metadata

license: mit
task_categories:
  - text-to-image
  - text-to-video
language:
  - en
  - ja
tags:
  - animation
  - image
  - video
  - webdataset
pretty_name: Sakugabooru 2025
size_categories:
  - 100K<n<1M

Sakugabooru2025: Curated Animation Clips from Enthusiasts

Sakugabooru.com is a booru-style imageboard dedicated to collecting and sharing noteworthy animation clips, emphasizing Japanese anime but open to creators worldwide. Over the years, it has amassed more than 240,000 animation clips, alongside informative blog posts for anime fans everywhere.

With the growing interest in generative video models and AI animations, the scarcity of proper animation-related video datasets has become a challenge. This dataset was created to provide resources for animation enthusiasts and industry researchers, to help advancing the frontier of animation-video research.

Potential uses include:

text-to-video training
multimodal animation research
video quality/aesthetics analysis

Dataset Details

This dataset includes over 155,238 video clips and 8,680 images of unique or noteworthy animation:

image files: 8680, video files: 155238, json files: 240242
total unique files: 404160
last post id: 273264

Note that Sakugabooru’s archives have been reduced by DMCA strikes over the years. Popular titles like Mob Psycho 100 may not appear because they were removed from the site.

Curated by: trojblue
Language(s) (NLP): English, Japanese
License: MIT

Dataset Structure

Files are split into tar archives by post ID modulo 1,000:

Example: post ID 42 goes to 0000.tar, post ID 5306 goes to 0005.tar.
Each tar is about 1GB, suitable for large-batch data processing.

The dataset follows WebDataset conventions, with each tar containing media files (~1GB) plus metadata:

./media/0.tar
# ./train/0.tar/{sakuga_5306.webm}
# ./train/0.tar/{sakuga_5306.json}

The index file ./sakugabooru-index.json is generated using widsindex (see WebDataset FAQ):

widsindex *.tar > mydataset-index.json

Dataset Creation

This dataset is sourced directly from Sakugabooru by querying IDs from 0 up to the latest, by Dec.28, 2024:

https://www.sakugabooru.com/post/show/{id}

Data Collection and Processing

Since Sakugabooru lacks a direct JSON API, metadata were gathered by visiting each page and extracting relevant parts. The scraper code is available here: arot-devs/sakuga-scraper.

Each post’s JSON metadata follows a format like:

{
    "post_id": 100112,
    "post_url": "https://www.sakugabooru.com/post/show/100112",
    "image_url": null,
    "tags": {},
    ...
    "status_notice": [
        "This post was deleted.\n\n      Reason: ..."
    ]
}

No additional processing was performed beyond metadata collection.

Bias, Risks, and Limitations

Owing to DMCA takedowns, duplicates, or other factors, only about 60% (163,918 of 273,264) post IDs have attached media. The selection also tends to favor older Japanese animations.

Recommendations

For more interesting reads about animation, visit: Sakuga Blog – The Art of Japanese Animation

You can support Sakugabooru through: Sakugabooru | Patreon