[Mirror]Danbooru2023: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset

Danbooru2023 is an extension of Danbooru2021, featuring over 6.8 million anime-style images, totaling more than 8.3 TB. Each image is accompanied by community-contributed tags that provide detailed descriptions of its content, including characters, artists, copyright information, concepts, and attire. This makes it a crucial resource for stylized computer vision tasks and transfer learning.

This dataset mirrors nyanko7/danbooru2023. It has been converted to Parquet format to optimize compatibility with the Hugging Face ecosystem and enhance streaming performance. We gratefully acknowledge the excellent work contributed by Nyanko Devs (nyanko7).

To ensure research reproducibility, this dataset will not receive future updates, even if the original dataset (which has not been updated for over a year) undergoes changes. As the license is inherited from the original dataset, any copyright disputes affecting it will lead to corresponding adjustments in this mirror. For DOI requests, please open an issue on the repository.

The dataset includes:

• Image Files: Numerous anime-style images, varying in resolution, stored as raw bytes within Parquet files.
• Tag Data: Textual tags associated with each image, manually added and reviewed by community users.
• Other Metadata: Includes image upload time, ratings, favorite counts, and original URLs.

Dataset Details

• Curated by: Nyanko Devs (nyanko7)
• Language(s): Primarily English, also containing some Japanese or other Romaji spellings.
• License: MIT
• Repository: https://huggingface.co/datasets/zenless-lab/danbooru2023
• Original Repository: https://huggingface.co/datasets/nyanko7/danbooru2023

Intended Use

This dataset was initially employed for research into catastrophic forgetting during transfer learning, especially when substantial distribution disparities exist between pre-training and fine-tuning datasets.

It is also well-suited for training anime-style image generation models and other research objectives, including image analysis.

Usage

Given the substantial size of this dataset, it can be accessed using the datasets library's streaming functionality (in 1GB blocks) or by leveraging the glob method to acquire a local, random subset.

from datasets import load_dataset

ds = load_dataset("zenless-lab/danbooru2023", split='train', streaming=True)
next(iter(ds))

Alternatively, directly use glob to acquire a local, random subset:

from datasets import load_dataset

# The method below will read the first and second Parquet blocks, downloading approximately 2GB of data.
ds = load_dataset("parquet", data_files="hf://datasets/zenless-lab/danbooru2023/data/train-0000[12]-*.parquet", split="train")
next(iter(ds))

Bias, Risks, and Limitations

The image quality within the Danbooru2023 dataset is inconsistent, and it contains a significant proportion of R18 and R18G content. As not all image origins can be traced, proper authorization cannot be guaranteed; therefore, users are strongly advised to exercise discretion prior to use. The dataset may exhibit a strong bias towards styles preferred by the English-speaking community, a reflection of the original uploaders' preferences.

While images have unique UIDs, duplicates may still be present as fuzzy deduplication was not performed.

Licensing

This dataset operates under the MIT license, inherited from nyanko7/danbooru2023. However, no independent copyright investigation of the included images has been conducted. Users are strongly advised to ensure their intended use adheres to the copyright laws of their respective jurisdictions, determining if it qualifies for applicable copyright exemptions (e.g., fair use in the U.S., Article 30-4 of the Japanese Copyright Act in Japan, or the Article 3/4(TDM exemption) of DSM in EU). This document does not constitute legal advice; consultation with a qualified legal professional is recommended to evaluate potential risks.