Intro

Here's a sampling of some of the tags you can use in generation and/or training for an influence on style and a short summary of their effects. I also include any notes I have on if they're useful in generation, training (e.g., to minimize inherent style), or both, though these are mainly anecdote based.

Had these notes written out written out for a while (June 2025) and realized I'd never get around to making the full article since I don't know enough art to effectively test or describe the art term heavy stuff that are also well used Danbooru tags and should produce a useful effect (e.g., "ligne claire", or "art nouveau"), but figured the first part would still be useful as a beginner's guide (if you do have the art background to do it, I encourage you to write an article about these kinds of tags).

The Tags

Anime screencap

Makes an image look "anime", but this is overwhelmingly based on broadcast anime and thus includes a lot of “animation shortcuts” like lower detail in lighting and design (with face and hands in non-closeups are particularly vulnerable to degredation, needing addetailer or regional regen more often). The tag “anime screenshot” is often used in its place because Danbooru has renamed "anime screencap" to that after Illustrious was trained, but I've got no idea if it's actually effective or just read as "anime". Using it in training might isolate the shortcuts, but if you're training off only screencaps (instead of a mix of it with e.g., official art and/or fanart) wouldn't recommend it for characters (but might be useful for poses, clothing etc.).

Official Art

Art by a work’s creators from outside the work itself, typically for promotional or merch purposes, though development art and even just random sketches by the author (typically creator owned IP: Art by an official artist can still be non-official when dealing with corporate owned IP) are often included in this tag as well. In practice the overwhelming majority of the training data with this tag was for anime and Japanese video games which means this makes things “anime” styled, but with certain “higher quality” traits like shading that get skipped in anime screenshot. It's very solid and I'd recommend it over anime screencap for "anime" in most cases. When training I prefer to use it for ("anime" style) images that do qualify for it (or were by official artist in official style outside of official capacity) and are actually high quality “exemplar” images (rather than just sketches) so those will disproportionately influence the output when "official art" is use.

1980s (style), 1990s (style), 2000s (style), retro artstyle

Indicates specific vintage of art, or imitations thereof. As with many tags, the nature of the dataset means that these will generally be strongly anime influenced. "Retro artstyle" is broad enough its better to use the decade tags when applicable. “1970s (style)” and older (back to 1920s) are also Danbooru tags but the usage is significantly less so its effect will be less dramatic. These can be used in training with vintage animation to containerize inherent style (especially useful if tape is the best quality for the source material so the inherent style will have artifacts), preferably with at least some more modern art mixed in and not tagged this.

Toon (Style)

Style of western animation. Due to nature of using Danbooru as a datasource for base model training, most of the data is of parodies (generally by Japanese artists, though "western" cartoons of at least the 80s onward were outsourcing animation to Japan or the same Asian countries Japan was itself outsourcing to anyways) rather than actual western animation. IMO this comes off a bit generic and fake when used for "generic" characters (those made with only tags), but it can help push character or style LoRAs that were trained off this a bit closer to being accurate. I’d only bother using this tag in training if you have multiple style types and want to isolate this, as most such characters come off very "generic" or "wrong" when freed of their style.

Western comics (style)

Like the above, the nature of boorus mean this is generally made of parodies of the style by eastern artists, with the occasional drawing of a Japanese character by a western artist. The usage rate on Danbooru is rather low, so the effect may not be very pronounced. I’d only use it for training if the style is the exception among data, but I doubt it has much effect.

Comic, speech bubble, blank speech bubble, dialogue box

Not a style, but so frequently part of certain styles I should mention them. Comic is used for literal comics with multiple scenes in separate panels (for Danbooru images, generally short ones that fit in a single image), but I’ve found it effective for isolating issues with comic artifacts like panel dividers when there’s no way to crop those out. Likewise whiting out (or, occasionally, blacking out) a speech bubble’s text and using “blank speech bubble” lets you use pictures with those effects without issue. I don’t expect any of these to be useful for generation unless a LoRA calls for them however.

Pixel art

Text to image is unable to follow the rules needed to make this totally accurate without extensions (that the on-site generator doesn’t have access to), but some style LoRAs can get fairly close. I don’t expect this to work well without a dedicated LoRAs to refine it since it encompasses so many styles and formats. Works well when used in training to isolate the style (even fairly crude sprites have produced surprisingly good results for me in LoRA making when other styles are called in generation), and should be applied to all such images, both for styles and characters as I don’t expect untagged data in this style to be good for the resulting LoRA.

Chibi, chibi only

A version of a character with a small body, and disproportionately large (a third of the total height or more) head, often with some details like fingers ommited (for hands ending in stub). This works well for both generation and training, but with one fatal flaw: Both Pony and Illustrious (maybe SD1.5) have an annoying tendency to make copies of characters spawn. I've had some (but not great) improvement with using the tag "chibi only" over chibi. Another “must use” on applicable training data for characters that are normally not chibi, though I’m not sure if it’s needed for all chibi styles or specific characters who are inherently chibi.

