Booru Tags to Natural Language - Open Source Dataset

DATASET: https://huggingface.co/datasets/Mexes/booru-tags-to-natural-language

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

HOW THE CONVERTER WORKS: https://civitai.red/articles/37453/tags-to-natural-language-how-booru-prompt-gallery-writes-prompts-for-flux2-krea-2-and-moss-image

TL;DR: I released the data behind Booru Prompt Gallery's Tags to Natural Language feature as a free, open source dataset. It has 26,302 Danbooru and Rule34 tags, each with a short English phrase and the place it takes in a sentence (wariza → "sitting with her legs splayed to the sides"). For acts between two people it also says who does what. It also has 4,633 character profiles that say how each character is usually drawn. CC BY 4.0, so use it for whatever you want.

Hello! Booru Prompt Gallery v10.4 can show a card's prompt as plain sentences for FLUX.2, Krea 2 and MOSS-Image, or as tags plus a short description for Anima. It doesn't run any AI when you use it, just a set of rules that read this data. I explained how they work in another article.

What is this?

The dataset has two parts.

1. Phrases

Booru tags are great for searching, but many of them don't read as English. Think of wariza, gendou_pose, :3 or cowboy_shot. Others do read as English, but need a verb or an article before they fit in a sentence.

This table gives each tag a phrase and a kind. The kind says which sentence the phrase continues, so a program knows where to put it.

• is: "She is ___" → looking_at_viewer → looking at the viewer

• with: "..., with ___" → hand_on_own_hip → one hand on her hip

• has: "She has ___" → thick_eyebrows → thick eyebrows

• wears: "She wears ___" → wrist_cuffs → wrist cuffs

• framing: "The image is ___" → cowboy_shot → a cowboy shot from the thighs up

• scene: light and atmosphere → dappled_sunlight → dappled sunlight

In the data, "her" is written as {pos}, so the same phrase works for her, his or their.

Put a few together and you get "She is sitting, looking at the viewer, with one hand on her hip."

There are 15,711 Danbooru tags and 10,591 Rule34-only tags (ones the Danbooru taxonomy doesn't have). A source column tells them apart, because the Rule34 ones are less curated.

Most hair and clothing tags have no phrase, because long hair and black skirt already read as English once the underscores become spaces. Most phrases are for poses, gestures, expressions, the body, accessories, the camera and the atmosphere. There are also phrases for clothing tags that aren't something you wear, like bare_shoulders or jacket_on_shoulders, and for tags the taxonomy never sorted.

Acts and poses also say who they belong to when there are two or more people. The who column says whether the phrase describes the one who does the act, the one it's done to, both of them, or the person with a certain body, and partner says what body the other person needs, so a converter can write who does what.

2. Character profiles

With two characters in a prompt, the tags are still one flat list, so in white hair, black hair, ponytail, sword there's no way to tell who has what.

For each character, the profiles give the share of their Danbooru posts that have each hair, eye, body, clothing and equipment tag.

hatsune_miku has twintails in 83% of her posts. If a prompt has Miku and someone else, the twintails are probably hers.

Each profile also has a gender signal, which is how many of the character's solo posts are tagged 1girl or 1boy. I use solo posts because in a group picture the 1girl is often someone else.

Rule34 and Danbooru often spell the same character differently, so there's also a character map with 2,461 Rule34 character tags and their Danbooru names. With it, Rule34 posts can use the same profiles.

Why did I make it public?

The phrases and the character profiles took a lot of work, and until now they only lived inside my app. I think they can be useful for a lot more than Booru Prompt Gallery, and I'm sure people will come up with better uses than mine.

How was it made?

Phrases

1. Candidates: tags from the Danbooru Tag Taxonomy and the Rule34 Tag Taxonomy, mostly in the slots that need rewording, plus tags the taxonomy never sorted. For those, the AI picked the slot too.

2. Writing and review: one AI model wrote a kind and a phrase for each tag in plain, neutral words, describing only what the tag shows. A second one checked every entry against what the tag means on its site and rewrote the ones that were off. The review column says which.

3. Unclear tags: tags marked as not visual or too ambiguous were left without a phrase, because I'd rather have no phrase than a wrong one.

4. Who does what: Jev, TypeSafe's classification model, labeled the acts and poses with who and partner.

5. Real posts: I ran the converter on real posts and fixed by hand the phrases that came out wrong. Those rows have review set to by_hand.

Rule34 tags come from a taxonomy no person has checked, so a Rule34 tag can have a fine phrase and still sit in the wrong slot. If you only want the most reliable part, keep the rows where source is danbooru.

Character profiles

1. Characters: the 3,000 most used character tags on Danbooru, plus the Danbooru names of popular Rule34 characters.

2. Related tags: from Danbooru's related-tag lists, I kept the hair, eye, body, clothing and equipment tags that appear in at least 10% of the character's posts.

3. Character map: Rule34 character tags were matched to Danbooru names, keeping only clear matches (same words in another order, a longer full name, another way of writing the series...) and official Danbooru aliases.

What's inside?

Most phrases are is (poses, actions, expressions), then with and scene.

Each character profile has about 45 tags.

Limitations

• AI wrote the phrases. No person read all 26,302, and I only fixed the ones I ran into. Some will be wrong or clumsy, mostly for rare tags.

• A phrase describes one tag. Joining phrases into a natural sentence is left to the converter, including merging long hair and black hair into "long black hair" and choosing the order.

• Tags that come from a partner. A profile counts every post the character appears in, including posts with other people. So a male character picks up tags from the girls he's drawn with. kamado_tanjirou has long_hair at 49% and breasts at 11%. Check the gender columns before giving a character a body or clothing tag.

• Profiles mix outfits. A character with several costumes gets the clothing tags of all of them, each at a middling share.

• Name mix-ups. The character map can still pair two characters with very similar names.

• Snapshot. The shares and post counts are from the date of the release.

• English tags only.

Content warning: Danbooru's vocabulary includes tags for nudity and sexual content, and most Rule34 tags are about sexual content. The phrases describe them in plain, neutral words like any other tag. The dataset contains text only, no images.

How to use it

If you don't code, the dataset includes CSV files you can open in Excel or Google Sheets.

If you do, it's a couple of lines.

from datasets import load_dataset

phrases = load_dataset("Mexes/booru-tags-to-natural-language", "phrases", split="train")
traits = load_dataset("Mexes/booru-tags-to-natural-language", "character_traits", 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 data.

The slot column uses the slots of the Danbooru and Rule34 tag taxonomies, so you can join them by tag.

The license is CC BY 4.0. You can use it for anything, including commercial projects, as long as you give credit. The tag names and the numbers behind the profiles come from Danbooru and Rule34, so please credit them too.

FAQ

Found a wrong phrase?

Leave a comment with the tag and what it should say, and I'll fix it in the next version.

My favourite character isn't there

The profiles cover the most used characters. If one you need is missing, tell me in the comments.

Where are the Rule34 tags and their categories?

In their own dataset, the Rule34 Tag Taxonomy. It's separate from the Danbooru one because an AI sorted those tags and no person checked them.

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:

• Tags to Natural Language: how Booru Prompt Gallery writes prompts for FLUX.2, Krea 2 and MOSS-Image | Civitai

• Open Source Dataset: Rule34 Tag Taxonomy | Civitai

• Open Source Dataset: Danbooru Tag Taxonomy | Civitai

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

That's it!

Stay hydrated and don't forget to blink.