| --- |
| license: mit |
| task_categories: |
| - image-classification |
| - image-to-image |
| - text-to-image |
| language: |
| - en |
| - ja |
| tags: |
| - anime |
| - danbooru |
| - art |
| pretty_name: danbooru2025 |
| size_categories: |
| - 10M<n<100M |
| viewer: false |
| --- |
| |
| <img src="https://huggingface.co/datasets/nyanko-devs/danbooru2026/resolve/main/banner-5.png" alt="Danbooru2026 Cover" width="750"/> |
|
|
| # Danbooru2026: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset [WIP] |
|
|
| ## Dataset Description |
|
|
| Danbooru2026 is a large-scale anime illustration dataset containing over 10 million community-annotated images. It is intended for research and development in anime-style image generation, image classification, multimodal learning, and related tasks. |
|
|
| Danbooru is a long-running image board known for its extensive tagging system and community-maintained metadata. |
|
|
| **Key Specifications:** |
|
|
| * **Shared by:** Nyanko Devs |
| * **Source:** Danbooru Community |
| * **Language(s):** English, Japanese, Mixed |
| * **Dataset License:** MIT, Original images retain their respective copyrights |
| * **Image Format:** WebP |
| * **Image Count:** 11M+ (IDs up to #11,740,611) |
|
|
| This dataset builds upon [danbooru2023](https://huggingface.co/datasets/nyanko7/danbooru2023) and expands the collection by more than 4.5 million images. |
|
|
| ## Image Processing |
|
|
| Danbooru2026 contains images only. Non-image files, animations, videos, audio files, archives, documents, and other unsupported media are excluded. |
|
|
| All included images are: |
|
|
| * decoded from their original source files, converted to the WebP format; |
| * resized to a maximum of 4 million pixels while preserving the original aspect ratio; |
| * stored without upscaling images that are already below the resolution limit. |
|
|
| For an image with width (W) and height (H), resizing is applied only when: |
|
|
| ```text |
| W × H > 4096 × 4096 |
| ``` |
|
|
| ## Dataset Structure and Format |
|
|
| The dataset is designed to be straightforward to use, avoiding obscure storage formats while supporting efficient downloading, local access, research, and torrent seeding. |
|
|
| Images are stored as WebP files and distributed across 1,000 zero-padded buckets named `0000` through `0999`. The bucket for each image is determined by the Danbooru post ID modulo 1,000. |
|
|
| For example: |
|
|
| ```bash |
| BUCKET=$(printf "%04d" $((ID % 1000))) |
| ``` |
|
|
| The corresponding image path is: |
|
|
| ```text |
| original/$BUCKET/$ID.webp |
| ``` |
|
|
| An image with post ID `210001` is therefore stored at: |
|
|
| ```text |
| original/0001/210001.webp |
| ``` |
|
|
| ### File Tree |
|
|
| ```text |
| / |
| └── danbooru2026/ |
| ├── README.md |
| ├── metadata/ |
| │ ├── posts.parquet |
| │ └── ... |
| └── original/ |
| ├── 0000/ -> data-0000.tar |
| ├── 0001/ -> data-0001.tar |
| │ ├── 10001.webp |
| │ ├── 210001.webp |
| │ └── ... |
| └── ... (up to 0999) |
| ``` |
|
|
| ## Notes |
|
|
| The processed dataset does not preserve the original source file formats. Every successfully decoded item is provided as a WebP image with a maximum resolution of 4 megapixels. |
|
|
| Files that cannot be safely decoded as images are excluded from the processed image collection. This includes corrupted files, incorrectly labeled non-image files, unsupported media, and other invalid source data. |
|
|
