| --- |
| tags: |
| - danbooru |
| - e621 |
| - furry |
| size_categories: |
| - 10K<n<100K |
| configs: |
| - config_name: default |
| data_files: |
| - split: train |
| path: data/*.parquet |
| --- |
| |
| A highly condensed, diversity-maximized image dataset distilled from booru-style image boards. |
| The original version used 2024 tag definition and is captioned with old outdated captioners. |
|
|
| [RicemanT](https://huggingface.co/datasets/RicemanT/booru-essence-2026) updated the tag definition to July/August 2026 and most importantly: |
|
|
| - Tagged additional tags using the https://huggingface.co/animetimm/convnextv2_huge.dbv4-full model for Danbooru, and https://huggingface.co/RedRocket/Hydra for E621. |
| |
| - Recaptioned the whole thing with the current best vision captioner available **Gemini 3.7 Flash** while using **Grok 4.6** and **Qwen 3.8 27B** as good backups for NSFW refusals. |
| |
| Unlike the original dataset, the artist trigger tag format was changed from **by artistname** to **Drawn by artistname**. This is my personal preference for my own training projects, you can use scripts to change the prepend format to your liking. |
| |
| The images can be obtained here [RicemanT/booru-essence-2026-tars](https://huggingface.co/datasets/RicemanT/booru-essence-2026-tars) |
| |
| This repo also contain the CSV files listing all characters and artists tag present within the dataset, and their count. |
| |
| The cost of this whole operation on my own pocket was around 70 dollars, but actual usage if considering free vertex credits would be 400-450 dollars ;-; |
| |
| ## Overview |
| A dataset of **~41,685 images**, curated through a maximum variance selection strategy. Rather than collecting images in bulk, each image was chosen to contribute unique tag coverage — ensuring that all **74,000+ booru tags** have at least one representative image in the dataset. |
| The result is a dense, high-signal dataset where every sample carries informational weight. |
| |
| ## Key Properties |
| | Property | Value | Value | Value | |
| |---|---|---|---| |
| | Total images | ~41,000 | |
| | Unique booru tags covered | 74,000+ | |
| | Captioned by | Gemini 3.7 Flash (~37,000 images) | Grok 4.6 (~4,000 images) | Qwen 3.8 27B FP8 (296 images) | |
| | Selection strategy | Maximum variance (diversity-maximized) | |
| |
| ## Curation Methodology |
| The dataset was built around a **coverage-first principle**: |
| - Every one of the 74K booru tags is represented by **at least one image** |
| - Images were selected to **maximize diversity** across the tag space, avoiding redundancy |
| - The goal was the smallest possible dataset that still spans the full breadth of booru tag semantics |
| This makes Booru Essence well-suited for tasks where **tag coverage and concept diversity** matter more than raw volume. |
| |
| ## Intended Use |
| - Research on furry & anime/illustration style recognition and diffusion training |
| - Tag classification and multi-label learning |
| - Concept coverage benchmarking |
| - Low-resource or compute-efficient training scenarios |
| |
| ## License |
| This dataset is released under [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/). |
| |
| But how do you licence a computer-generated dataset? |
| |
