#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ JoyCaption Alpha One This module provides functionality for generating captions for images using a combination of CLIP, LLM, and custom image adapters. It supports various caption types, tones, and lengths. The main components include: - Loading and initializing models (CLIP, LLM, image adapter) - Processing images and generating captions - Command-line interface for batch processing images in a directory """ import os import argparse import re import random from pathlib import Path from typing import List, Tuple, Dict from PIL import Image import pillow_jxl import torch import torchvision.transforms.functional as TVF from transformers import ( AutoModel, AutoTokenizer, AutoModelForCausalLM, PreTrainedTokenizer, PreTrainedTokenizerFast, ) from torch import nn from e6db_reader import TagSetNormalizer, tag_category2id, tag_rank_to_freq CLIP_PATH = "google/siglip-so400m-patch14-384" MODEL_PATH = "meta-llama/Meta-Llama-3.1-8B" CHECKPOINT_PATH = Path(__file__).resolve().parent / "9em124t2-499968" CAPTION_TYPE_MAP = { ("descriptive", "formal", False, False): [ "Write a descriptive caption for this image in a formal tone." ], ("descriptive", "formal", False, True): [ "Write a descriptive caption for this image in a formal tone within " "{word_count} words." ], ("descriptive", "formal", True, False): [ "Write a {length} descriptive caption for this image in a formal tone." ], ("descriptive", "informal", False, False): [ "Write a descriptive caption for this image in a casual tone." ], ("descriptive", "informal", False, True): [ "Write a descriptive caption for this image in a casual tone within " "{word_count} words." ], ("descriptive", "informal", True, False): [ "Write a {length} descriptive caption for this image in a casual tone." ], ("training_prompt", "formal", False, False): [ "Write a stable diffusion prompt for this image." ], ("training_prompt", "formal", False, True): [ "Write a stable diffusion prompt for this image within " + "{word_count} words." ], ("training_prompt", "formal", True, False): [ "Write a {length} stable diffusion prompt for this image." ], ("rng-tags", "formal", False, False): [ "Write a list of Booru tags for this image." ], ("rng-tags", "formal", False, True): [ "Write a list of Booru tags for this image within {word_count} words." ], ("rng-tags", "formal", True, False): [ "Write a {length} list of Booru tags for this image." ], } HF_TOKEN = os.environ.get("HF_TOKEN", None) class ImageAdapter(nn.Module): """ Custom image adapter module for processing CLIP vision outputs. This module adapts the output of a CLIP vision model to be compatible with a text model. It supports optional layer normalization, positional embeddings, and deep feature extraction. Args: input_features (int): Number of input features from the vision model. output_features (int): Number of output features to match the text model. ln1 (bool): Whether to use layer normalization. pos_emb (bool): Whether to use positional embeddings. num_image_tokens (int): Number of image tokens. deep_extract (bool): Whether to use deep feature extraction. """ def __init__( self, input_features: int, output_features: int, ln1: bool, pos_emb: bool, num_image_tokens: int, deep_extract: bool, ): super().__init__() self.deep_extract = deep_extract if self.deep_extract: input_features = input_features * 5 self.linear1 = nn.Linear(input_features, output_features) self.activation = nn.GELU() self.linear2 = nn.Linear(output_features, output_features) self.ln1 = nn.Identity() if not ln1 else nn.LayerNorm(input_features) self.pos_emb = ( None if not pos_emb else nn.Parameter(torch.zeros(num_image_tokens, input_features)) ) self.other_tokens = nn.Embedding(3, output_features) self.other_tokens.weight.data.normal_(mean=0.0, std=0.02) def forward(self, vision_outputs: torch.Tensor): """ Forward pass of the image adapter. Args: vision_outputs (torch.Tensor): Output tensor from the CLIP vision model. Returns: torch.Tensor: Adapted image features. """ if self.deep_extract: x = torch.concat( ( vision_outputs[-2], vision_outputs[3], vision_outputs[7], vision_outputs[13], vision_outputs[20], ), dim=-1, ) assert len(x.shape) == 3, f"Expected 3, got {len(x.shape)}" expected_shape = vision_outputs[-2].shape[-1] * 5 assert ( x.shape[-1] == expected_shape ), f"Expected {expected_shape}, got {x.shape[-1]}" else: x = vision_outputs[-2] x = self.ln1(x) if self.pos_emb is not None: assert ( x.shape[-2:] == self.pos_emb.shape ), f"Expected {self.pos_emb.shape}, got {x.shape[-2:]}" x = x + self.pos_emb x = self.linear1(x) x = self.activation(x) x = self.linear2(x) other_tokens = self.other_tokens( torch.tensor([0, 1], device=self.other_tokens.weight.device) .expand(x.shape[0], -1) ) assert other_tokens.shape == ( x.shape[0], 2, x.shape[2], ), f"Expected {(x.shape[0], 2, x.shape[2])}, got {other_tokens.shape}" x = torch.cat((other_tokens[:, 0:1], x, other_tokens[:, 1:2]), dim=1) return x def get_eot_embedding(self): """ Get the end-of-text embedding. Returns: torch.Tensor: The end-of-text embedding. """ return self.other_tokens( torch.tensor([2], device=self.other_tokens.weight.device) ).squeeze(0) class JoyCaptionModel: """ A class for generating captions for images using CLIP, LLM, and custom image adapters. This class encapsulates the functionality to load and initialize various models (CLIP, LLM, image adapter) and use them to process images and generate captions. It supports different caption types, tones, and lengths. Attributes: clip_model: The CLIP vision model for processing images. text_model: The language model for generating captions. image_adapter: Custom adapter for processing CLIP vision outputs. tokenizer: Tokenizer for the language model. Methods: load_models(): Load and initialize all required models. process_image(input_image, caption_type, caption_tone, caption_length): Process an input image and generate a caption based on specified parameters. """ def __init__(self): self.clip_model = None self.text_model = None self.image_adapter = None self.tokenizer = None def load_models(self): """ Load and initialize all required models (CLIP, LLM, image adapter). """ print("Loading CLIP") self.clip_model = AutoModel.from_pretrained(CLIP_PATH) self.clip_model = self.clip_model.vision_model if (CHECKPOINT_PATH / "clip_model.pt").exists(): print("Loading VLM's custom vision model") checkpoint = torch.load( CHECKPOINT_PATH / "clip_model.pt", map_location="cpu" ) checkpoint = { k.replace("_orig_mod.module.", ""): v for k, v in checkpoint.items() } self.clip_model.load_state_dict(checkpoint) del checkpoint self.clip_model.eval() self.clip_model.requires_grad_(False) self.clip_model.to("cuda") print("Loading tokenizer") self.tokenizer = AutoTokenizer.from_pretrained( MODEL_PATH, use_fast=False ) assert isinstance( self.tokenizer, (PreTrainedTokenizer, PreTrainedTokenizerFast) ) print("Loading LLM") if (CHECKPOINT_PATH / "text_model").exists(): print("Loading VLM's custom text model") self.text_model = AutoModelForCausalLM.from_pretrained( CHECKPOINT_PATH / "text_model", device_map=0, torch_dtype=torch.bfloat16 ) else: self.text_model = AutoModelForCausalLM.from_pretrained( MODEL_PATH, device_map="auto", torch_dtype=torch.bfloat16 ) self.text_model.eval() print("Loading image adapter") self.image_adapter = ImageAdapter( self.clip_model.config.hidden_size, self.text_model.config.hidden_size, False, False, 38, False, ) self.image_adapter.load_state_dict( torch.load( CHECKPOINT_PATH / "image_adapter.pt", map_location="cpu" ) ) self.image_adapter.eval() self.image_adapter.to("cuda") @torch.no_grad() def process_image( self, input_image: Image.Image, caption_type: str, caption_tone: str, caption_length: str | int, custom_prompt: str | None = None, ) -> str: """ Process an input image and generate a caption based on specified parameters. """ torch.cuda.empty_cache() if custom_prompt is not None: prompt_str = custom_prompt else: prompt_str = self._get_prompt_string( caption_type, caption_tone, caption_length ) print(f"Prompt: {prompt_str}") pixel_values = self._preprocess_image(input_image) prompt = self._tokenize_prompt(prompt_str) embedded_images = self._embed_image(pixel_values) inputs_embeds, input_ids, attention_mask = self._construct_inputs( embedded_images, prompt ) generate_ids = self._generate_caption(inputs_embeds, input_ids, attention_mask) caption = self._decode_caption(generate_ids, input_ids) return caption.strip() def generate_valid_caption( self, input_image: Image.Image, caption_type: str, caption_tone: str, caption_length: str | int, custom_prompt: str | None = None, ) -> str: """ Generate a valid caption, retrying if the caption contains only special characters, does not end with a period, exclamation mark, or question mark, contains the word fluffy more than once, repeats any word longer than 4 characters multiple times, or contains only one sentence. """ while True: caption = self.process_image( input_image, caption_type, caption_tone, caption_length, custom_prompt ) words = re.findall(r'\b\w{5,}\b', caption.lower()) word_counts = {word: words.count(word) for word in set(words)} sentence_count = len(re.findall(r'[.!?]', caption)) if (re.search(r'\w', caption) and caption[-1] in {'.', '!', '?'} and caption.lower().count('fluffy') <= 1 and all(count == 1 for count in word_counts.values()) and sentence_count > 1): return caption print("Generated caption is invalid. Retrying...") def _get_prompt_string(self, caption_type, caption_tone, caption_length): length = None if caption_length == "any" else caption_length if isinstance(length, str): try: length = int(length) except ValueError: pass if caption_type in {"rng-tags", "training_prompt"}: caption_tone = "formal" prompt_key = ( caption_type, caption_tone, isinstance(length, str), isinstance(length, int), ) if prompt_key not in CAPTION_TYPE_MAP: raise ValueError(f"Invalid caption type: {prompt_key}") prompt_str = CAPTION_TYPE_MAP[prompt_key][0].format( length=length, word_count=length ) return prompt_str def _preprocess_image(self, input_image): image = input_image.resize((384, 384), Image.LANCZOS) pixel_values = TVF.pil_to_tensor(image).unsqueeze(0) / 255.0 pixel_values = TVF.normalize(pixel_values, [0.5], [0.5]) pixel_values = pixel_values.to("cuda") return pixel_values def _tokenize_prompt(self, prompt_str): prompt = self.tokenizer.encode( prompt_str, return_tensors="pt", padding=False, truncation=False, add_special_tokens=False, ) return prompt def _embed_image(self, pixel_values): with torch.amp.autocast_mode.autocast("cuda", enabled=True): vision_outputs = self.clip_model( pixel_values=pixel_values, output_hidden_states=True ) image_features = vision_outputs.hidden_states embedded_images = self.image_adapter(image_features) embedded_images = embedded_images.to("cuda") return embedded_images def _construct_inputs(self, embedded_images, prompt): prompt_embeds = self.text_model.model.embed_tokens(prompt.to("cuda")) assert prompt_embeds.shape == ( 1, prompt.shape[1], self.text_model.config.hidden_size, ), ( f"Prompt shape is {prompt_embeds.shape}, expected " f"{(1, prompt.shape[1], self.text_model.config.hidden_size)}" ) embedded_bos = self.text_model.model.embed_tokens( torch.tensor( [[self.tokenizer.bos_token_id]], device=self.text_model.device, dtype=torch.int64, ) ) eot_embed = ( self.image_adapter.get_eot_embedding() .unsqueeze(0) .to(dtype=self.text_model.dtype) ) inputs_embeds = torch.cat( [ embedded_bos.expand(embedded_images.shape[0], -1, -1), embedded_images.to(dtype=embedded_bos.dtype), prompt_embeds.expand(embedded_images.shape[0], -1, -1), eot_embed.expand(embedded_images.shape[0], -1, -1), ], dim=1, ) input_ids = torch.cat( [ torch.tensor( [[self.tokenizer.bos_token_id]], dtype=torch.long ), torch.zeros( (1, embedded_images.shape[1]), dtype=torch.long ), prompt, torch.tensor( [[self.tokenizer.eos_token_id]], dtype=torch.long ), ], dim=1, ).to("cuda") attention_mask = torch.ones_like(input_ids) return inputs_embeds, input_ids, attention_mask def _generate_caption(self, inputs_embeds, input_ids, attention_mask): generate_ids = self.text_model.generate( input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=True, suppress_tokens=None, repetition_penalty=1.2, ) return generate_ids def _decode_caption(self, generate_ids, input_ids): generate_ids = generate_ids[:, input_ids.shape[1]:] if (generate_ids[0][-1] == self.tokenizer.eos_token_id or generate_ids[0][-1] == self.tokenizer.convert_tokens_to_ids( "<|eot_id|>")): generate_ids = generate_ids[:, :-1] caption = self.tokenizer.batch_decode( generate_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False )[0] return caption def main(): """ Generate captions for images in a directory and save them as .caption files. """ parser = argparse.ArgumentParser( description=( "Generate captions for images in a directory and save them as " ".caption files." ) ) parser.add_argument( "directory", type=str, help="Target directory containing images." ) parser.add_argument( "--caption_type", type=str, default="descriptive", choices=["descriptive", "training_prompt", "rng-tags", "custom"], help="Type of caption to generate.", ) parser.add_argument( "--caption_tone", type=str, default="formal", choices=["formal", "informal"], help="Tone of the caption.", ) parser.add_argument( "--caption_length", type=str, default="any", help="Length