokenizer", - # "tokenizer_2", - "text_encoder", - # "text_encoder_2", - "image_encoder", - "feature_extractor", - ] - _callback_tensor_inputs = [ - "latents", - "prompt_embeds", - "negative_prompt_embeds", - "add_text_embeds", - "add_time_ids", - "negative_pooled_prompt_embeds", - "negative_add_time_ids", - ] + _optional_components = ["safety_checker", "feature_extractor"] def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, - # text_encoder_2: CLIPTextModelWithProjection, tokenizer: CLIPTokenizer, - # tokenizer_2: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: KarrasDiffusionSchedulers, - feature_extractor: Optional[CLIPImageProcessor] = None, - image_encoder: Optional[CLIPVisionModelWithProjection] = None, - force_zeros_for_empty_prompt: bool = True, - add_watermarker: Optional[bool] = None, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + requires_safety_checker: bool = True, ): super().__init__() + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse( + version.parse(unet.config._diffusers_version).base_version + ) < version.parse("0.9.0.dev0") + is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5" + " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) self.register_modules( vae=vae, text_encoder=text_encoder, - text_encoder_2=text_encoder, tokenizer=tokenizer, - tokenizer_2=tokenizer, unet=unet, scheduler=scheduler, + safety_checker=safety_checker, feature_extractor=feature_extractor, - image_encoder=image_encoder, ) - self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) - self.mask_processor = VaeImageProcessor( - vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True + self.register_to_config( + requires_safety_checker=requires_safety_checker, ) - self.default_sample_size = self.unet.config.sample_size - - add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() - if add_watermarker: - self.watermark = StableDiffusionXLWatermarker() - else: - self.watermark = None - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing def enable_vae_slicing(self): r""" - Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to - compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + Enable sliced VAE decoding. + + When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several + steps. This is useful to save some memory and allow larger batch sizes. """ self.vae.enable_slicing() - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing def disable_vae_slicing(self): r""" - Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to computing decoding in one step. """ self.vae.disable_slicing() - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling def enable_vae_tiling(self): r""" - Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to - compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow - processing larger images. + Enable tiled VAE decoding. + + When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in + several steps. This is useful to save a large amount of memory and to allow the processing of larger images. """ self.vae.enable_tiling() - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling def disable_vae_tiling(self): r""" - Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + Disable tiled VAE decoding. If `enable_vae_tiling` was previously invoked, this method will go back to computing decoding in one step. """ self.vae.disable_tiling() + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_sequential_cpu_offload + def enable_sequential_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, + text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a + `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. + Note that offloading happens on a submodule basis. Memory savings are higher than with + `enable_model_cpu_offload`, but performance is lower. + """ + if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"): + from accelerate import cpu_offload + else: + raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher") + + device = torch.device(f"cuda:{gpu_id}") + + if self.device.type != "cpu": + self.to("cpu", silence_dtype_warnings=True) + torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist) + + for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]: + cpu_offload(cpu_offloaded_model, device) + + if self.safety_checker is not None: + cpu_offload(self.safety_checker, execution_device=device, offload_buffers=True) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload def enable_model_cpu_offload(self, gpu_id=0): r""" Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared @@ -701,80 +612,64 @@ class SDXLLongPromptWeightingPipeline( self.to("cpu", silence_dtype_warnings=True) torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist) - model_sequence = ( - [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] - ) - model_sequence.extend([self.unet, self.vae]) - hook = None - for cpu_offloaded_model in model_sequence: + for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]: _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook) + if self.safety_checker is not None: + _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook) + # We'll offload the last model manually. self.final_offload_hook = hook - # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt + @property + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device + def _execution_device(self): + r""" + Returns the device on which the pipeline's models will be executed. After calling + `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module + hooks. + """ + if not hasattr(self.unet, "_hf_hook"): + return self.device + for module in self.unet.modules(): + if ( + hasattr(module, "_hf_hook") + and hasattr(module._hf_hook, "execution_device") + and module._hf_hook.execution_device is not None + ): + return torch.device(module._hf_hook.execution_device) + return self.device + def encode_prompt( self, - prompt: str, - prompt_2: Optional[str] = None, - device: Optional[torch.device] = None, - num_images_per_prompt: int = 1, - do_classifier_free_guidance: bool = True, - negative_prompt: Optional[str] = None, - negative_prompt_2: Optional[str] = None, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + max_embeddings_multiples=10, 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, - lora_scale: Optional[float] = None, ): r""" Encodes the prompt into text encoder hidden states. Args: - prompt (`str` or `List[str]`, *optional*): + prompt (`str` or `list(int)`): prompt to be encoded - prompt_2 (`str` or `List[str]`, *optional*): - The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is - used in both text-encoders device: (`torch.device`): torch device num_images_per_prompt (`int`): number of images that should be generated per prompt do_classifier_free_guidance (`bool`): whether to use classifier free guidance or not - negative_prompt (`str` or `List[str]`, *optional*): - The prompt or prompts not to guide the image generation. If not defined, one has to pass - `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is - less than `1`). - negative_prompt_2 (`str` or `List[str]`, *optional*): - The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and - `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders - prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not - provided, text embeddings will be generated from `prompt` input argument. - negative_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt - weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input - argument. - pooled_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. - If not provided, pooled text embeddings will be generated from `prompt` input argument. - negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt - weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` - input argument. - lora_scale (`float`, *optional*): - A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + negative_prompt (`str` or `List[str]`): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. """ - device = device or self._execution_device - - # set lora scale so that monkey patched LoRA - # function of text encoder can correctly access it - if lora_scale is not None and isinstance(self, LoraLoaderMixin): - self._lora_scale = lora_scale - if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): @@ -782,187 +677,57 @@ class SDXLLongPromptWeightingPipeline( else: batch_size = prompt_embeds.shape[0] - # 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 - # textual inversion: procecss multi-vector tokens if necessary - prompt_embeds_list = [] - prompts = [prompt, prompt_2] - for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): - if isinstance(self, TextualInversionLoaderMixin): - prompt = self.maybe_convert_prompt(prompt, tokenizer) - - text_inputs = tokenizer( - prompt, - padding="max_length", - max_length=tokenizer.model_max_length, - truncation=True, - return_tensors="pt", - ) - - text_input_ids = text_inputs.input_ids - untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids - - if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( - text_input_ids, untruncated_ids - ): - removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) - logger.warning( - "The following part of your input was truncated because CLIP can only handle sequences up to" - f" {tokenizer.model_max_length} tokens: {removed_text}" - ) - - prompt_embeds = text_encoder( - text_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) - - # get unconditional embeddings for classifier free guidance - zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt - if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: - negative_prompt_embeds = torch.zeros_like(prompt_embeds) - negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) - elif do_classifier_free_guidance and negative_prompt_embeds is None: - negative_prompt = negative_prompt or "" - negative_prompt_2 = negative_prompt_2 or negative_prompt - - uncond_tokens: List[str] - if prompt is not None and type(prompt) is not type(negative_prompt): - raise TypeError( - f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" - f" {type(prompt)}." - ) + if negative_prompt_embeds is None: + if negative_prompt is None: + negative_prompt = [""] * batch_size elif isinstance(negative_prompt, str): - uncond_tokens = [negative_prompt, negative_prompt_2] - elif batch_size != len(negative_prompt): + negative_prompt = [negative_prompt] * batch_size + if batch_size != len(negative_prompt): raise ValueError( f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" " the batch size of `prompt`." ) - else: - uncond_tokens = [negative_prompt, negative_prompt_2] - - negative_prompt_embeds_list = [] - for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): - if isinstance(self, TextualInversionLoaderMixin): - negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) - - max_length = prompt_embeds.shape[1] - uncond_input = tokenizer( - negative_prompt, - padding="max_length", - max_length=max_length, - truncation=True, - return_tensors="pt", - ) - - negative_prompt_embeds = text_encoder( - uncond_input.input_ids.to(device), - output_hidden_states=True, - ) - # We are only ALWAYS interested in the pooled output of the final text encoder - negative_pooled_prompt_embeds = negative_prompt_embeds[0] - negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] - - negative_prompt_embeds_list.append(negative_prompt_embeds) - - negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + if prompt_embeds is None or negative_prompt_embeds is None: + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + if do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = self.maybe_convert_prompt(negative_prompt, self.tokenizer) + + prompt_embeds1, negative_prompt_embeds1 = get_weighted_text_embeddings( + pipe=self, + prompt=prompt, + uncond_prompt=negative_prompt if do_classifier_free_guidance else None, + max_embeddings_multiples=max_embeddings_multiples, + ) + if prompt_embeds is None: + prompt_embeds = prompt_embeds1 + if negative_prompt_embeds is None: + negative_prompt_embeds = negative_prompt_embeds1 - prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) 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) if do_classifier_free_guidance: - # duplicate unconditional embeddings for each generation per prompt, using mps friendly method - seq_len = negative_prompt_embeds.shape[1] - negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + bs_embed, seq_len, _ = negative_prompt_embeds.shape negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) - negative_prompt_embeds = negative_prompt_embeds.view(batch_size * 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 - ) - if do_classifier_free_guidance: - negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( - bs_embed * num_images_per_prompt, -1 - ) - - return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image - def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): - dtype = next(self.image_encoder.parameters()).dtype - - if not isinstance(image, torch.Tensor): - image = self.feature_extractor(image, return_tensors="pt").pixel_values - - image = image.to(device=device, dtype=dtype) - if output_hidden_states: - image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] - image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) - uncond_image_enc_hidden_states = self.image_encoder( - torch.zeros_like(image), output_hidden_states=True - ).hidden_states[-2] - uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( - num_images_per_prompt, dim=0 - ) - return image_enc_hidden_states, uncond_image_enc_hidden_states - else: - image_embeds = self.image_encoder(image).image_embeds - image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) - uncond_image_embeds = torch.zeros_like(image_embeds) - - return image_embeds, uncond_image_embeds - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs - def prepare_extra_step_kwargs(self, generator, eta): - # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature - # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. - # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 - # and should be between [0, 1] - - accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) - extra_step_kwargs = {} - if accepts_eta: - extra_step_kwargs["eta"] = eta + negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) - # check if the scheduler accepts generator - accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) - if accepts_generator: - extra_step_kwargs["generator"] = generator - return extra_step_kwargs + return prompt_embeds def check_inputs( self, prompt, - prompt_2, height, width, strength, callback_steps, negative_prompt=None, - negative_prompt_2=None, prompt_embeds=None, negative_prompt_embeds=None, - pooled_prompt_embeds=None, - negative_pooled_prompt_embeds=None, - callback_on_step_end_tensor_inputs=None, ): if height % 8 != 0 or width % 8 != 0: raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") @@ -970,48 +735,31 @@ class SDXLLongPromptWeightingPipeline( if strength < 0 or strength > 1: raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") - if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): raise ValueError( f"`callback_steps` has to be a positive integer but is {callback_steps} of type" f" {type(callback_steps)}." ) - if callback_on_step_end_tensor_inputs is not None and not all( - k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs - ): - raise ValueError( - f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" - ) - if prompt is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) - elif prompt_2 is not None and prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" - " only forward one of the two." - ) elif prompt is None and prompt_embeds is None: raise ValueError( "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." ) elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") - elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): - raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") if negative_prompt is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) - elif negative_prompt_2 is not None and negative_prompt_embeds is not None: - raise ValueError( - f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" - f" {negative_prompt_embeds}. Please make sure to only forward one of the two." - ) if prompt_embeds is not None and negative_prompt_embeds is not None: if prompt_embeds.shape != negative_prompt_embeds.shape: @@ -1021,163 +769,69 @@ class SDXLLongPromptWeightingPipeline( f" {negative_prompt_embeds.shape}." ) - if prompt_embeds is not None and pooled_prompt_embeds is None: - raise ValueError( - "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." - ) - - if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: - raise ValueError( - "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." - ) - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_freeu - def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): - r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497. - - The suffixes after