v1.0: 17k images, 6k tags. Added rating-separated versions. Deduped.

Try the random prompt generator with better visual experience here.

This is a set of wildcards based on real tags from images on Danbooru. The images were randomly chosen from artworks created by artists in a curated list. Each image matches at least the tags 1girl solo original rating:g,s.

The general tags are broken down into 4 rough categories:

• Character-specific: These are attributes that would remain consistent even if the depicted character were engaged in a different activity or wearing a different outfit. This category includes body attributes and hairstyles, but it excludes accessories.

• Outfit-specific: This pertains to what the character is wearing or not wearing. It encompasses anything that indirectly describes the outfit or implies the presence or absence of a particular outfit.

• Action-specific: This refers to the actions that the character is performing. It covers poses, expressions, and other actions, as long as they do not directly reference the outfit.

• Scene-specific: This relates to the image's setting, encompassing the location, background elements, composition, and essentially everything that does not fall under the other three categories.

Since there are only 4 categories, this set of wildcards is more coarse-grained than most others. Note that some tags are inherently ambiguous and tags from different categories may contradict each other.

The tags were first zero-shot categorized by GPT-3.5 Turbo with the definitions above, but the human ended up redoing ~15% of the tags. The Danbooru tag wiki is nice but sadly the tag groups aren't as comprehensive as desired.

Made to make testing models easier.

All sample images used the prompt template __characters__ , __outfits__ , __actions__ , __scenes__ ,(general variants for v1.0).