of the caption." ) parser.add_argument( "--dont-strip-commas", action="store_true", help=( "If set, commas will not be stripped from the generated captions." ), ) parser.add_argument( "--custom_prompt", type=str, help=( "Custom prompt for the captioner. " "Use with --caption_type custom." ), ) parser.add_argument( "--add-commas-to-sentence-ends", action="store_true", help="Add commas after periods in sentences", ) parser.add_argument( "--feed-from-tags", type=int, nargs="?", const=-1, help=( "Use .txt files with the same base filename " "as the images as input to the captioner. " "Optionally specify the number of tags to use." ), ) parser.add_argument( "--random-tags", type=int, help=( "Randomly select n number of tags. " "Only works if --feed-from-tags is enabled." ), ) args = parser.parse_args() # Validate random-tags usage if args.random_tags is not None and args.feed_from_tags is None: parser.error( "--random-tags can only be used when --feed-from-tags is enabled" ) print("Loading e621 tag data") tagset_normalizer = make_tagset_normalizer() # Initialize and load models joy_caption_model = JoyCaptionModel() joy_caption_model.load_models() # Validate custom prompt usage if args.caption_type == "custom" and not args.custom_prompt: parser.error( "--custom_prompt is required when using --caption_type custom" ) elif args.caption_type != "custom" and args.custom_prompt: parser.error( "--custom_prompt can only be used with --caption_type custom" ) image_extensions = {".webp", ".png", ".jpeg", ".jpg", ".jxl"} for image_path in Path(args.directory).rglob("*"): if image_path.suffix.lower() in image_extensions: caption_file = image_path.with_suffix(".caption") # Skip if the caption file already exists if caption_file.exists(): print(f"Skipping {image_path}: Caption file already exists.") continue input_image = Image.open(image_path).convert("RGB") # Use custom prompt if specified custom_prompt = None if args.caption_type == "custom": custom_prompt = args.custom_prompt elif args.feed_from_tags is not None: custom_prompt = prompt_from_tags( args, image_path, tagset_normalizer ) print(f"Custom prompt: {custom_prompt}") caption = joy_caption_model.generate_valid_caption( input_image, args.caption_type, args.caption_tone, args.caption_length, custom_prompt=custom_prompt, ) # Strip commas if the --dont-strip-commas flag is not set if not args.dont_strip_commas: # Existing comma stripping logic caption = re.sub(r",\s*([^\d])", r" \1", caption) # New feature: Add commas after periods if specified if args.add_commas_to_sentence_ends: caption = re.sub(r"(\.)(\s+)([A-Z])", r"\1,\2\3", caption) # Remove all newline characters caption = caption.replace("\n", " ") print(f"Caption for {image_path}:\n\n{caption}\n\n") # Save the caption to a .caption file with open(caption_file, "w", encoding="utf-8") as f: f.write(caption) print(f"Caption saved to {caption_file}") RE_PARENS_SUFFIX = re.compile(r"_\([^)]+\)$") E6DB_DATA = Path(__file__).resolve().parent / "data" def make_tagset_normalizer(): """ Create a TagSetNormalizer for encoding/decoding tags to and from integers. Configures it based on the provided config. """ # This loads all the aliases and implications tagset_normalizer = TagSetNormalizer(E6DB_DATA) tagid2cat = tagset_normalizer.tag_normalizer.tag_categories cat_artist = tag_category2id["artist"] cat2suffix = { tag_category2id["character"]: "_(character)", tag_category2id["lore"]: "_(lore)", tag_category2id["species"]: "_(species)", tag_category2id["copyright"]: "_(copyright)", } # Create additional aliases for tags using simple rules def input_map(tag, tid): # Make an alias without parentheses, it might conflict but we'll handle # it depending on `on_alias_conflict` config value. without_suffix = RE_PARENS_SUFFIX.sub("", tag) had_suffix = tag != without_suffix if had_suffix: yield without_suffix # Add an alias with the suffix (special case for artist) cat = tagid2cat[tid] if tid is not None else -1 if cat == cat_artist: artist = without_suffix.removeprefix("by_") if artist != without_suffix: yield artist if not had_suffix: yield f"{artist}_(artist)" else: yield f"by_{artist}" if not had_suffix: yield f"by_{artist}_(artist)" elif not had_suffix: suffix = cat2suffix.get(cat) if suffix is not None: yield f"{without_suffix}{suffix}" # Recognize tags where ':' were replaced by a space (aspect ratio) if ":" in tag: yield tag.replace(":", "_") return