DeepDanbooru

Why we included this project

Anyone building a tagger for anime or illustration collections will find DeepDanbooru one of the more established options. It is a multi-label classifier that estimates hundreds of tags from a single image in one pass, trained on the tagging metadata the Danbooru community maintains. Everything runs from the command line: you can fetch the current tag list, filter a SQLite-annotated image database, train a model on your own corpus, and evaluate single images or entire folders. Training your own model means you are not locked into a fixed pretrained tagger, which matters when your collection uses different vocabulary or conventions. The project also ships Grad-CAM visualization tooling, so you can see which parts of an image drove a tag decision before you rely on the result.