the scaling factors represent the stages where they are being applied. - - Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values - that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. - - Args: - s1 (`float`): - Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to - mitigate "oversmoothing effect" in the enhanced denoising process. - s2 (`float`): - Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to - mitigate "oversmoothing effect" in the enhanced denoising process. - b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. - b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. - """ - if not hasattr(self, "unet"): - raise ValueError("The pipeline must have `unet` for using FreeU.") - self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_freeu - def disable_freeu(self): - """Disables the FreeU mechanism if enabled.""" - self.unet.disable_freeu() - - # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections - def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): - """ - Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, - key, value) are fused. For cross-attention modules, key and value projection matrices are fused. - - - - This API is 🧪 experimental. - - - - Args: - unet (`bool`, defaults to `True`): To apply fusion on the UNet. - vae (`bool`, defaults to `True`): To apply fusion on the VAE. - """ - self.fusing_unet = False - self.fusing_vae = False - - if unet: - self.fusing_unet = True - self.unet.fuse_qkv_projections() - self.unet.set_attn_processor(FusedAttnProcessor2_0()) - - if vae: - if not isinstance(self.vae, AutoencoderKL): - raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") - - self.fusing_vae = True - self.vae.fuse_qkv_projections() - self.vae.set_attn_processor(FusedAttnProcessor2_0()) - - # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections - def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): - """Disable QKV projection fusion if enabled. - - - - This API is 🧪 experimental. - - - - Args: - unet (`bool`, defaults to `True`): To apply fusion on the UNet. - vae (`bool`, defaults to `True`): To apply fusion on the VAE. - - """ - if unet: - if not self.fusing_unet: - logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") - else: - self.unet.unfuse_qkv_projections() - self.fusing_unet = False - - if vae: - if not self.fusing_vae: - logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") - else: - self.vae.unfuse_qkv_projections() - self.fusing_vae = False - - def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): - # get the original timestep using init_timestep - if denoising_start is None: + def get_timesteps(self, num_inference_steps, strength, device, is_text2img): + if is_text2img: + return self.scheduler.timesteps.to(device), num_inference_steps + else: + # get the original timestep using init_timestep init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) - else: - t_start = 0 + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] - timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + return timesteps, num_inference_steps - t_start - # Strength is irrelevant if we directly request a timestep to start at; - # that is, strength is determined by the denoising_start instead. - if denoising_start is not None: - discrete_timestep_cutoff = int( - round( - self.scheduler.config.num_train_timesteps - - (denoising_start * self.scheduler.config.num_train_timesteps) - ) + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is not None: + safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) ) + else: + has_nsfw_concept = None + return image, has_nsfw_concept + + def decode_latents(self, latents): + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents).sample + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image - num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item() - if self.scheduler.order == 2 and num_inference_steps % 2 == 0: - # if the scheduler is a 2nd order scheduler we might have to do +1 - # because `num_inference_steps` might be even given that every timestep - # (except the highest one) is duplicated. If `num_inference_steps` is even it would - # mean that we cut the timesteps in the middle of the denoising step - # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1 - # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler - num_inference_steps = num_inference_steps + 1 + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] - # because t_n+1 >= t_n, we slice the timesteps starting from the end - timesteps = timesteps[-num_inference_steps:] - return timesteps, num_inference_steps + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta - return timesteps, num_inference_steps - t_start + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs def prepare_latents( self, image, - mask, - width, - height, - num_channels_latents, timestep, - batch_size, num_images_per_prompt, + batch_size, + num_channels_latents, + height, + width, dtype, device, - generator=None, - add_noise=True, + generator, latents=None, - is_strength_max=True, - return_noise=False, - return_image_latents=False, ): - batch_size *= num_images_per_prompt - if image is None: + batch_size = batch_size * num_images_per_prompt shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) if isinstance(generator, list) and len(generator) != batch_size: raise ValueError( @@ -1192,397 +846,91 @@ class SDXLLongPromptWeightingPipeline( # scale the initial noise by the standard deviation required by the scheduler latents = latents * self.scheduler.init_noise_sigma - return latents - - elif mask is None: - if not isinstance(image, (torch.Tensor, Image.Image, list)): - raise ValueError( - f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" - ) - - # Offload text encoder if `enable_model_cpu_offload` was enabled - if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: - self.text_encoder_2.to("cpu") - torch.cuda.empty_cache() - - image = image.to(device=device, dtype=dtype) - - if image.shape[1] == 4: - init_latents = image - - else: - # make sure the VAE is in float32 mode, as it overflows in float16 - if self.vae.config.force_upcast: - image = image.float() - self.vae.to(dtype=torch.float32) - - if isinstance(generator, list) and len(generator) != batch_size: - raise ValueError( - f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" - f" size of {batch_size}. Make sure the batch size matches the length of the generators." - ) - - elif isinstance(generator, list): - init_latents = [ - retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) - for i in range(batch_size) - ] - init_latents = torch.cat(init_latents, dim=0) - else: - init_latents = retrieve_latents(self.vae.encode(image), generator=generator) - - if self.vae.config.force_upcast: - self.vae.to(dtype) - - init_latents = init_latents.to(dtype) - init_latents = self.vae.config.scaling_factor * init_latents - - if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: - # expand init_latents for batch_size - additional_image_per_prompt = batch_size // init_latents.shape[0] - init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) - elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: - raise ValueError( - f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." - ) - else: - init_latents = torch.cat([init_latents], dim=0) - - if add_noise: - shape = init_latents.shape - noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) - # get latents - init_latents = self.scheduler.add_noise(init_latents, noise, timestep) - - latents = init_latents - return latents - + return latents, None, None else: - shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) - if isinstance(generator, list) and len(generator) != batch_size: - raise ValueError( - f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" - f" size of {batch_size}. Make sure the batch size matches the length of the generators." - ) - - if (image is None or timestep is None) and not is_strength_max: - raise ValueError( - "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise." - "However, either the image or the noise timestep has not been provided." - ) - - if image.shape[1] == 4: - image_latents = image.to(device=device, dtype=dtype) - image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) - elif return_image_latents or (latents is None and not is_strength_max): - image = image.to(device=device, dtype=dtype) - image_latents = self._encode_vae_image(image=image, generator=generator) - image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1) - - if latents is None and add_noise: - noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) - # if strength is 1. then initialise the latents to noise, else initial to image + noise - latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep) - # if pure noise then scale the initial latents by the Scheduler's init sigma - latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents - elif add_noise: - noise = latents.to(device) - latents = noise * self.scheduler.init_noise_sigma - else: - noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) - latents = image_latents.to(device) - - outputs = (latents,) - - if return_noise: - outputs += (noise,) - - if return_image_latents: - outputs += (image_latents,) - - return outputs - - def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator): - dtype = image.dtype - if self.vae.config.force_upcast: - image = image.float() - self.vae.to(dtype=torch.float32) - - if isinstance(generator, list): - image_latents = [ - retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i]) - for i in range(image.shape[0]) - ] - image_latents = torch.cat(image_latents, dim=0) - else: - image_latents = retrieve_latents(self.vae.encode(image), generator=generator) - - if self.vae.config.force_upcast: - self.vae.to(dtype) - - image_latents = image_latents.to(dtype) - image_latents = self.vae.config.scaling_factor * image_latents - - return image_latents - - def prepare_mask_latents( - self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance - ): - # resize the mask to latents shape as we concatenate the mask to the latents - # we do that before converting to dtype to avoid breaking in case we're using cpu_offload - # and half precision - mask = torch.nn.functional.interpolate( - mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) - ) - mask = mask.to(device=device, dtype=dtype) - - # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method - if mask.shape[0] < batch_size: - if not batch_size % mask.shape[0] == 0: - raise ValueError( - "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" - f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" - " of masks that you pass is divisible by the total requested batch size." - ) - mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) - - mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask - - if masked_image is not None and masked_image.shape[1] == 4: - masked_image_latents = masked_image - else: - masked_image_latents = None - - if masked_image is not None: - if masked_image_latents is None: - masked_image = masked_image.to(device=device, dtype=dtype) - masked_image_latents = self._encode_vae_image(masked_image, generator=generator) - - if masked_image_latents.shape[0] < batch_size: - if not batch_size % masked_image_latents.shape[0] == 0: - raise ValueError( - "The passed images and the required batch size don't match. Images are supposed to be duplicated" - f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." - " Make sure the number of images that you pass is divisible by the total requested batch size." - ) - masked_image_latents = masked_image_latents.repeat( - batch_size // masked_image_latents.shape[0], 1, 1, 1 - ) - - masked_image_latents = ( - torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents - ) - - # aligning device to prevent device errors when concating it with the latent model input - masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) - - return mask, masked_image_latents - - def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype): - add_time_ids = list(original_size + crops_coords_top_left + target_size) - - passed_add_embed_dim = ( - self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim - ) - expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features - - if expected_add_embed_dim != passed_add_embed_dim: - raise ValueError( - f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." - ) - - add_time_ids = torch.tensor([add_time_ids], dtype=dtype) - return add_time_ids - - # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae - def upcast_vae(self): - dtype = self.vae.dtype - self.vae.to(dtype=torch.float32) - use_torch_2_0_or_xformers = isinstance( - self.vae.decoder.mid_block.attentions[0].processor, - ( - AttnProcessor2_0, - XFormersAttnProcessor, - LoRAXFormersAttnProcessor, - LoRAAttnProcessor2_0, - ), - ) - # if xformers or torch_2_0 is used attention block does not need - # to be in float32 which can save lots of memory - if use_torch_2_0_or_xformers: - self.vae.post_quant_conv.to(dtype) - self.vae.decoder.conv_in.to(dtype) - self.vae.decoder.mid_block.to(dtype) - - # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding - def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32): - """ - See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 - - Args: - timesteps (`torch.Tensor`): - generate embedding vectors at these timesteps - embedding_dim (`int`, *optional*, defaults to 512): - dimension of the embeddings to generate - dtype: - data type of the generated embeddings - - Returns: - `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` - """ - assert len(w.shape) == 1 - w = w * 1000.0 - - half_dim = embedding_dim // 2 - emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) - emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) - emb = w.to(dtype)[:, None] * emb[None, :] - emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) - if embedding_dim % 2 == 1: # zero pad - emb = torch.nn.functional.pad(emb, (0, 1)) - assert emb.shape == (w.shape[0], embedding_dim) - return emb - - @property - def guidance_scale(self): - return self._guidance_scale - - @property - def guidance_rescale(self): - return self._guidance_rescale - - @property - def clip_skip(self): - return self._clip_skip - - # 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. - @property - def do_classifier_free_guidance(self): - return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None - - @property - def cross_attention_kwargs(self): - return self._cross_attention_kwargs + image = image.to(device=self.device, dtype=dtype) + init_latent_dist = self.vae.encode(image).latent_dist + init_latents = init_latent_dist.sample(generator=generator) + init_latents = self.vae.config.scaling_factor * init_latents - @property - def denoising_end(self): - return self._denoising_end + # Expand init_latents for batch_size and num_images_per_prompt + init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0) + init_latents_orig = init_latents - @property - def denoising_start(self): - return self._denoising_start - - @property - def num_timesteps(self): - return self._num_timesteps + # add noise to latents using the timesteps + noise = randn_tensor(init_latents.shape, generator=generator, device=self.device, dtype=dtype) + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + latents = init_latents + return latents, init_latents_orig, noise @torch.no_grad() - @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, - prompt: str = None, - prompt_2: Optional[str] = None, - image: Optional[PipelineImageInput] = None, - mask_image: Optional[PipelineImageInput] = None, - masked_image_latents: Optional[torch.FloatTensor] = None, - height: Optional[int] = None, - width: Optional[int] = None, - strength: float = 0.8, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + image: Union[torch.FloatTensor, PIL.Image.Image] = None, + mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None, + height: int = 512, + width: int = 512, num_inference_steps: int = 50, - timesteps: List[int] = None, - denoising_start: Optional[float] = None, - denoising_end: Optional[float] = None, - guidance_scale: float = 5.0, - negative_prompt: Optional[str] = None, - negative_prompt_2: Optional[str] = None, + guidance_scale: float = 7.5, + strength: float = 0.8, num_images_per_prompt: Optional[int] = 1, + add_predicted_noise: Optional[bool] = False, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, - ip_adapter_image: Optional[PipelineImageInput] = 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, + max_embeddings_multiples: Optional[int] = 3, output_type: Optional[str] = "pil", return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, 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, - clip_skip: Optional[int] = None, - callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, - callback_on_step_end_tensor_inputs: List[str] = ["latents"], - **kwargs, ): r""" Function invoked when calling the pipeline