3D

Makes character made of polygons, like CGI cartoons or a video game character. This one really only works in generation if you’re using a LoRA (character trained with it, or a style focused on it) and is very checkpoint dependent, but it will work. This should always be used on applicable images in training, even if it’s the entirety of the data set, or else it will contaminate everything and give nasty results (but with tagging its fine). The tag “low poly” is often used alongside this, but does not implicate it as it includes 2D drawings of low poly forms so both are needed. “cel rendering” is a particular style of 3D rendering, but like low poly is often used for 2D images imitating that style. I’d strongly recommend using it (and low poly or cel rendering as applicable) if you have to use such data for any kind of model.

Photo (Medium), photorealistic

Images made from photos, and images intended to look like one respectively. Note that photos of artificial objects (e.g., anime figurine), and frames of live action motion pictures (for poses/clothes/objects/location/etc.) can be "photo (medium)", while "photorealistic" isn't actually a photo. I don’t really have experience or commentary generating these. Using these tags on all applicable data when when training with photos can help avoid everything made with the LoRA having a weird pseudo-realism look.

Realistic

Despite what the name might imply at first, the "realistic" tag is actually for what in text to image is generally called "semi-real": An image with realistic proportions, lighting, etc. but still unmistakably an illustration (though one possibly directly based on reality). Something like Mona Lisa, Girl with a Pearl Earring, Norman Rockwell art, or official portraits of heads of state would easily qualify as "realistic" but never be mistaken for a photograph. Use "photo (medium)" and/or "photorealistic" for something intended to look like a photograph. Some video game 3D models (mostly seventh gen onward, though close ups in some sixth generation titles and a few pre-rendered scenes from fifth gen games could get it) might qualify for this tag, but they are first and foremost "3D" and I don't know if stacking the two for training is worth it.

Monochrome, greyscale

Technically a style. Rarely used in generation as a positive unless paired with a style LoRA, though a common negative. Most manga (outside of cover pages and some special pages) is monochrome and greyscale. Occasionally you’ll run into monochrome images in a non-greyscale, for these use “[color] theme”, with “blue theme” being the most common (note that these tags don’t necessarily have to be monochrome). Sometimes an otherwise monochrome image has a small part (such as a character’s eyes or a single piece of clothing) given color, this is the tag “spot color” used on top of it. These should be used when training with applicable data, though remember to adjust/drop any tags relating to color accordingly.

Monochrome background, greyscale with colored background, 3d background, photo background

A few styles isolated to the background alone. Likely useful in training (I use when applicable), but no comment on general use.

Line art

A subtype of monochrome images with no shading. I can see reasons to use this in generation, but hardly a “daily driver”. Since internal example/reference artwork (e.g., settei) generally falls under this tag you’ll often find cause to use it in training.

Flat color

Images with no shading or color variance within a block of color. Works well in generation, though I understand one's choice of sampler is also a big part in how shading is applied. Seems useable in training to isolate shading of older color comics without minimizing their composition. Doesn't seem to add more shading when used as a negative prompt.

No lineart

Images without outlines and no or extremely minimal inner lines. Unlike flat color, often regularly uses shading to break up shapes. Works well in generation from my tests, though I've only ever had one LoRA where it composed any meaningful number of images in training data.

Sketch

A low detail image often made with a single art tool (e.g., pencil or pen). A common negative and I doubt people will be using it much in generation without using a very specific style LoRA. I wouldn’t recommend using this kind of image in training data if you can help it, but when you have to, this seems to isolate the issues fairly well.

Traditional media

Tag for stuff made with non-computer tools then scanned (or occasionally photographed). In practice, the main effect this has is to enhance the “texture” or “grain” of an image. A lot of specific styles actually have their own tags (check the Danbooru wiki entry for this tag) that are worth using when applicable. I think it’s worth applying this and/or the subtag to training data when applicable.

Oekaki

Images from a class of web based drawing tool with relatively simple features. Their sharp, thin lines and limited detail have a distinct appearance. Haven't tested generation with it much, though it seems to really only add the jaggy linework. Tagging these images in training seems to isolate their style, though I haven’t done a lot with them. Beware Civitai trainer’s auto-label is very prone to false positives here, identifying random images as this so be sure to clear them.

Misc

Cover images for this article

If you care, the cover image is simply a composite of the images in this post. I've since added a few more tests.

Changelog

Feb 6 2026: Changes to intro. Clarified realistic vs photorealistic vs. photo (medium)
Feb 7 2026: Various small changes.
Feb 21 2026: Added flat color, no lineart. Fixed some typos in changelog. Added some tests to the linked example gallery above. Altered formatting so there will be a table of contents. Some changes to intro.