tagset_normalizer.map_inputs(input_map, on_conflict="ignore") def format_nl_list(word_list): """ Takes a list of words and generates a natural language output. """ n = len(word_list) assert n > 0 if n == 1: return word_list[0] if n == 2: return f"{word_list[0]} and {word_list[1]}" # n > 2 *head, last = word_list return ", ".join(head) + ", and " + last TAG_SPECIES = tag_category2id["species"] TAG_CHARACTER = tag_category2id["character"] TAG_ARTIST = tag_category2id["artist"] TAG_COPYRIGHT = tag_category2id["copyright"] TAG_META = tag_category2id["meta"] TAG_FREQ_THRESH = 0 def prompt_from_tags(args, image_path: Path, tagset_normalizer: TagSetNormalizer): """ Generates a prompt from tags associated with the given image. Args: args: Additional arguments for the function. image_path (Path): The path to the image file. tagset_normalizer (TagSetNormalizer): An instance to normalize the tag set. Returns: None """ tag_file = find_tag_file(image_path) if tag_file is None: return None with open(tag_file, "r", encoding="utf-8") as f: tags = f.read().lower().split(",") tag_id_to_cat_id = tagset_normalizer.tag_normalizer.tag_categories encode = tagset_normalizer.tag_normalizer.encode # These lists contain tuples (freq, tag, tag_id) tag_by_category: Dict[int, List[Tuple[int, str, int]]] = { cat: [] for cat in [TAG_ARTIST, TAG_CHARACTER, TAG_COPYRIGHT, TAG_SPECIES] } other_tags: List[Tuple[int, str, int]] = [] implied: set = set() for tag in tags: tag = tag.strip() # Encode the tag into a numerical id tag_id = encode(tag.replace(" ", "_")) if tag_id is None: other_tags.append((0, tag, 0)) implied.update(tagset_normalizer.implications_rej.get(0, ())) continue # Get the category of the tag cat_id = tag_id_to_cat_id[tag_id] # Skip meta tags if cat_id == TAG_META: continue implied.update(tagset_normalizer.implications.get(tag_id, ())) # Get the frequency of the tag freq = tag_rank_to_freq(tag_id) if freq < TAG_FREQ_THRESH: continue tag_by_category.get(cat_id, other_tags).append( (int(freq), tag, tag_id) ) other_tags = sorted( (int(freq), tag, tag_id) for freq, tag, tag_id in other_tags if tag_id not in implied ) for cat_id, cat_list in tag_by_category.items(): tag_by_category[cat_id] = sorted( (int(freq), tag, tag_id) for freq, tag, tag_id in cat_list if tag_id not in implied ) if args.random_tags is not None: # Randomly select tags if --random-tags is specified num_tags = min(args.random_tags, len(other_tags)) other_tags = random.sample( [ (i, tag, 0) for i, tag in enumerate(tags[: round(args.random_tags * 1.5)]) ], num_tags, ) elif args.feed_from_tags > 0: # Use specified number of tags if --feed-from-tags has a positive value other_tags = other_tags[: args.feed_from_tags] # Prepare sentence pieces artist_tag = tag_by_category[TAG_ARTIST] if artist_tag: artist_list = [str(tp[1]).removeprefix('by ') for tp in artist_tag[:4]] artist_txt = f"by {format_nl_list(artist_list)}" else: artist_txt = "" character_tag = tag_by_category[TAG_CHARACTER] if character_tag: tags = [tag for _, tag, _ in character_tag[:4]] character_txt = f"named {format_nl_list(tags)}" else: character_txt = "" species_tag = tag_by_category[TAG_SPECIES] if species_tag: species_txt = ( "of a " if len(character_tag) <= 1 and len(species_tag) <= 1 else "of " ) species_txt += format_nl_list([tp[1] for tp in species_tag[:4]]) else: if character_tag: species_txt = ( " a character" if len(character_tag) <= 1 else " characters" ) else: species_txt = "" copyright_tag = tag_by_category[TAG_COPYRIGHT] if copyright_tag: tags = [tag for _, tag, *_ in copyright_tag[:4]] copyright_txt = f"from {format_nl_list(tags)}" else: copyright_txt = "" tag_string = ", ".join(tp[1] for tp in other_tags) custom_prompt = ' '.join(s for s in [ "Write a descriptive caption for this image", artist_txt, species_txt, character_txt, copyright_txt, "in a formal tone. Use these tags to construct your caption:", tag_string, ] if s) return custom_prompt def find_tag_file(image_path): """ Find the corresponding .txt file for the given image path. Handles cases where the image has a -(number) suffix. """ base_name = image_path.stem tag_file = image_path.with_suffix(".txt") if tag_file.exists(): return tag_file # Handle -(number) suffix match = re.match(r"(.+)-\d+$", base_name) if match: base_name = match.group(1) tag_file = image_path.with_name(base_name).with_suffix(".txt") if tag_file.exists(): return tag_file return None if __name__ == "__main__": main()