Danbooru tags, explained

If you have trained a LoRA, used an anime-style Stable Diffusion checkpoint or run an auto-tagger, you have already met Danbooru tags. They look strange at first: 1girl, looking_at_viewer, upper_body. They follow a small set of rules, though, and once you know those rules, tagger output becomes much easier to read, prune and trust.

This guide covers the parts of the system that matter when you are tagging your own images: what a tag is, the categories, the ratings, the naming conventions, and the habits that make booru tags work so well as training captions.

Where booru tags come from

"Booru" is the community's name for a family of image-board sites that organize images by tags instead of folders or captions. Danbooru is the best-known one. It has been running since 2005, and as of October 2026 it hosts over 12 million posts. Every post is tagged by volunteers following a written set of guidelines. Over many years that has produced a large, consistent and openly searchable label vocabulary.

That consistency is why machine-learning projects adopted it. The WD14 tagger models that ImgTagger runs were trained to predict Danbooru tags from pixels. Many anime-focused image generators were trained on images captioned with the same tags. When a model "speaks booru", it is because booru tags are what it saw during training.

Anatomy of a tag

A Danbooru tag is a single token with no spaces. The conventions are simple and strict:

• Lowercase only. blue_eyes, never Blue_Eyes.
• Underscores instead of spaces. On the site, the space character separates one tag from the next, so multi-word concepts are joined: looking_at_viewer.
• Parenthetical qualifiers tell apart names that would otherwise collide: star_(sky) versus star_(symbol), or a series name for characters, as in ganyu_(genshin_impact). In the tag list ImgTagger's default model uses, 1,165 of the 9,083 tags contain a qualifier of this form.
• Japanese name order (family name first) for Japanese characters and people, e.g. hakurei_reimu.
• Numbers for counts. 1girl, 2girls, 1boy and so on describe how many characters are visible.

The five tag categories

Every Danbooru tag belongs to exactly one category. The category tells you what kind of fact the tag records. That matters a lot when you decide which tags to keep in a training caption.

General tags are by far the largest group. They are also the only category a vision model can reliably learn from pixels alone. A model can see a smile. It cannot see who drew the picture, and it can only guess resolution from a resized copy. This shows in the WD14 tag lists. The ConvNeXt V2 and SwinV2 taggers predict 6,947 general tags and 2,132 character tags. The older ViT tagger predicts only general tags. None of them output artist, copyright or meta tags. That is why tags such as highres and absurdres never appear in ImgTagger results, even on huge images.

The four content ratings

Separately from tags, every Danbooru post carries one rating. Since 2022 there have been four:

• general: safe for any audience.
• sensitive: suggestive or mildly revealing, but not explicit.
• questionable: clearly not safe for work, but not explicit.
• explicit: explicit content.

Older documentation and older models mention a three-level "safe / questionable / explicit" scheme. The four-level scheme replaced it by splitting the old "safe" into general and sensitive.

WD14 taggers treat the ratings as four extra outputs, alongside the tags, in a category of their own. In ImgTagger they appear in the tag list like any other tag, so an image may come back with general at the front of its tags. Many people keep the rating tag in training captions so they can steer it at generation time. Others delete it. If you don't want it, add general, sensitive, questionable, explicit to the tagger's blacklist and it will never be written.

Counting tags and the "solo" family

One of the most useful booru conventions is how it records the number of characters. The pattern is:

• 1girl, 2girls, 3girls… and 1boy, 2boys… record exact counts by apparent gender. 1other exists for characters who don't fit either.
• multiple_girls and multiple_boys mean two or more. 2girls automatically implies multiple_girls.
• solo means exactly one character is in the image. no_humans means none, as in landscapes, still lifes and animal-only pictures.

These are some of the most common tags in the whole vocabulary. When the WD14 v2 tag list was built, 1girl was on 4,225,150 Danbooru posts and solo on 3,515,897. Because they are so frequent, models learn them very well. They are also strong levers in a training caption, so make sure they are right. A dataset where half the "solo" images actually show two people teaches the model the wrong thing.

Aliases and implications

Two mechanisms keep the vocabulary tidy, and both help you prune tags.

Aliases map an alternative spelling to the official tag. Someone typing grayscale on Danbooru gets greyscale. Aliased tags don't exist as separate labels, so a tagger never outputs them.

Implications say that one tag automatically adds another. Some real examples from Danbooru:

• very_long_hair → long_hair
• cat_ears → animal_ears
• 2girls → multiple_girls

Implications explain a pattern you will see constantly in tagger output: pairs of tags where one is a more specific version of the other. Whether to keep both is a judgement call. Some trainers drop the general parent tag to keep captions short. Others keep both because that is how the base model saw them during training. Our tag pruning guide walks through the trade-offs.

"Tag what you see"

The most important rule in Danbooru's tagging guidelines is often shortened to "tag what you see". Tags describe what is visible in the image, not what you happen to know about it. If a character normally has a hat but isn't wearing it in this picture, there is no hat tag. If a scene is set at night but the image doesn't show it, there is no night tag.

This rule is a big part of why booru tags make good training data. Each tag is a claim about the pixels, so a model can actually learn the link between the label and what it looks like. Keep the rule in mind when you hand-edit captions too. Adding tags for things that aren't visible, such as a story detail or a character's backstory, quietly teaches your model associations that aren't there.

Reading auto-tagger output with these rules in mind

Put it together and a typical tagger result becomes easy to read. Say an image comes back from ImgTagger with:

general, 1girl, solo, long hair, smile, looking at viewer, upper body, scarf, outdoors, snow, blurry background

You can now read this as:

• a rating (general);
• a count (1girl, solo);
• appearance (long hair, smile);
• composition (looking at viewer, upper body);
• clothing (scarf);
• setting (outdoors, snow, blurry background).

ImgTagger orders new tags by confidence, so the most certain predictions come first. That makes it easy to scan for errors near the end of the list, where the weaker guesses sit. The confidence threshold guide explains how to set the cut-off.

Where booru tags fall short

Booru tags are excellent for describing what is in an image. They are weak at describing relationships. 1girl, 1boy, red_shirt, blue_shirt doesn't say who is wearing which shirt. Tags also can't capture mood, story or exact spatial layout the way a sentence can. And because the vocabulary grew up around anime-style illustration, it has deep coverage of hairstyles and outfits but thinner coverage of, say, product photography or architecture.

That is why many modern pipelines mix tags with natural-language captions, or use one for some models and the other for others. If you are deciding between the two, read booru tags vs natural-language captions. If you want to see which everyday tags are the most useful in practice, the booru tag glossary groups the common ones by what they describe.