Michael Limberger
Need me? Email mike@limberger.ca
AI
Danbooru Tagging System
A tagging language, not a sentence
Danbooru is a big anime image board that has been cataloging and tagging artwork for years. Every image gets standardized tags: character details, art style, composition, everything.
When people trained AI models on this data, the models learned to connect specific tags with visual stuff. Those tags became a programming language for making images.
Why tags beat a paragraph
Danbooru tags are precise. Instead of writing "a young woman with long flowing blonde hair wearing a red dress," you break it into chunks the model actually recognizes:
Show me.
1girl, long hair, blonde hair, red dressEach tag maps to something the model learned. It combines them. Way more control than natural language.
Kinds of tags
| Kind | Examples |
|---|---|
| Character count | 1girl, 1boy, 2girls, solo, multiple girls |
| Physical traits | blonde hair, blue eyes, tall, muscular |
| Clothing | dress, suit, armor, school uniform |
| Poses | standing, sitting, arms crossed, looking at viewer |
| Expressions | smile, serious, crying, surprised |
| Backgrounds | outdoors, classroom, forest, simple background |
| Meta and quality | masterpiece, best quality, highres |
Tags versus plain English
Some models (especially photorealistic ones) like natural language better. Anime models and most SDXL checkpoints expect tags.
Mix and match works too:
1girl, detective, woman in her 30s, noir style, standing in rainThe model grabs what it knows and figures out the rest from context.
Order is weight
Tags up front carry more weight.
Quality tags first, to set the baseline. Character count early, to establish who is in the shot. Physical stuff next, general (age, skin) to specific (hair, eyes). Clothing, expression, and pose come after.
Tags right after a character description bind to that character.