# Copyright 2022 Jay Wang, Evan Montoya, David Munechika, Alex Yang, Ben Hoover, Polo Chau # MIT License """Loading script for DiffusionDB.""" import re import numpy as np import pandas as pd from json import load, dump from os.path import join, basename from huggingface_hub import hf_hub_url import datasets # Find for instance the citation on arxiv or on the dataset repo/website _CITATION = """\ @article{wangDiffusionDBLargescalePrompt2022, title = {{{DiffusionDB}}: {{A}} Large-Scale Prompt Gallery Dataset for Text-to-Image Generative Models}, author = {Wang, Zijie J. and Montoya, Evan and Munechika, David and Yang, Haoyang and Hoover, Benjamin and Chau, Duen Horng}, year = {2022}, journal = {arXiv:2210.14896 [cs]}, url = {https://arxiv.org/abs/2210.14896} } """ # You can copy an official description _DESCRIPTION = """ DiffusionDB is the first large-scale text-to-image prompt dataset. It contains 2 million images generated by Stable Diffusion using prompts and hyperparameters specified by real users. The unprecedented scale and diversity of this human-actuated dataset provide exciting research opportunities in understanding the interplay between prompts and generative models, detecting deepfakes, and designing human-AI interaction tools to help users more easily use these models. """ _HOMEPAGE = "https://poloclub.github.io/diffusiondb" _LICENSE = "CC0 1.0" _VERSION = datasets.Version("0.9.1") # Programmatically generate the URLs for different parts # hf_hub_url() provides a more flexible way to resolve the file URLs # https://huggingface.co/datasets/poloclub/diffusiondb/resolve/main/images/part-000001.zip _URLS = {} _URLS_LARGE = {} _PART_IDS = range(1, 2001) _PART_IDS_LARGE = range(1, 14001) for i in _PART_IDS: _URLS[i] = hf_hub_url( "poloclub/diffusiondb", filename=f"images/part-{i:06}.zip", repo_type="dataset", ) for i in _PART_IDS_LARGE: if i < 10001: _URLS_LARGE[i] = hf_hub_url( "poloclub/diffusiondb", filename=f"diffusiondb-large-part-1/part-{i:06}.zip", repo_type="dataset", ) else: _URLS_LARGE[i] = hf_hub_url( "poloclub/diffusiondb", filename=f"diffusiondb-large-part-2/part-{i:06}.zip", repo_type="dataset", ) # Add the metadata parquet URL as well _URLS["metadata"] = hf_hub_url( "poloclub/diffusiondb", filename="metadata.parquet", repo_type="dataset" ) _URLS_LARGE["metadata"] = hf_hub_url( "poloclub/diffusiondb", filename="metadata-large.parquet", repo_type="dataset", ) _SAMPLER_DICT = { 1: "ddim", 2: "plms", 3: "k_euler", 4: "k_euler_ancestral", 5: "ddik_heunm", 6: "k_dpm_2", 7: "k_dpm_2_ancestral", 8: "k_lms", 9: "others", } class DiffusionDBConfig(datasets.BuilderConfig): """BuilderConfig for DiffusionDB.""" def __init__(self, part_ids, is_large, **kwargs): """BuilderConfig for DiffusionDB. Args: part_ids([int]): A list of part_ids. is_large(bool): If downloading data from DiffusionDB Large (14 million) **kwargs: keyword arguments forwarded to super. """ super(DiffusionDBConfig, self).__init__(version=_VERSION, **kwargs) self.part_ids = part_ids self.is_large = is_large class DiffusionDB(datasets.GeneratorBasedBuilder): """A large-scale text-to-image prompt gallery dataset based on Stable Diffusion.""" BUILDER_CONFIGS = [] # Programmatically generate configuration options (HF requires to use a string # as the config key) for num_k in [1, 5, 10, 50, 100, 500, 1000]: for sampling in ["first", "random"]: for is_large in [False, True]: num_k_str = f"{num_k}k" if num_k < 1000 else f"{num_k // 1000}m" subset_str = "large_" if is_large else "2m_" if sampling == "random": # Name the config cur_name = subset_str + "random_" + num_k_str # Add a short description for each config cur_description = ( f"Random {num_k_str} images with their prompts and parameters" ) # Sample part_ids total_part_ids = _PART_IDS_LARGE if is_large else _PART_IDS