quarterturn/danbooru-1024-eq-captioned

Danbooru 1024 e/q Captioned Dataset

This repository contains 59,495 anime-style images downsampled to 1024px on the longest edge, drawn exclusively from Danbooru's e (explicit) and q (questionable) rated highscore pool. Each image is accompanied by a structured JSON caption generated via OpenRouter MiniMax-M3 ($0.29/image total).

Structure

danbooru-1024-eq-captioned/
├── danbooru-1024-eq-captioned.parquet   ← metadata manifest (queryable)
├── train/                               ← raw images organized by batch subdirectories
│   ├── batch_0001/3211105.webp
│   ├── batch_0001/3472553.webp
│   └── ...
├── captions/                            ← JSON captions 1:1 with train/ files
│   ├── batch_0001/3211105.json
│   ├── batch_0001/3472553.json
│   └── ...
└── proxies/                             ← heavily pixelated versions of ALL images for safe HF preview
    ├── batch_0001/321105.jpg
    └── ...

About the Proxy Images

Every image in proxies/ is a globally blanket-pixelated JPEG copy of its corresponding raw file in train/. The pixelation process works by shrinking the entire image to 1/28th resolution then scaling back up with nearest-neighbor interpolation and saving at quality=40 JPEG. This produces a recognizable thumbnail-level preview while destroying all fine anatomical detail — including explicit body parts like breasts, genitals, nipples, and penises — regardless of character pose or angle.

The Dataset Viewer on HuggingFace points at proxies/ instead of train/. The raw unfiltered images only exist inside train/ for model training purposes after you pull the dataset locally. Proxy images are never intended to be used directly in any training pipeline and are provided solely as a safety mechanism so public dataset previews remain safe-to-browse without risking automatic takedowns or community policy violations from displaying uncensored explicit anime art at render scale on HuggingFace's UI.

Captions

Each JSON caption contains structured fields including:

• subject_and_action — detailed natural language description of the image content, characters, poses, clothing states
• characters[] — per-character breakdowns with visible_body_parts, state_of_dress, action, gaze direction
• style_description — color palette hex codes, mood tag, lighting notes, camera framing and depth-of-field
• post_processing — effects and filter tags (cel-shaded, wet looks, etc)

The parquet file consolidates all key metadata into a single queryable manifest including Danbooru artist names, source works, subject/character tags, ratings, fav counts, image dimensions, token usage stats, and pointers to both raw train/ paths and safe proxies/ preview paths.

Model Used

MiniMax-M3 via huggingface_openrouter with structured JSON prompt requiring explicit per-character state-of-dress reporting, visible body part inventories, camera/focus notes, captured text transcription when present, color palette sampling, mood/lighting tags, and post-processing effect descriptions for deep character comprehension during fine-tuning.

License & Notes

• Images are sourced from Danbooru under individual artist copyright. This dataset is provided for personal training/creative use only.
• All files inside proxies/ are derivative blur copies not intended for reproduction or inclusion in any trained model weights whatsoever — they exist purely as safe preview thumbnails against automatic flagging systems on HuggingFace viewer UI alone.

Statistics

Metric	Count
Total images	59,495
e (explicit) rated	~18K
q (questionable) rated	~41K
Average tokens/image	varies significantly depending on multi-character compositions
Total API cost	$106.92 via OpenRouter MiniMax-M3

Future Plans

Safe and s-rated image batches will be added in parallel after this e/q run completes fully, keeping ratings cleanly separated but following identical JSON caption structures for unified training pipelines downstream later on once everything merges together into one complete master dataset finally someday soon hopefully. ❤️