for generation. Args: - prompt (`str`): - The prompt to guide the image generation. If not defined, one has to pass `prompt_embeds`. - instead. - prompt_2 (`str`): - The prompt to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is - used in both text-encoders - image (`PipelineImageInput`, *optional*): + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + image (`torch.FloatTensor` or `PIL.Image.Image`): `Image`, or tensor representing an image batch, that will be used as the starting point for the process. - mask_image (`PipelineImageInput`, *optional*): + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. - height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + height (`int`, *optional*, defaults to 512): The height in pixels of the generated image. - width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + width (`int`, *optional*, defaults to 512): The width in pixels of the generated image. - strength (`float`, *optional*, defaults to 0.8): - Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. - `image` will be used as a starting point, adding more noise to it the larger the `strength`. The - number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added - noise will be maximum and the denoising process will run for the full number of iterations specified in - `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. num_inference_steps (`int`, *optional*, defaults to 50): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. - timesteps (`List[int]`, *optional*): - Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument - in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is - passed will be used. Must be in descending order. - denoising_start (`float`, *optional*): - When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be - bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and - it is assumed that the passed `image` is a partly denoised image. Note that when this is specified, - strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline - is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image - Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). - denoising_end (`float`, *optional*): - When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be - completed before it is intentionally prematurely terminated. As a result, the returned sample will - still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be - denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the - final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline - forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refine Image - Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality). - guidance_scale (`float`, *optional*, defaults to 5.0): + guidance_scale (`float`, *optional*, defaults to 7.5): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. - negative_prompt (`str`): - The prompt not to guide the image generation. If not defined, one has to pass - `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is - less than `1`). - negative_prompt_2 (`str`): - The prompt not to guide the image generation to be sent to `tokenizer_2` and - `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + `image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. + add_predicted_noise (`bool`, *optional*, defaults to True): + Use predicted noise instead of random noise when constructing noisy versions of the original image in + the reverse diffusion process eta (`float`, *optional*, defaults to 0.0): Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to [`schedulers.DDIMScheduler`], will be ignored for others. @@ -1593,8 +941,6 @@ class SDXLLongPromptWeightingPipeline( Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will ge generated by sampling using the supplied random `generator`. - ip_adapter_image: (`PipelineImageInput`, *optional*): - Optional image input to work with IP Adapters. prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. @@ -1602,110 +948,45 @@ class SDXLLongPromptWeightingPipeline( Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. - pooled_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. - If not provided, pooled text embeddings will be generated from `prompt` input argument. - negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): - Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt - weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` - input argument. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generate image. Choose between [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. return_dict (`bool`, *optional*, defaults to `True`): - Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead - of a plain tuple. + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. cross_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - guidance_rescale (`float`, *optional*, defaults to 0.0): - Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are - Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of - [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). - Guidance rescale factor should fix overexposure when using zero terminal SNR. - original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): - If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. - `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as - explained in section 2.2 of - [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). - crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): - `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position - `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting - `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of - [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). - target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): - For most cases, `target_size` should be set to the desired height and width of the generated image. If - not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in - section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). - clip_skip (`int`, *optional*): - Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that - the output of the pre-final layer will be used for computing the prompt embeddings. - callback_on_step_end (`Callable`, *optional*): - A function that calls at the end of each denoising steps during the inference. The function is called - with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, - callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by - `callback_on_step_end_tensor_inputs`. - callback_on_step_end_tensor_inputs (`List`, *optional*): - The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list - will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the - `._callback_tensor_inputs` attribute of your pipeine class. - - Examples: Returns: - [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: - [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a - `tuple`. When returning a tuple, the first element is a list with the generated images. + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. """ - - callback = kwargs.pop("callback", None) - callback_steps = kwargs.pop("callback_steps", None) - - if callback is not None: - deprecate( - "callback", - "1.0.0", - "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", - ) - if callback_steps is not None: - deprecate( - "callback_steps", - "1.0.0", - "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", - ) - # 0. Default height and width to unet - height = height or self.default_sample_size * self.vae_scale_factor - width = width or self.default_sample_size * self.vae_scale_factor - - original_size = original_size or (height, width) - target_size = target_size or (height, width) + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor # 1. Check inputs. Raise error if not correct self.check_inputs( - prompt, - prompt_2, - height, - width, - strength, - callback_steps, - negative_prompt, - negative_prompt_2, - prompt_embeds, - negative_prompt_embeds, - pooled_prompt_embeds, - negative_pooled_prompt_embeds, - callback_on_step_end_tensor_inputs, + prompt, height, width, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds ) - self._guidance_scale = guidance_scale - self._guidance_rescale = guidance_rescale - self._clip_skip = clip_skip - self._cross_attention_kwargs = cross_attention_kwargs - self._denoising_end = denoising_end - self._denoising_start = denoising_start - # 2. Define call parameters if prompt is not None and isinstance(prompt, str): batch_size = 1 @@ -1715,606 +996,477 @@ class