tokenize-anything

Why we included this project

Teams building interactive image tools often need a model that handles several region-level jobs at once, and Tokenize Anything does them in a single pass. Give it a point, box, or sketch prompt, and it segments the area, identifies what it is, and returns a caption, which saves you from chaining a segmentation model to a separate captioner. The code is modular: predictors are decoupled from the model, and a registry lets you plug in your own inference logic instead of editing a monolithic script. