Common booru tags: a reference glossary
The short answer
Most of what a WD14 tagger writes comes from a few hundred everyday Danbooru tags. They cover subject counts, framing, pose, hair, eyes, clothing, setting, lighting and style. This glossary groups the most useful ones and gives each a plain-English meaning, its usual mix-ups and the number of Danbooru posts behind it. File-quality tags such as lowres and absurdres are not in the model's vocabulary, so a tagger never outputs them.
Tagger output is much easier to check when you know exactly what each tag claims. cowboy_shot has nothing to do with cowboys, v is a hand gesture, and portrait means a specific crop rather than "a picture of a person". Use this page to decide quickly whether a tag fits an image and whether it belongs in your caption.
How to read the tables
The Posts column shows how many Danbooru posts carried each tag when the tag list for the ConvNeXt V2 and SwinV2 taggers was built. That count is a rough measure of the training data behind a tag. Of the 6,947 general tags, only 266 have 100,000 posts or more, and the median has just 2,735. Most tags here sit far above that. Danbooru's live counts have grown since, so treat these numbers as a snapshot.
Tags are shown in Danbooru form, with underscores. ImgTagger writes them with spaces (upper body), except for a few emoticon tags covered below.
People and subjects
Count tags usually lead tagger output because the model is most confident about them. Our guide to how Danbooru tags work explains the counting system. These are the subject tags people most often get wrong.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
1girl | 4,225,150 | One female character | Groups count separately: a mixed pair is 1girl plus 1boy. |
solo | 3,515,897 | Exactly one character | Wrong if anyone else is visible, even far back. |
solo_focus | 246,572 | Several people, one clear subject | Use it instead of solo for crowds and bystanders. |
male_focus | 371,503 | Only or mainly male characters | An emphasis tag, not a count. |
no_humans | 87,715 | No characters at all | Never belongs next to a count tag. |
animal_focus | 7,469 | An animal is the subject | A pet portrait is no_humans plus, say, cat (53,145). |
Framing and camera angle
Framing tags say how much of the figure is in the frame. Angle tags say where the camera sits. A crop that appears in most of your images but is never tagged tends to become a trained model's default.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
portrait | 35,713 | Mostly the face, cut off around the chest | Tighter still is close-up (28,288). |
upper_body | 434,670 | Roughly waist up | Most often confused with portrait. |
cowboy_shot | 336,374 | Thighs up, face included | A Western film framing that showed the holster. |
full_body | 463,008 | Head to feet | Cropped feet make it feet_out_of_frame (73,883). |
from_side | 122,971 | Camera sees the subject side-on | Camera position. profile (78,791) describes the face. |
from_behind | 150,284 | Camera behind the subject | Often with looking_back. |
from_above / from_below | 64,438 / 54,898 | Camera looking down / up | For a clear angle, not a slight tilt. |
dutch_angle | 89,835 | The whole frame is tilted | Look for a sloping horizon. |
Gaze, expression and pose
These tags vary from image to image even within one character's dataset, so they are some of the most useful to keep in a caption. Emoticon-style tags such as :d and ^_^ are real tags. ImgTagger keeps the underscore in the 13 emoticon tags where it is part of the face, so ^_^ stays ^_^.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
looking_at_viewer | 2,089,971 | Eyes on the camera | The gaze. facing_viewer (31,063) is the head and body. |
looking_back | 208,023 | Looking over the shoulder | Usually with from_behind. |
closed_eyes | 455,512 | Both eyes shut | A wink is one_eye_closed (294,831). |
smile | 1,903,619 | Smiling | Pairs with open_mouth or closed_mouth. |
:d | 387,533 | A wide, open-mouthed smile | Means ":D". Tag names are lowercase. |
blush | 2,040,471 | Reddened cheeks | Often drawn out of habit, so it appears on calm faces too. |
v | 84,806 | The two-finger V sign | Easy to delete by mistake as a typo. |
head_tilt | 102,674 | Head tilted sideways | Not dutch_angle, where the frame tilts. |
Hair
Hair is where the vocabulary is most detailed. Danbooru draws fixed length lines. Short hair ends above the shoulders, medium is shoulder length, long runs to the waist, and very long goes past it. Check borderline lengths first.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
short_hair | 1,568,265 | Above the shoulders | Chin-length bobs sit on the line. |
medium_hair | 170,863 | About shoulder length | About a ninth as many examples as short_hair. |
long_hair | 2,982,517 | Below the shoulders, to the waist | The third most common general tag. |
very_long_hair | 622,920 | Past the waist | Implies long_hair, so expect the pair. |
bangs | 1,516,840 | Hair over the forehead | Danbooru has retired it because nearly everyone has bangs, but the model still predicts it. |
hair_between_eyes | 692,014 | A lock between the eyes | Frequent, rarely important. |
blonde_hair / brown_hair / black_hair | 1,098,200 / 1,072,209 / 981,413 | Hair color | Neighboring shades swap in dim or tinted light. |
hair_ornament | 939,495 | A decorative clip or accessory | hairclip (224,901) implies it. Bows don't count: that's hair_bow. |
Eyes and face
Every image is shrunk to 448×448 pixels before the model sees it. In a full_body picture the eyes may be only a few pixels across, so check eye colors on wide shots by hand. Swaps almost always happen between neighboring colors.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
blue_eyes / aqua_eyes / green_eyes | 1,225,129 / 120,894 / 584,783 | Iris color | Aqua is the in-between shade. |
red_eyes / pink_eyes / purple_eyes | 897,316 / 168,674 / 536,623 | Iris color | Warm lighting pushes them around. |
brown_eyes / black_eyes | 645,874 / 136,390 | Iris color | Dark eyes in shadow drift between them. |
yellow_eyes / orange_eyes | 447,582 / 103,149 | Iris color | Gold or amber can land on either. |
heterochromia | 67,128 | Two different eye colors | Both colors get tagged too. |
Clothing basics
Danbooru tags clothing at two levels, generic (shirt) and specific (white_shirt), and taggers output both. Our tag pruning guide covers when to collapse them.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
shirt / white_shirt | 1,001,030 / 435,867 | A shirt / a white one | Generic and specific arrive together. |
skirt | 1,100,620 | A skirt | A skirt with a separate top is not a dress. |
dress | 838,290 | A one-piece dress | Belts and aprons can make it read as two garments. |
school_uniform | 585,679 | Any school uniform | serafuku (250,901) is the sailor-collar style. Blazer uniforms add blazer (50,589). |
long_sleeves | 883,900 | Long sleeves | Accurate but rarely the point. A common pruning candidate. |
bow / ribbon | 795,194 / 793,922 | A bow or ribbon anywhere | Placement tags like hair_bow imply the generic one. |
holding | 732,899 | Something held in the hand | Comes with specifics like holding_umbrella (17,024). |
Background and setting
ImgTagger requires a score of at least 0.5 before writing simple_background, white_background or grey_background, instead of the usual 0.35. It also flattens transparent areas onto white, because the model takes an ordinary color image with no transparency channel.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
simple_background | 1,074,818 | A plain or minimal background | Covers flat colors and gradients alike. |
white_background | 874,291 | A plain white background | Cut-out PNGs land here after flattening. |
transparent_background | 61,514 | A truly transparent background | Rare in ImgTagger output. Add it by hand if needed. |
outdoors / indoors | 320,641 / 197,331 | Outside / inside | Plain backgrounds usually get neither. |
scenery | 34,026 | The setting is a main subject | Not just "has a background". |
flower | 411,874 | Flowers anywhere | Includes flowers worn in the hair. |
blurry_background / depth_of_field | 73,931 / 74,158 | Out-of-focus background / shallow focus | Close cousins. One is usually enough. |
Lighting and time of day
Lighting tags have far fewer examples, so treat them as softer signals. Danbooru draws careful lines here. sunset needs the sun visible and an orange or red sky. evening is late afternoon. twilight is the glow while the sun is just below the horizon, before sunrise or after sunset.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
day / night | 243,849 / 74,665 | Time of day, when evident | Indoor shots often get neither. |