SDXLLongPromptWeightingPipeline( batch_size = prompt_embeds.shape[0] device = self._execution_device - - if ip_adapter_image is not None: - output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True - image_embeds, negative_image_embeds = self.encode_image( - ip_adapter_image, device, num_images_per_prompt, output_hidden_state - ) - if self.do_classifier_free_guidance: - image_embeds = torch.cat([negative_image_embeds, image_embeds]) + # 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 # 3. Encode input prompt - (self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None) - - negative_prompt = negative_prompt if negative_prompt is not None else "" - - ( - prompt_embeds, - negative_prompt_embeds, - pooled_prompt_embeds, - negative_pooled_prompt_embeds, - ) = get_weighted_text_embeddings_sdxl( - pipe=self, - prompt=prompt, - neg_prompt=negative_prompt, - num_images_per_prompt=num_images_per_prompt, - clip_skip=clip_skip, + prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + max_embeddings_multiples, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, ) dtype = prompt_embeds.dtype - if isinstance(image, Image.Image): - image = self.image_processor.preprocess(image, height=height, width=width) + # 4. Preprocess image and mask + if isinstance(image, PIL.Image.Image): + image = preprocess_image(image, batch_size) if image is not None: image = image.to(device=self.device, dtype=dtype) - - if isinstance(mask_image, Image.Image): - mask = self.mask_processor.preprocess(mask_image, height=height, width=width) - else: - mask = mask_image + if isinstance(mask_image, PIL.Image.Image): + mask_image = preprocess_mask(mask_image, batch_size, self.vae_scale_factor) if mask_image is not None: - mask = mask.to(device=self.device, dtype=dtype) - - if masked_image_latents is not None: - masked_image = masked_image_latents - elif image.shape[1] == 4: - # if image is in latent space, we can't mask it - masked_image = None - else: - masked_image = image * (mask < 0.5) + mask = mask_image.to(device=self.device, dtype=dtype) + mask = torch.cat([mask] * num_images_per_prompt) else: mask = None - # 4. Prepare timesteps - def denoising_value_valid(dnv): - return isinstance(self.denoising_end, float) and 0 < dnv < 1 - - timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) - if image is not None: - timesteps, num_inference_steps = self.get_timesteps( - num_inference_steps, - strength, - device, - denoising_start=self.denoising_start if denoising_value_valid else None, - ) - - # check that number of inference steps is not < 1 - as this doesn't make sense - if num_inference_steps < 1: - raise ValueError( - f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline" - f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline." - ) - + # 5. set timesteps + self.scheduler.set_timesteps(num_inference_steps, device=device) + timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device, image is None) latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) - is_strength_max = strength == 1.0 - add_noise = True if self.denoising_start is None else False - - # 5. Prepare latent variables - num_channels_latents = self.vae.config.latent_channels - num_channels_unet = self.unet.config.in_channels - return_image_latents = num_channels_unet == 4 - latents = self.prepare_latents( - image=image, - mask=mask, - width=width, - height=height, - num_channels_latents=num_channels_unet, - timestep=latent_timestep, - batch_size=batch_size, - num_images_per_prompt=num_images_per_prompt, - dtype=prompt_embeds.dtype, - device=device, - generator=generator, - add_noise=add_noise, - latents=latents, - is_strength_max=is_strength_max, - return_noise=True, - return_image_latents=return_image_latents, + # 6. Prepare latent variables + latents, init_latents_orig, noise = self.prepare_latents( + image, + latent_timestep, + num_images_per_prompt, + batch_size, + self.unet.config.in_channels, + height, + width, + dtype, + device, + generator, + latents, ) - if mask is not None: - if return_image_latents: - latents, noise, image_latents = latents - else: - latents, noise = latents - - # 5.1 Prepare mask latent variables - if mask is not None: - mask, masked_image_latents = self.prepare_mask_latents( - mask=mask, - masked_image=masked_image, - batch_size=batch_size * num_images_per_prompt, - height=height, - width=width, - dtype=prompt_embeds.dtype, - device=device, - generator=generator, - do_classifier_free_guidance=self.do_classifier_free_guidance, - ) - - # Check that sizes of mask, masked image and latents match - if num_channels_unet == 9: - # default case for runwayml/stable-diffusion-inpainting - num_channels_mask = mask.shape[1] - num_channels_masked_image = masked_image_latents.shape[1] - if num_channels_latents + num_channels_mask + num_channels_masked_image != num_channels_unet: - raise ValueError( - f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects" - f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +" - f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}" - f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of" - " `pipeline.unet` or your `mask_image` or `image` input." - ) - elif num_channels_unet != 4: - raise ValueError( - f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}." - ) - - # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) - # 6.1 Add image embeds for IP-Adapter - added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else {} - - height, width = latents.shape[-2:] - height = height * self.vae_scale_factor - width = width * self.vae_scale_factor - - original_size = original_size or (height, width) - target_size = target_size or (height, width) - - # 7. Prepare added time ids & embeddings - add_text_embeds = pooled_prompt_embeds - add_time_ids = self._get_add_time_ids( - original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype - ) - - if self.do_classifier_free_guidance: - prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) - add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) - add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) - - prompt_embeds = prompt_embeds.to(device) - add_text_embeds = add_text_embeds.to(device) - add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) - - num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) - - # 7.1 Apply denoising_end - if ( - self.denoising_end is not None - and self.denoising_start is not None - and denoising_value_valid(self.denoising_end) - and denoising_value_valid(self.denoising_start) - and self.denoising_start >= self.denoising_end - ): - raise ValueError( - f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: " - + f" {self.denoising_end} when using type float." - ) - elif self.denoising_end is not None and denoising_value_valid(self.denoising_end): - discrete_timestep_cutoff = int( - round( - self.scheduler.config.num_train_timesteps - - (self.denoising_end * self.scheduler.config.num_train_timesteps) - ) - ) - num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) - timesteps = timesteps[:num_inference_steps] - - # 8. Optionally get Guidance Scale Embedding - timestep_cond = None - if self.unet.config.time_cond_proj_dim is not None: - guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) - timestep_cond = self.get_guidance_scale_embedding( - guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim - ).to(device=device, dtype=latents.dtype) - - self._num_timesteps = len(timesteps) - - # 9. Denoising loop + # 8. 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): # expand the latents if we are doing classifier free guidance - latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents - + 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 mask is not None and num_channels_unet == 9: - latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) - # predict the noise residual - added_cond_kwargs.update({"text_embeds": add_text_embeds, "time_ids": add_time_ids}) noise_pred = self.unet( latent_model_input, t, encoder_hidden_states=prompt_embeds, - timestep_cond=timestep_cond, - cross_attention_kwargs=self.cross_attention_kwargs, - added_cond_kwargs=added_cond_kwargs, - return_dict=False, - )[0] + cross_attention_kwargs=cross_attention_kwargs, + ).sample # perform guidance - if self.do_classifier_free_guidance: + if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) - noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) - - if self.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) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) # 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] - - if mask is not None and num_channels_unet == 4: - init_latents_proper = image_latents + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample - if self.do_classifier_free_guidance: - init_mask, _ = mask.chunk(2) - else: - init_mask = mask - - if i < len(timesteps) - 1: - noise_timestep = timesteps[i + 1] + if mask is not None: + # masking + if add_predicted_noise: init_latents_proper = self.scheduler.add_noise( - init_latents_proper, noise, torch.tensor([noise_timestep]) + init_latents_orig, noise_pred_uncond, torch.tensor([t]) ) - - latents = (1 - init_mask) * init_latents_proper + init_mask * latents - - if callback_on_step_end is not None: - callback_kwargs = {} - for k in callback_on_step_end_tensor_inputs: - callback_kwargs[k] = locals()[k] - callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) - - latents = callback_outputs.pop("latents", latents) - prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) - negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) - add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds) - negative_pooled_prompt_embeds = callback_outputs.pop( - "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds - ) - add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids) + else: + init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t])) + latents = (init_latents_proper * mask) + (latents * (1 - mask)) # 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: - step_idx = i // getattr(self.scheduler, "order", 1) - callback(step_idx, t, latents) - - if not output_type == "latent": - # make sure the VAE is in float32 mode, as it overflows in float16 - needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast - - if needs_upcasting: - self.upcast_vae() - latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + if i % callback_steps == 0: + if callback is not None: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + if is_cancelled_callback is not None and is_cancelled_callback(): + return None + + if output_type == "latent": + image = latents + has_nsfw_concept = None + elif output_type == "pil": + # 9. Post-processing + image = self.decode_latents(latents) - image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + # 10. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) - # cast back to fp16 if needed - if needs_upcasting: - self.vae.to(dtype=torch.float16) + # 11. Convert to PIL + image = self.numpy_to_pil(image) else: - image = latents - return StableDiffusionXLPipelineOutput(images=image) - - # apply watermark if available - if self.watermark is not None: - image = self.watermark.apply_watermark(image) + # 9. Post-processing + image = self.decode_latents(latents) - image = self.image_processor.postprocess(image, output_type=output_type) + # 10. Run safety checker + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) # 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 image, has_nsfw_concept - return StableDiffusionXLPipelineOutput(images=image) + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) def text2img( self, - prompt: str = None, - prompt_2: Optional[str] = None, - height: Optional[int] = None, - width: Optional[int] = None, + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, + height: int = 512, + width: int = 512, num_inference_steps: int = 50, - timesteps: List[int] = None, - denoising_start: Optional[float] = None, - denoising_end: Optional[float] = None, - guidance_scale: float = 5.0, - negative_prompt: Optional[str] = None, - negative_prompt_2: Optional[str] = None, + guidance_scale: float = 7.5, 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, - ip_adapter_image: Optional[PipelineImageInput] = 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, + max_embeddings_multiples: Optional[int] = 3, output_type: Optional[str] = "pil", return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, 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, - clip_skip: Optional[int] = None, - callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, - callback_on_step_end_tensor_inputs: List[str] = ["latents"], - **kwargs, ): r""" - Function invoked when calling pipeline for text-to-image. + Function for text-to-image generation. + Args: + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + height (`int`, *optional*, defaults to 512): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to 512): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - Refer to the documentation of the `__call__` method for parameter descriptions. + Returns: + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. """ return self.__call__( prompt=prompt, - prompt_2=prompt_2, + negative_prompt=negative_prompt, height=height, width=width, num_inference_steps=num_inference_steps, - timesteps=timesteps, - denoising_start=denoising_start, - denoising_end=denoising_end, guidance_scale=guidance_scale, - negative_prompt=negative_prompt, - negative_prompt_2=negative_prompt_2, num_images_per_prompt=num_images_per_prompt, eta=eta, generator=generator, latents=latents, - ip_adapter_image=ip_adapter_image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, - pooled_prompt_embeds=pooled_prompt_embeds, - negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + max_embeddings_multiples=max_embeddings_multiples, output_type=output_type, return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, cross_attention_kwargs=cross_attention_kwargs, - guidance_rescale=guidance_rescale, - original_size=original_size, - crops_coords_top_left=crops_coords_top_left, - target_size=target_size, - clip_skip=clip_skip, - callback_on_step_end=callback_on_step_end, - callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, - **kwargs, ) def img2img( self, - prompt: str = None, - prompt_2: Optional[str] = None, - image: Optional[PipelineImageInput] = None, - height: Optional[int] = None, - width: Optional[int] = None, + image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, strength: float = 0.8, - num_inference_steps: int = 50, - timesteps: List[int] = None, - denoising_start: Optional[float] = None, - denoising_end: Optional[float] = None, - guidance_scale: float = 5.0, - negative_prompt: Optional[str] = None, - negative_prompt_2: Optional[str] = None, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, num_images_per_prompt: Optional[int] = 1, - eta: float = 0.0, + eta: Optional[float] = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, - latents: Optional[torch.FloatTensor] = None, - ip_adapter_image: Optional[PipelineImageInput] = 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, + max_embeddings_multiples: Optional[int] = 3, output_type: Optional[str] = "pil", return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, 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, - clip_skip: Optional[int] = None, - callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, - callback_on_step_end_tensor_inputs: List[str] = ["latents"], - **kwargs, ): r""" - Function invoked when calling pipeline for image-to-image. + Function for image-to-image generation. + Args: + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. + `image` will be used as a starting point, adding more noise to it the larger the `strength`. The + number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added + noise will be maximum and the denoising process will run for the full number of iterations specified in + `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. This parameter will be modulated by `strength`. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - Refer to the documentation of the `__call__` method for parameter descriptions. + Returns: + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. """ return self.