Booru tags

The WD14 tag list, by the numbers

The short answer

The WD14 v2 tag list behind ImgTagger's default models has 9,083 tags: 6,947 general, 2,132 character and 4 rating tags. The tagger can output only those. Every tag had at least 600 Danbooru posts when the list was built, and about a fifth had fewer than 1,000. Series that took off after 2022, artists and file-quality tags such as highres are not in it, and about 5.5% of its tags have since been renamed or retired on Danbooru.

A WD14 tagger can only output tags that appear in its tag list, a fixed file named selected_tags.csv. No threshold or setting adds a tag to it. That makes the list the real boundary of what the tagger can say about an image, so it is worth knowing what is inside.

This guide counts it. It covers the list used by ImgTagger's two v2 models, ConvNeXt V2 (the default) and SwinV2, which share one byte-identical file. Every figure below was computed with a script from two files we ship: that pinned CSV and a snapshot of Danbooru's tag data taken on October 15, 2026. The numbers describe the vocabulary, not how well the model uses it.

The list at a glance

The file has four columns: tag_id, name, category and count. The category column holds one of three codes, and the count is the number of Danbooru posts that carried the tag when the list was built.

CategoryTagsShare of listMedian postsMost-posted tag
General6,94776.5%2,7351girl (4,225,150)
Character2,13223.5%1,362hatsune_miku (78,616)
Rating4under 0.1%n/an/a
Total9,083

The four rating rows are the content ratings general, sensitive, questionable and explicit. Danbooru has five tag categories, but the list has only these three. There are no artist, copyright or meta tags, so the tagger never outputs an artist or a series as a tag of its own. Our Danbooru tags guide explains the categories.

The older ViT model has a smaller list: 6,549 tags, the four ratings plus 6,545 general tags, with no characters. The v2 list has 2,560 tags that ViT lacks (all 2,132 characters and 428 general tags), and ViT has 26 that v2 lacks. The model comparison has the details. The rest of this guide covers the v2 list.

Where the edges of the list come from

The ConvNeXt V2 model card gives two limits, and the SwinV2 list is the same file. Training used Danbooru images up to post ID 5,944,504, and tags with fewer than 600 images were filtered out. Danbooru's own API shows that post 5,944,504 was created in the final seconds of December 31, 2022 (US Eastern time). So the vocabulary is frozen at the end of 2022 and has a hard floor of 600 posts. The smallest count in the file is exactly 600, a character tag, and 334 tags sit between 600 and 699. These two limits explain the gaps described later.

The long tail: most tags are rare

Counts across the 9,079 general and character tags run from 600 to 4,225,150, and the bands below show how uneven that is.

Posts when builtGeneralCharacterBothShare
1,000,000 or more180180.2%
100,000 to 999,99924802482.7%
10,000 to 99,9991,420711,49116.4%
1,000 to 9,9994,0841,3835,46760.2%
Under 1,0001,1776781,85520.4%

To see how top-heavy the list is, add up the count column (each post counted once for every tag it carries). The 128 most-posted tags, 1.4% of the list, make up half of that total. The top 845 tags (9.3%) make up 80%, and the top 1,939 (21.4%) make up 90%. The other 7,140 tags, 78.6% of the list, share the last tenth.

In practice, a few tags have millions of posts behind them, while a tag in the bottom fifth has under a thousand. Review the rarer tags more closely, because they usually rest on less data. The confidence threshold guide covers how to be stricter about them.

Character tags are less top-heavy than general ones. Half of all general tag uses come from 1.7% of the general tags, but it takes the top 13.3% of character tags to reach half of character uses. Still, no character tag has 100,000 posts, and 678 of the 2,132 (31.8%) have under 1,000.

The counts have aged

The count column dates from when the list was built. Across the 8,547 tags with a live count in our October 2026 snapshot, the median tag now has about 2.2 times the posts the CSV records (2.4 times for general tags, 1.4 for characters), and 1girl went from 4,225,150 to 8,545,948. The order held up, with a rank correlation of 0.91 between old and new counts. Use the counts to compare tags with each other, not as current totals.

