Mexes/danbooru-tag-taxonomy

Danbooru Tag Taxonomy

112,570 Danbooru tags labelled with a category and a slot from a fixed taxonomy of 7 categories and 33 slots, for example clothing:footwear, pose:expression or scenery:atmosphere. The labels come from booru-prompt-gallery, which uses them to colour tags, score prompts and generate prompt variations.

Danbooru marks most descriptive tags as general and does not say what part of the image they describe. This dataset adds that. Tags that would replace each other in a prompt share a slot: boots and sneakers are both clothing:footwear. Possible uses:

• grouping or colouring tags in prompt editors and autocomplete,
• wildcard and prompt randomizers that change one slot at a time,
• filtering or balancing captions for image model training,
• checking which parts of an image a caption describes.

Try it in the browser: the Danbooru Tag Taxonomy Space sorts a tag, a list of tags or a whole prompt by category and slot.

Snapshot: 2026-10-02.

Files

File	Format
tags.parquet	Parquet, aliases as list<string> (read by the dataset viewer and datasets)
tags.csv	UTF-8 CSV, aliases comma-joined in one quoted field

from datasets import load_dataset

ds = load_dataset("Mexes/danbooru-tag-taxonomy", split="train")
general = ds.filter(lambda r: r["danbooru_category"] == "general" and r["status"] == "approved")

Each snapshot is tagged (v1.0, ...). Pass revision="v1.0" to load_dataset to pin one.

Columns

Column	Type	Description
tag	string	Danbooru tag name, underscores kept (long_hair)
danbooru_category	string	Danbooru's tag type: general, copyright, character, meta (artist tags are excluded)
category	string	One of the 7 categories below
subcategory	string	Slot within the category; unclassified while pending review
slot	string	category:subcategory in one column
confidence	float	Classifier confidence (0–1); 1.0 for rule-derived rows, hand-set rows and rows where both external sources agree
status	string	approved or needs_review (see below)
post_count	int	Danbooru post count at the snapshot date, never lowered between snapshots (0 when unknown)
aliases	list<string>	Danbooru aliases that resolve to this tag, plus Danbooru's current name for renamed tags

Taxonomy

Counts are approved general tags per slot.

Category	Covers	Slots
appearance	Physical traits like eye color, hair style, skin tone	hair (564), eyes (279), anatomy (2,381), demographics (272)
clothing	Attire, accessories, and footwear	outfit (3,989), top (1,334), bottom (497), underwear (517), headwear (1,366), eyewear (311), legwear (368), footwear (358), accents (2,161)
equipment	Weapons, handheld props, and character items	weapon (1,448), handheld (3,376)
pose	Body position, gestures, actions, and angles	camera (254), posture (462), gesture (516), expression (413), kinetic (246), combat (210), interaction (628), physiological (1,013)
scenery	Background, location, and environmental elements	setting (1,070), props (1,295), atmosphere (610)
creature	Animals, mythical beasts, and pets	animal (808)
other	Who and where a tag comes from, art style, layout and image metadata	character (440), copyright (776), artist (29), style (624), layout (290), meta (207)

Non-general Danbooru tags are assigned by type: copyright → other:copyright, character → other:character, meta → other:meta.

Composition

Danbooru type	Tags
general	40,525
copyright	12,581
character	58,840
meta	624

• approved: 101,157 rows
• needs_review: 11,413 rows, all general tags parked in other:unclassified because the model's best slot scored below 0.7 or no slot fit. Drop them unless you plan to label them yourself.
• 29,112 of 40,525 general tags have a slot; 7,342 of those have confidence = 1, the rest are model labels at or above 0.7.

How it was built

1. Tag names, aliases, types and post counts come from Danbooru's public tag data.
2. General tags were labelled with Jev, TypeSafe's classification model, using the taxonomy definitions as context. Labels below 0.7 confidence were kept as needs_review. Copyright, character and meta tags were assigned by rule from their Danbooru type, and general tags of the form X_(cosplay) (dressed as character X) are clothing:outfit.
3. Labels were cross-checked against two public sources: Danbooru's tag-group wiki, via danbooru-tag-groups-scraper, and Jio7/danbooru-tags-classified. Pending tags were approved where both sources pointed to the same slot. Where only Jio7 gave a category, Jev chose again among that category's slots. Approved tags that a wiki group placed in another slot were asked again with only the two slots as options, and changed when Jev picked the wiki's.
4. Labels reported as wrong in the app, or fixed in its admin review queue, were corrected by hand. Hand corrections replace the model's label.
5. Each snapshot adds the tags Danbooru gained since the previous one (25 posts or more) and refreshes post counts and aliases.

Tag names

Tags keep the name they had when they entered the dataset, even after Danbooru renames or deprecates them. Models such as Illustrious and Anima were trained on those names, so china_dress stays china_dress and Danbooru's current qipao is listed among its aliases. A renamed tag's post_count is that of the tag it now resolves to. Counts are never lowered, so a tag Danbooru has emptied or deprecated keeps its last count (black_footwear stays at 287K). Danbooru searches resolve aliases, so renamed tags still find the same posts.

Limitations

• Each tag has one slot. Some tags fit two: holding_sword is labelled equipment:weapon, although it also describes a gesture.
• post_count is a snapshot. Query Danbooru for current counts.
• Recent tags, such as characters from recent games, are newer than most image models and are often not recognized by them.
• Most rows with confidence < 1 were labelled by the model and not checked by hand, so some slots are wrong.
• English Danbooru vocabulary only.

Excluded tags

Danbooru artist tags are not included. Their slot would only repeat the Danbooru type, and the dataset is not meant to be a list of artist names.

Content warning

Danbooru's vocabulary includes tags for sexual and other adult content, and they are labelled like any other tag. The dataset contains tag names only, no images.

Related datasets

• Jio7/danbooru-tags-classified: Danbooru tags in 14 flat categories labelled by an LLM. Used here as a second opinion.
• deepghs/site_tags: tag lists, types and aliases for Danbooru and other sites, without semantic categories.
• Danbooru's tag groups, exported monthly as JSON by danbooru-tag-groups-scraper: hierarchical groups curated by the Danbooru community. Used here to check and fill slots.

This dataset differs in giving each tag one slot, so tags of the same kind can replace each other, and in recording a confidence and a review status per tag.

License and attribution

The classification (category, subcategory, slot, confidence, status) is released under CC BY 4.0. Tag names, aliases and post counts originate from Danbooru. Credit both when you use it. The cross-check used danbooru-tag-groups-scraper and Jio7/danbooru-tags-classified, both MIT-licensed.

@misc{danbooru_tag_taxonomy,
  author = {Mexes},
  title  = {Danbooru Tag Taxonomy},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Mexes/danbooru-tag-taxonomy}
}