Booru captioning is a specific style of image tagging used primarily for training AI models—like Stable Diffusion and Pony Diffusion—based on the structured, comma-separated metadata found on imageboard sites like Danbooru. Unlike natural language descriptions, Booru captions use a flat hierarchy of standardized tags (e.g., 1girl, solo, long_hair, blue_eyes) to help AI models precisely identify and replicate specific visual elements. Why Use Booru Captions?
This paper proposes Caption Booru, an open, privacy-aware platform for collecting, curating, and evaluating image captions at scale. Caption Booru combines moderated community contribution, automated captioning models, and structured metadata to create a searchable dataset for research and application in multimodal AI. We present system design, dataset schema, moderation policy, model-in-the-loop curation, evaluation methodology, and initial experimental results. Caption Booru
"First time," Elias said, sliding onto a stool. "I heard you can make anything real here. If you tag it right." Booru captioning is a specific style of image
The site’s real utility, however, lies in its rule structure. Caption Booru has notoriously strict posting guidelines: images must contain a caption, tags must follow a precise format, and certain content requires warning labels. This rigorous, volunteer-enforced system demonstrates how a community can maintain high quality and accessibility without corporate oversight. It is a working model of "self-governing digital commons," where usability (finding exactly what you want via tags) depends entirely on collective adherence to rules. "First time," Elias said, sliding onto a stool
Narrative Depth: Instead of just looking at a static character, the caption provides a "voice," transforming the viewer into a reader.