What the general tags look like

General tags are short. Counting the pieces between underscores, 2,110 of the 6,947 (30.4%) are one piece, like smile. Another 3,940 (56.7%) have two, like long_hair, 738 (10.6%) have three, and only 159 (2.3%) have four or more.

Counting tags end at 6+

The list has 16 tags for the number of people: 1girl to 5girls, 1boy to 5boys, 1other to 3others, and a 6+girls, 6+boys and 6+others bucket for bigger groups. An image with eight girls and an image with six get the same tag.

Hair and eye colors

There are 12 basic color tags for hair and 12 for eyes. The sets differ by one color: hair has blonde_hair where eyes have yellow_eyes. The spread is wide. blue_eyes has about 164 times as many posts as white_eyes (7,469).

Hair tagPostsEye tagPosts
blonde_hair1,098,200blue_eyes1,225,129
brown_hair1,072,209red_eyes897,316
black_hair981,413brown_eyes645,874
blue_hair564,360green_eyes584,783
pink_hair427,100purple_eyes536,623
purple_hair426,385yellow_eyes447,582
white_hair407,226pink_eyes168,674
grey_hair375,103black_eyes136,390
red_hair338,682aqua_eyes120,894
green_hair286,703orange_eyes103,149
orange_hair147,142grey_eyes98,579
aqua_hair82,429white_eyes7,469

Beyond these, it has three shade tags (light_brown_hair, light_blue_hair and dark_blue_hair) and tags for mixed colors: multicolored_hair, two-tone_hair, gradient_hair, streaked_hair and heterochromia. Alternate spellings are not listed, only Danbooru's chosen name. On October 18, 2026, its alias list mapped yellow_hair and gold_hair to blonde_hair and gray_hair to grey_hair, and none of those three old names is in the tag list.

Qualifiers for words with two meanings

Of the 1,165 tags with a parenthetical qualifier, 155 are general tags that separate words with several meanings: star_(symbol) and star_(sky), bow_(weapon), arrow_(projectile).

What the 2,132 character tags look like

Character tags are 23.5% of the list but only 3.3% of its total tag uses, because even the biggest ones are small next to 1girl. The first-ranked hatsune_miku has 78,616 posts, and 16 of the top 20 characters are from Touhou. Names follow Danbooru's pattern: a character with an unambiguous name usually has no qualifier, as in hakurei_reimu, while an ambiguous one carries its series.

The split is almost even: 1,010 character tags (47.4%) carry a qualifier and 1,122 do not. The most common qualifiers are these:

QualifierCharacter tags
kancolle (Kantai Collection)234
fate141
pokemon58
azur_lane53
genshin_impact45
arknights45

Reading the final parenthetical on each of them (one tag with a nested qualifier is skipped) gives 155 distinct qualifiers. A few, like female, are not series names. Only 14 appear on ten or more tags, and 89 appear exactly once. Use the character tagger search box to check any name. It matches words in any order and also matches series, such as kancolle.

What the list cannot contain

  • Anything that took off after 2022. Honkai: Star Rail was released internationally on April 26, 2023, and no tag in the list contains star_rail. When we checked on October 18, 2026, Danbooru had 9,792 posts for march_7th_(honkai:_star_rail) and 8,212 for kafka_(honkai:_star_rail). Both are far above the 600-post floor today but came too late for the list. The tagger can describe such a character's hair, eyes and outfit but cannot name them.
  • Anything under 600 posts at the time. Niche characters, props and poses below the floor were left out, even where Danbooru has a tag for them today.
  • Artist, copyright and meta tags. lowres, highres, absurdres and jpeg_artifacts are not in the list. See the end of the tag glossary for why.

The list reflects Danbooru as a whole, which includes mature material, and it has tags for that too. This guide's examples stick to everyday tags. To keep particular tags out of a dataset, use the blacklist described below.

Renamed and retired tags

Danbooru keeps editing its tags, and the list does not change. Our snapshot asked Danbooru about the 9,079 general and character tags on October 15, 2026.

CategoryTagsLive post countAliasedDeprecatedNo live count
General6,9476,529176 (2.5%)218 (3.1%)26
Character2,1322,01894 (4.4%)15 (0.7%)5
Both9,0798,547 (94.1%)270 (3.0%)233 (2.6%)31

Two general tags are in both the aliased and deprecated lists, so the rows add up to two more than the totals. Counted once, 501 tags (5.5%) have been renamed or retired. The 31 in the last column came back with no live count and no flag. We have not checked why.

The two terms mean different things. An alias makes Danbooru treat two names as one tag: its help page says an alias redirects searches for the alternate name to the main tag. A deprecated tag is, per Danbooru's deprecation notice, locked and cannot be added to new posts. The tagger can still output either kind, because its vocabulary stopped changing in 2022.

  • Aliased examples: hand_on_hip (100,931 posts in the CSV) is now hand_on_own_hip, hairpin is hairclip, doughnut is donut and china_dress is qipao.
  • Character renames are often small: annie_leonhardt became annie_leonhart, and ahri_(league_of_legends) became ahri_(league). Of the 94 aliased characters, 58 involve a series qualifier on one side or the other.
  • Deprecated examples: bangs is the most-posted deprecated tag (1,516,840 posts in the CSV). Danbooru's wiki says the base tag is no longer used because it would apply to almost every post. striped, black_footwear, uniform, gradient and plaid are deprecated too. Seven deprecated tags had 100,000 or more posts.

Whether this matters depends on your use. For your own training captions, consistency usually counts for more than current spelling, so keep one set of names across the dataset. To paste tags into a booru or search one, you want the current names. The Danbooru tagger does that swap for you: its "Use Danbooru's current tag names" option is on by default, and with it on, retired tags are left out of the text. The batch auto-tagger writes the names from the list, with no renaming.

Using this when you tag

  • Look up characters before you tag. If a name is not in the 2,132, the tagger cannot give it, whatever the threshold. The character tagger search is the fastest check.
  • Blacklist what you don't want. The batch tagger's blacklist is comma-separated and matches whole tags, ignoring case and underscores. Adding general, sensitive, questionable, explicit drops the four rating tags.
  • Be stricter in the tail. The default threshold is 0.35, and a few tags have built-in minimums (monochrome, greyscale, sketch and blurry need 0.9). On the quick taggers, character tags need 0.85. Raise the threshold if you see rare tags on pictures that do not fit.
  • Prune the common ones. The head of the list is full of generic tags. Our tag pruning guide shows what to cut.
Put this into practice

Run WD14 tagger models on a whole folder, then search, filter and bulk-edit the tags before exporting .txt caption files. It runs in your browser. Open the Batch WD14 Tagger & Editor.

Frequently asked questions

How do I get a tag that is not in the list into my captions?

Add it yourself. The list only limits what the model can predict, not what a caption can hold. In the batch tagger, the Prepend box places a tag first on every image you tag, and the Add Tags panel adds one to the selected images or to all of them, at the start or the end. That works for a new character, a trigger word or resolution tags.

Where can I download the tag list?

It is the file the tagger loads: selected_tags.csv from the ConvNeXt V2 model repository, pinned to the commit ImgTagger uses. No tag name contains a comma, but one row quotes its name because the name itself contains quotation marks, so a CSV library is safer than a hand-written split. The SwinV2 file is identical.

Is the post count the number of images the model trained on?

Treat it as a rough guide. The count is the number of Danbooru posts with the tag when the list was built. The model card says images with fewer than 10 general tags were also filtered out, and we cannot tell from the files how that filter relates to the counts. The smallest count, 600, does match the card's cut-off for tags.

Will ImgTagger update to a newer tag list?

Not automatically. Each model loads from a fixed commit, so its list cannot change without a site update. Newer v3 taggers exist, but ImgTagger does not offer them yet.