Danbooru Tag Taxonomy DB

DATASET: https://huggingface.co/datasets/Mexes/danbooru-tag-taxonomy

WEB TOOL: https://huggingface.co/spaces/Mexes/danbooru-tag-taxonomy

BOORU PROMPT GALLERY: https://booru-prompt-gallery.vercel.app/

TL;DR: I released the tag categories behind Booru Prompt Gallery as a free, open source dataset. 85,042 Danbooru tags, each one sorted into a category (Appearance, Clothing, Equipment, Pose, Scenery, Creature, Other) and one of 33 slots (hair, top, footwear, weapon, expression, setting...). It comes with a small web tool where you can paste a tag or a whole prompt, see it sorted, and export it clean. CC BY 4.0, use it for whatever you want.

Hello! If you've used Booru Prompt Gallery, you've already used this data without knowing it. It's what lets the app copy a single category, merge posts, and build packs without putting two outfits on the same character. Now you can use it in your own tools too.

What is this?

Danbooru splits its tags into 5 types: general, character, copyright, artist and meta. The problem is that almost everything useful is "general". long_hair, boots, sitting, classroom, holding_sword... all "general". Danbooru doesn't tell you if a tag is about the hair, the outfit, the pose or the background.

This dataset adds exactly that. Every tag has a category and a slot, and the main idea is simple:

Tags in the same slot can usually replace each other.

boots and sneakers are both clothing:footwear; smile and crying are both pose:expression. If you swap one for the other, the rest of the prompt still makes sense.

• Appearance: hair, eyes, anatomy, subject (1girl, solo...)

• Clothing: outfit, top, bottom, underwear, headwear, eyewear, legwear, footwear, accents

• Equipment: weapon, handheld

• Pose: camera, posture, gesture, expression, kinetic, combat, interaction, physiological

• Scenery: setting, props, atmosphere

• Creature: animal

• Other: character, copyright, artist, style, layout, meta

Clothing is by far the biggest category.

Why did I make it public?

The categories took a lot of work, and they were only living inside my app. I think they can be useful for a lot more than Booru Prompt Gallery,

I'm sure people will come up with better uses than mine.

The Web Tool

If you don't want to touch any code, I made a small web page for it. Paste a single tag, a list of tags or a whole prompt, and every tag gets sorted by category and slot.

Things it handles:

• Prompt weights and brackets: (white sundress:1.2) works fine.

• Danbooru aliases: holding spear shows as → holding polearm, which is the real tag on Danbooru.

• Unknown combinations: white sundress is not a Danbooru tag, but sundress is. The tool guesses the slot from the shorter tag and marks it with a dashed border and "≈", so you know it's a guess and not a confirmed label.

• Tags that aren't in the dataset: quality tags like masterpiece aren't Danbooru tags, so they go to their own box instead of disappearing.

Export

At the bottom there's an Export box that gives you everything sorted and ready to copy, in two formats:

• Sorted tags, one line: subject first, weights kept. Handy for cleaning up messy prompts: paste something you wrote in a hurry and get it back in a consistent order.

• Grouped by category and slot: one line per slot, like the image below.

Everything runs in your browser, nothing you type is sent anywhere.

How was it made?

Sorting tags into categories couldn't be done by magic (I said that in the app article too, and it's still true):

1. Tag data: names, aliases and post counts come from Danbooru's public tag data.

2. AI classification: general tags were classified with Jev, an AI model from TypeSafe made for this kind of judgment call. If Jev was less than 70% sure, the tag was not approved and was set aside for humans.

3. Rules: character, copyright and meta tags were assigned by rule. A character tag is always other:character, no AI needed.

4. Cross-check: everything was compared against two public sources, Danbooru's own tag group wiki and the Jio7/danbooru-tags-classified dataset. Where both agreed, the tag was approved; where they disagreed with Jev, Jev was asked again.

5. Human fixes: tags reported as wrong in Booru Prompt Gallery, and the ones you classified with Quick Teach, were fixed by hand. Thank you to everyone who helped!

This is how sure Jev was about each general tag. Everything left of the dashed line is waiting for a human (more on that below).

What's inside?

• Character: 42,616

• General: 32,125 (the ones that actually needed classifying)

• Copyright: 9,841

• Meta: 460

Limitations

I want to be honest here, this is not perfect:

• One slot per tag: some tags fit in two. holding_sword is equipment:weapon, but it's also a gesture.

• Not every label was checked by a human: while preparing this release I went through some of the smaller slots and still found mistakes, like a gun brand and a Japanese votive plaque sitting in the "artist" slot. (They're fixed now, but there are surely more.)

• 8,036 general tags are still unclassified: Jev wasn't sure, so I preferred to leave them empty instead of guessing.

• Post counts are not live: they're from when the tags were imported.

• English Danbooru tags only.

How to use it

If you don't code: the dataset includes a CSV file you can open in Excel or Google Sheets and filter by slot.

If you do, it's one line:

from datasets import load_dataset

ds = load_dataset("Mexes/danbooru-tag-taxonomy", split="train")

Releases are tagged, so you can pin a version with revision="v1.0" and your tool won't break when I update the labels.

License: CC BY 4.0. You can use it for anything, including commercial projects, as long as you give credit. The tag names come from Danbooru, so please credit them too.

FAQ

Is this the same data the app uses?

Yes, it comes from the same database. When a tag gets fixed in the app, it will be fixed in the next version of the dataset.

Found a tag in the wrong slot?

Leave a comment with the tag and where you think it belongs, and I'll fix it.

Want to help with the unclassified tags?

Those 8,036 tags are exactly the ones Quick Teach shows you in Booru Prompt Gallery. Every answer you give there ends up in the next version of the dataset. It's a few keypresses per tag, and it helps both the app and anyone using the dataset.

Think a slot is missing?

Probably! For example, jobs (nurse, photographer, mangaka) don't have a clear place right now. Suggestions are welcome.

Built something with it?

Please share it in the comments, I'd love to see it.

Want to support me?

Thank you so much! You can Buy Me a Coffee, or send Buzz. It also helps a lot if you:

• Leave a comment and a rating.

Other articles that may interest you:

• Web App: Booru Prompt Gallery V10.0 | Civitai

That's it!

Stay hydrated and don't forget to blink.