part_ids = np.random.choice( total_part_ids, num_k, replace=False ).tolist() else: # Name the config cur_name = subset_str + "first_" + num_k_str # Add a short description for each config cur_description = f"The first {num_k_str} images in this dataset with their prompts and parameters" # Sample part_ids total_part_ids = _PART_IDS_LARGE if is_large else _PART_IDS part_ids = total_part_ids[1 : num_k + 1] # Create configs BUILDER_CONFIGS.append( DiffusionDBConfig( name=cur_name, part_ids=part_ids, is_large=is_large, description=cur_description, ), ) # Add few more options for Large only for num_k in [5000, 10000]: for sampling in ["first", "random"]: num_k_str = f"{num_k // 1000}m" subset_str = "large_" if sampling == "random": # Name the config cur_name = subset_str + "random_" + num_k_str # Add a short description for each config cur_description = ( f"Random {num_k_str} images with their prompts and parameters" ) # Sample part_ids total_part_ids = _PART_IDS_LARGE part_ids = np.random.choice( total_part_ids, num_k, replace=False ).tolist() else: # Name the config cur_name = subset_str + "first_" + num_k_str # Add a short description for each config cur_description = f"The first {num_k_str} images in this dataset with their prompts and parameters" # Sample part_ids total_part_ids = _PART_IDS_LARGE part_ids = total_part_ids[1 : num_k + 1] # Create configs BUILDER_CONFIGS.append( DiffusionDBConfig( name=cur_name, part_ids=part_ids, is_large=True, description=cur_description, ), ) # Need to manually add all (2m) and all (large) BUILDER_CONFIGS.append( DiffusionDBConfig( name="2m_all", part_ids=_PART_IDS, is_large=False, description="All images with their prompts and parameters", ), ) BUILDER_CONFIGS.append( DiffusionDBConfig( name="large_all", part_ids=_PART_IDS_LARGE, is_large=True, description="All images with their prompts and parameters", ), ) # We also prove a text-only option, which loads the meatadata parquet file BUILDER_CONFIGS.append( DiffusionDBConfig( name="2m_text_only", part_ids=[], is_large=False, description="Only include all prompts and parameters (no image)", ), ) BUILDER_CONFIGS.append( DiffusionDBConfig( name="large_text_only", part_ids=[], is_large=True, description="Only include all prompts and parameters (no image)", ), ) # Add a random 1k from 2M as the first entry point to show on HF data viewer # Sample part_ids part_ids = np.random.choice(_PART_IDS, 1000, replace=False).tolist() BUILDER_CONFIGS.append( DiffusionDBConfig( name="1k_random_2m", part_ids=part_ids, is_large=False, description="Another random 1k images with meta data from DiffusionDB 2M", ), ) # Default to only load 1k random images DEFAULT_CONFIG_NAME = "2m_random_1k" def _info(self): """Specify the information of DiffusionDB.""" if "text_only" in self.config.name: features = datasets.Features( { "image_name": datasets.Value("string"), "prompt": datasets.Value("string"), "part_id": datasets.Value("uint16"), "seed": datasets.Value("uint32"), "step": datasets.Value("uint16"), "cfg": datasets.Value("float32"), "sampler": datasets.Value("string"), "width": datasets.Value("uint16"), "height": datasets.Value("uint16"), "user_name": datasets.Value("string"), "timestamp": datasets.Value("timestamp[us, tz=UTC]"), "image_nsfw": datasets.Value("float32"), "prompt_nsfw": datasets.Value("float32"), }, ) else: features = datasets.Features( { "image": datasets.Image(), "prompt": datasets.Value("string"), "seed": datasets.Value("uint32"), "step": datasets.Value("uint16"), "cfg": datasets.Value("float32"), "sampler": datasets.Value("string"), "width": datasets.Value("uint16"), "height": datasets.Value("uint16"), "user_name": datasets.Value("string"), "timestamp": datasets.Value("timestamp[us, tz=UTC]"), "image_nsfw": datasets.Value("float32"), "prompt_nsfw": datasets.Value("float32"), }, ) return