sunset | 21,004 | Setting sun, orange or red sky | About four times the examples of evening (4,487) or twilight (5,171). |
backlighting | 24,291 | Light from behind the subject | Look for a bright rim around the hair. |
light_rays | 18,405 | Visible shafts of light | Sun shafts specifically are sunbeam (8,334). |
sparkle / light_particles | 86,798 / 30,146 | Decorative light effects | Prune them if they don't matter to you. |
Style and medium
Style tags describe how an image was made. ImgTagger holds monochrome, greyscale and sketch to a strict 0.9 confidence. A wrong monochrome misdescribes the whole picture, so missing a few true ones is the better trade.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
monochrome | 478,584 | Shades of one color | Sepia is monochrome but not greyscale. |
greyscale | 377,967 | Shades of grey only | Implies monochrome, so they travel together. |
sketch | 112,175 | A rough drawing | Clean, unshaded lines are lineart (8,325). |
traditional_media | 61,046 | Physical materials | Specifics include watercolor_(medium) (8,390). |
comic / 4koma | 411,524 / 77,795 | Panels / a four-panel strip | A fast way to find pages to drop. |
chibi | 151,305 | Small, big-headed proportions | Keep it if you mix chibi and normal art. |
realistic | 16,156 | Realistic anatomy and lighting | Not a photograph. |
Text, signature and blur tags
These tags flag things you may want to crop out or clean up before training. Some trainers keep them in captions so the model ties watermarks to a tag it can avoid later. Either way, they make affected images easy to find with the Filter panel in the ImgTagger auto-tagger.
| Tag | Posts | Meaning | Watch for |
|---|---|---|---|
blurry | 140,047 | The image is out of focus | Not the deliberate blurry_background. Needs 0.9 in ImgTagger. |
signature | 183,546 | A prominent signature | Any written artist name is artist_name (200,583). |
watermark | 64,472 | Text or a logo laid over the image | Faint ones are easy to miss. |
twitter_username | 150,476 | A social media handle | Usually crops off cleanly. |
letterboxed | 43,050 | Bars above and below | Crop them so the model doesn't learn them. |
Tags the tagger can never output
Some of the best-known booru tags are missing from the vocabulary entirely. lowres, highres, absurdres and jpeg_artifacts are not in any of the three tag lists ImgTagger uses, so no threshold will produce them.
On Danbooru these are meta tags, which record facts about the file rather than what the picture shows. The tag lists hold only rating, general and character tags. The first three are added automatically from pixel size: lowres when both sides are 500 pixels or less, highres from 1600 wide or 1200 tall, and absurdres from 3200 wide or 2400 tall. The model couldn't judge size anyway, because every image is resized to 448×448 first.
- If you want resolution tags, add them yourself. ImgTagger shows each image's width and height in the preview pane. Select the images that qualify and use the Add Tags panel.
- For compression damage, judge by eye. Badly compressed images are usually better replaced than tagged.
- If these tags appear in your captions, they came from you or from imported .txt files, not the model.
To control how many tags reach each caption, see choosing a confidence threshold. To check which vocabulary you are using, see WD14 tagger models compared. To turn tags into training captions, continue with the LoRA dataset captions guide.
Drop in one picture and get its Danbooru-style tags in seconds, ready to copy into a prompt or a caption. It runs entirely in your browser. Open the Anime & Danbooru Tagger.
Frequently asked questions
Why does the tagger output both a generic tag and a specific one?
Danbooru applies both, and many specific tags add their parent automatically. The model learned from that data. Some people drop the generic tag to shorten captions. Keeping both is fine too, as long as you are consistent.
Are tags with low post counts unreliable?
Not necessarily, but they deserve a closer look. They were learned from far less data and are more easily confused with a common neighbor. Every tag in the list had at least 600 posts. Rarer tags were left out entirely.
Can I add tags that are not in the model's vocabulary?
Yes. Captions are plain text, and the vocabulary only limits what the model can predict. Use the Add Tags panel to add a trigger word, resolution tags or anything else.