__call__( prompt=prompt, - prompt_2=prompt_2, + negative_prompt=negative_prompt, image=image, - height=height, - width=width, - strength=strength, num_inference_steps=num_inference_steps, - timesteps=timesteps, - denoising_start=denoising_start, - denoising_end=denoising_end, guidance_scale=guidance_scale, - negative_prompt=negative_prompt, - negative_prompt_2=negative_prompt_2, + strength=strength, num_images_per_prompt=num_images_per_prompt, eta=eta, generator=generator, - latents=latents, - ip_adapter_image=ip_adapter_image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, - pooled_prompt_embeds=pooled_prompt_embeds, - negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + max_embeddings_multiples=max_embeddings_multiples, output_type=output_type, return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, cross_attention_kwargs=cross_attention_kwargs, - guidance_rescale=guidance_rescale, - original_size=original_size, - crops_coords_top_left=crops_coords_top_left, - target_size=target_size, - clip_skip=clip_skip, - callback_on_step_end=callback_on_step_end, - callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, - **kwargs, ) def inpaint( self, - prompt: str = None, - prompt_2: Optional[str] = None, - image: Optional[PipelineImageInput] = None, - mask_image: Optional[PipelineImageInput] = None, - masked_image_latents: Optional[torch.FloatTensor] = None, - height: Optional[int] = None, - width: Optional[int] = None, + image: Union[torch.FloatTensor, PIL.Image.Image], + mask_image: Union[torch.FloatTensor, PIL.Image.Image], + prompt: Union[str, List[str]], + negative_prompt: Optional[Union[str, List[str]]] = None, strength: float = 0.8, - num_inference_steps: int = 50, - timesteps: List[int] = None, - denoising_start: Optional[float] = None, - denoising_end: Optional[float] = None, - guidance_scale: float = 5.0, - negative_prompt: Optional[str] = None, - negative_prompt_2: Optional[str] = None, + num_inference_steps: Optional[int] = 50, + guidance_scale: Optional[float] = 7.5, num_images_per_prompt: Optional[int] = 1, - eta: float = 0.0, + add_predicted_noise: Optional[bool] = False, + eta: Optional[float] = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, - latents: Optional[torch.FloatTensor] = None, - ip_adapter_image: Optional[PipelineImageInput] = 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, + max_embeddings_multiples: Optional[int] = 3, output_type: Optional[str] = "pil", return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + is_cancelled_callback: Optional[Callable[[], bool]] = None, + callback_steps: int = 1, 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, - clip_skip: Optional[int] = None, - callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, - callback_on_step_end_tensor_inputs: List[str] = ["latents"], - **kwargs, ): r""" - Function invoked when calling pipeline for inpainting. + Function for inpaint. + Args: + image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, that will be used as the starting point for the + process. This is the image whose masked region will be inpainted. + mask_image (`torch.FloatTensor` or `PIL.Image.Image`): + `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be + replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a + PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should + contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. + prompt (`str` or `List[str]`): + The prompt or prompts to guide the image generation. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored + if `guidance_scale` is less than `1`). + strength (`float`, *optional*, defaults to 0.8): + Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength` + is 1, the denoising process will be run on the masked area for the full number of iterations specified + in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more + noise to that region the larger the `strength`. If `strength` is 0, no inpainting will occur. + num_inference_steps (`int`, *optional*, defaults to 50): + The reference number of denoising steps. More denoising steps usually lead to a higher quality image at + the expense of slower inference. This parameter will be modulated by `strength`, as explained above. + guidance_scale (`float`, *optional*, defaults to 7.5): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + add_predicted_noise (`bool`, *optional*, defaults to True): + Use predicted noise instead of random noise when constructing noisy versions of the original image in + the reverse diffusion process + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + max_embeddings_multiples (`int`, *optional*, defaults to `3`): + The max multiple length of prompt embeddings compared to the max output length of text encoder. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + is_cancelled_callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. If the function returns + `True`, the inference will be cancelled. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). - Refer to the documentation of the `__call__` method for parameter descriptions. + Returns: + `None` if cancelled by `is_cancelled_callback`, + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. + When returning a tuple, the first element is a list with the generated images, and the second element is a + list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" + (nsfw) content, according to the `safety_checker`. """ return self.__call__( prompt=prompt, - prompt_2=prompt_2, + negative_prompt=negative_prompt, image=image, mask_image=mask_image, - masked_image_latents=masked_image_latents, - height=height, - width=width, - strength=strength, num_inference_steps=num_inference_steps, - timesteps=timesteps, - denoising_start=denoising_start, - denoising_end=denoising_end, guidance_scale=guidance_scale, - negative_prompt=negative_prompt, - negative_prompt_2=negative_prompt_2, + strength=strength, num_images_per_prompt=num_images_per_prompt, + add_predicted_noise=add_predicted_noise, eta=eta, generator=generator, - latents=latents, - ip_adapter_image=ip_adapter_image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, - pooled_prompt_embeds=pooled_prompt_embeds, - negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + max_embeddings_multiples=max_embeddings_multiples, output_type=output_type, return_dict=return_dict, + callback=callback, + is_cancelled_callback=is_cancelled_callback, + callback_steps=callback_steps, cross_attention_kwargs=cross_attention_kwargs, - guidance_rescale=guidance_rescale, - original_size=original_size, - crops_coords_top_left=crops_coords_top_left, - target_size=target_size, - clip_skip=clip_skip, - callback_on_step_end=callback_on_step_end, - callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, - **kwargs, - ) - - # Overrride to properly handle the loading and unloading of the additional text encoder. - def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs): - # We could have accessed the unet config from `lora_state_dict()` too. We pass - # it here explicitly to be able to tell that it's coming from an SDXL - # pipeline. - state_dict, network_alphas = self.lora_state_dict( - pretrained_model_name_or_path_or_dict, - unet_config=self.unet.config, - **kwargs, - ) - self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet) - - text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k} - if len(text_encoder_state_dict) > 0: - self.load_lora_into_text_encoder( - text_encoder_state_dict, - network_alphas=network_alphas, - text_encoder=self.text_encoder, - prefix="text_encoder", - lora_scale=self.lora_scale, - ) - - text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k} - if len(text_encoder_2_state_dict) > 0: - self.load_lora_into_text_encoder( - text_encoder_2_state_dict, - network_alphas=network_alphas, - text_encoder=self.text_encoder_2, - prefix="text_encoder_2", - lora_scale=self.lora_scale, - ) - - @classmethod - def save_lora_weights( - self, - save_directory: Union[str, os.PathLike], - unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, - text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, - text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, - is_main_process: bool = True, - weight_name: str = None, - save_function: Callable = None, - safe_serialization: bool = False, - ): - state_dict = {} - - def pack_weights(layers, prefix): - layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers - layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()} - return layers_state_dict - - state_dict.update(pack_weights(unet_lora_layers, "unet")) - - if text_encoder_lora_layers and text_encoder_2_lora_layers: - state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder")) - state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2")) - - self.write_lora_layers( - state_dict=state_dict, - save_directory=save_directory, - is_main_process=is_main_process, - weight_name=weight_name, - save_function=save_function, - safe_serialization=safe_serialization, - ) - - def _remove_text_encoder_monkey_patch(self): - self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder) - self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2) \ No newline at end of file + ) \ No newline at end of file