Datasetpublic

quarterturn/danbooru-1024-eq-captioned

Danbooru 1024 e/q Captioned Dataset 59,495 high-resolution (1024px) anime-style images from Danbooru's explicit and questionable rated pools. Each image includes comprehensive JSON captions generated via MiniMax-M3 with structured per-character state-of-dress inventories, camera notes, mood palettes, and post-processing detections. Directory Structure danbooru-1024-eq-captioned.parquet <- consolidated metadata manifest originals/ <-… See the full description on the dataset page: https://huggingface.co/datasets/quarterturn/danbooru-1024-eq-captioned.

sourceHugging Facecc-by-sa-4.0updated 2mo agoView on Hugging Face

6likes22kdownloads

Dataset Card

Danbooru 1024 e/q Captioned Dataset

59,495 high-resolution (1024px) anime-style images from Danbooru's explicit and questionable rated pools. Each image includes comprehensive JSON captions generated via MiniMax-M3 with structured per-character state-of-dress inventories, camera notes, mood palettes, and post-processing detections.

Directory Structure

danbooru-1024-eq-captioned.parquet  <- consolidated metadata manifest
originals/                           <- original images (archived as tar.gz)
proxies/                             <- censored preview copies (jpg q=40) 
captions/                            <- structured JSON captions (1:1 with originals)

Components

`originals/` - Original Images (Archived)

Original unmodified Danbooru images organized into compressed tarball archives (batch_XXXX.tar.gz). Each archive contains approximately 200 webp/jpg files. These are for local downstream fine-tuning use.

To extract a batch:

bash

tar -xzf danbooru-1024-eq-captioned/originals/batch_0001.tar.gz
# ~200 image files now available in originals/batch_0001/

`proxies/` - Censored Preview Copies (Not for training)

Globally pixelated derivative JPEGs intended ONLY for public preview display. Generated via /28 shrink + nearest-neighbor upscale, saved as JPEG quality=40. These exist solely so the dataset viewer can show content without exposing uncensored originals. Do NOT use proxy images for model training.

`captions/` - Structured JSON Descriptions

One JSON file per image (59,495 total), containing rich per-character objects including: rating, subjectandaction, styledescription, stateofdress inventory, visiblebodyparts, mood tagging, and capturedtext OCR results wherever applicable during automated capture workflow.

`danbooru-1024-eq-captioned.parquet` - Metadata Manifest

Flat queryable manifest with 22 columns linking Danbooru artist names, character tags, source works, user score/favorites metrics, image dimensions, and token usage statistics across all records inside one unified file.

Statistics

Metric	Value
Total images	59,495
Explicit rated	19,367
Questionable rated	40,128
Resolution	1024px longest edge
Unique artists	10,881
Unique characters	5,044
Unique source works	3,367
Avg fav count	117.1
Avg score	66.9

Captioning Model

MiniMax-M3 via OpenRouter using structured JSON prompts with explicit per-character state-of-dress reporting and visible anatomical inventories.

License

Images remain under individual artist copyright as originally published on Danbooru itself. This dataset is provided for personal research offline usage purposes only. Proxy images are censored derivative copies intended exclusively for safe community browsing networks to prevent accidental exposure.

Donations

Donations in any amount are gratefully accepted to support future datasets and research costs.

Ethereum (ETH): 0x73a3fB9EB1fa259b7395F8cBBa3C1025453530d6