datasets.DatasetInfo( description=_DESCRIPTION, features=features, supervised_keys=None, homepage=_HOMEPAGE, license=_LICENSE, citation=_CITATION, ) def _split_generators(self, dl_manager): # If several configurations are possible (listed in BUILDER_CONFIGS), # the configuration selected by the user is in self.config.name # dl_manager is a datasets.download.DownloadManager that can be used to # download and extract URLS It can accept any type or nested list/dict # and will give back the same structure with the url replaced with path # to local files. By default the archives will be extracted and a path # to a cached folder where they are extracted is returned instead of the # archive # Download and extract zip files of all sampled part_ids data_dirs = [] json_paths = [] # Resolve the urls if self.config.is_large: urls = _URLS_LARGE else: urls = _URLS for cur_part_id in self.config.part_ids: cur_url = urls[cur_part_id] data_dir = dl_manager.download_and_extract(cur_url) data_dirs.append(data_dir) json_paths.append(join(data_dir, f"part-{cur_part_id:06}.json")) # Also download the metadata table metadata_path = dl_manager.download(urls["metadata"]) return [ datasets.SplitGenerator( name=datasets.Split.TRAIN, # These kwargs will be passed to _generate_examples gen_kwargs={ "data_dirs": data_dirs, "json_paths": json_paths, "metadata_path": metadata_path, }, ), ] def _generate_examples(self, data_dirs, json_paths, metadata_path): # This method handles input defined in _split_generators to yield # (key, example) tuples from the dataset. # The `key` is for legacy reasons (tfds) and is not important in itself, # but must be unique for each example. # Load the metadata parquet file if the config is text_only if "text_only" in self.config.name: metadata_df = pd.read_parquet(metadata_path) for _, row in metadata_df.iterrows(): yield row["image_name"], { "image_name": row["image_name"], "prompt": row["prompt"], "part_id": row["part_id"], "seed": row["seed"], "step": row["step"], "cfg": row["cfg"], "sampler": _SAMPLER_DICT[int(row["sampler"])], "width": row["width"], "height": row["height"], "user_name": row["user_name"], "timestamp": None if pd.isnull(row["timestamp"]) else row["timestamp"], "image_nsfw": row["image_nsfw"], "prompt_nsfw": row["prompt_nsfw"], } else: num_data_dirs = len(data_dirs) assert num_data_dirs == len(json_paths) # Read the metadata table (only rows with the needed part_ids) part_ids = [] for path in json_paths: cur_id = int(re.sub(r"part-(\d+)\.json", r"\1", basename(path))) part_ids.append(cur_id) # We have to use pandas here to make the dataset preview work (it # uses streaming mode) metadata_table = pd.read_parquet( metadata_path, filters=[("part_id", "in", part_ids)], ) # Iterate through all extracted zip folders for images for k in range(num_data_dirs): cur_data_dir = data_dirs[k] cur_json_path = json_paths[k] json_data = load(open(cur_json_path, "r", encoding="utf8")) for img_name in json_data: img_params = json_data[img_name] img_path = join(cur_data_dir, img_name) # Query the metadata query_result = metadata_table.query(f'`image_name` == "{img_name}"') # Yields examples as (key, example) tuples yield img_name, { "image": { "path": img_path, "bytes": open(img_path, "rb").read(), }, "prompt": img_params["p"], "seed": int(img_params["se"]), "step": int(img_params["st"]), "cfg": float(img_params["c"]), "sampler": img_params["sa"], "width": query_result["width"].to_list()[0], "height": query_result["height"].to_list()[0], "user_name": query_result["user_name"].to_list()[0], "timestamp": None if pd.isnull(query_result["timestamp"].to_list()[0]) else query_result["timestamp"].to_list()[0], "image_nsfw": query_result["image_nsfw"].to_list()[0], "prompt_nsfw": query_result["prompt_nsfw"].to_list()[0], }