NerissG/booru-essence-v2

Image mirror for lodestones/booru-essence-v2.

Dataset selection and annotations are by lodestones the original author. This repository provides the image files. All credit for creating and curating the dataset goes to him (and for giving me the few images I was missing <3).

This redistribution follows the original dataset’s CC BY-NC-SA 4.0 license.

This upload contains all the dataset. The danbooru subset is roughly ~356.8 GiB, and the e621 is ~433.0 GiB. For a total size of 789.8 GiB.

It is packaged as uncompressed, indexed tar archives. For both easy bulk downloads and individual-image retrieval.

Repository layout

Images are grouped by the first two characters of their MD5 hash, then split into numbered archives.

<subset>/
├── manifest.json              # Archive metadata, image counts, sizes, and SHA-256 checksums
└── images/
    ├── 00-00000.tar           # Image archive
    ├── 00-00000.json          # Image index (offsets, sizes, MD5, SHA-256)
    ├── 01-00000.tar
    ├── 01-00000.json
    ├── ...
    ├── ff-00000.tar
    └── ff-00000.json

Each image inside the TAR is named <md5>.<extension>.

Retrieving individual images

Offsets point directly to the image data within the uncompressed tar.

You can retrieve an image without downloading or extracting its entire archive:

1. Take the image_md5 from the booru-essence-v2 parquet.
2. Find the matching two-character prefix in manifest.json.
3. Search that prefix’s archive JSON indexes for the matching MD5.
4. Request the image’s byte range from the corresponding tar file with Range: bytes=<offset>-<offset + size - 1>

Example request:

To download this image:

// danbooru/images/10-00000.json
{
  "filesize": 1638383104,
  "hash_lfs": "f9481cebbdbcc5d4304ccb11d8d2bd904c9b7ba56f018926de4cfa9a44645bc8",
  "files": {
    // ...
    "1004f20da299e7277d0c8e16be2a1fc3.jpg": {
      "offset": 44211712,
      "size": 186721,
      "md5": "1004f20da299e7277d0c8e16be2a1fc3",
      "sha256": "3a6828e6cf186cb6b1f254109d66cff6029c79e8c8ed1f7f3ba1a1c42cb877f9"
    },
    // ...
  }
}

You can run the following curl request:

offset=44211712
size=186721
range="$offset-$((offset + size - 1))"

curl --fail --location --dump-header - \
    --header "Range: bytes=$range" \
    --output comrade.jpg \
    "https://huggingface.co/datasets/NerissG/booru-essence-v2/resolve/main/danbooru/images/10-00000.tar"

This should respond with 206 Partial Content, with a Content-Range header matching the requested range. And correctly save the image to your disk.

---

Booru Essence v2

A highly condensed, diversity-maximized image dataset distilled from booru-style image boards — now with 25 representatives per tag and full Qwen3.8-27B caption annotations.

the images can be obtained here (WIP — not uploaded yet) Clybius/booru-essence-v2-images

Overview

Booru Essence v2 scales up the original Booru Essence coverage-first strategy. Where v1 selected a single image per tag greedily, v2 keeps up to 25 samples per tag — 516,964 images spanning 170,925 unique booru tags, evenly drawn from Danbooru and e621, cleaned and with sketches excluded.

Every image is annotated by Qwen3.8-27B at maximum reasoning effort (xhigh), producing a four-level caption chain, a structured scene breakdown, and the full reasoning trace — built for distillation into smaller captioning models.

Key Properties

Property	Value
Total images	516,964
Unique booru tags covered	170,925
Selection strategy	Maximum variance, 25 samples per tag (v1: 1)
Sources	Danbooru (258,558) + e621 (258,406)
Caption model	Qwen3.8-27B, xhigh reasoning effort
Annotation success	99.9% (516,287 / 516,964)
License	CC BY-NC-SA 4.0

Annotation Schema

Each row carries four captions of decreasing fidelity, plus structured scene data:

Column	Content
cap_long_caption	Single-paragraph description, 200–350 words
cap_midjourney_style_summary	Midjourney-style prompt, 50–85 words
cap_brief_summary	Brief summary, 35–55 words
cap_10_word_summary	Exactly 10 words
cap_scene_json	Structured scene: subjects, characters, species, setting, composition
reasoning	The model's full reasoning trace (a deliverable, for distillation)
image_md5	Content hash of the source image (joins to the image archive)
status	ok for 99.9% of rows; filter status == 'ok' to drop truncated/refused

The caption bands are disjoint and enforced monotonic (long > midjourney > brief > 10-word) on 99.7% of rows. Original booru metadata (tags, rating, fav_count, artist/character/species weights, …) is preserved alongside.

Curation Methodology

Same coverage-first principle as v1, deepened rather than broadened:

• Every tag keeps up to 25 representative images instead of 1, giving each concept enough samples to train on rather than merely be present
• Selection still maximizes diversity across the tag space
• Cleaned and de-sketched upstream (tag_samples_clean_nosketch_25)

Intended Use

• Caption-model distillation (four granularities + reasoning traces per image)
• Research on furry & anime/illustration style recognition and diffusion training
• Tag classification and multi-label learning
• Concept coverage benchmarking

License

This dataset is released under CC BY-NC-SA 4.0.

• ✅ Free for non-commercial use
• ✅ Modifications and derivatives allowed
• ✅ Attribution required
• ❌ Commercial use not permitted
• 🔁 Derivatives must carry the same license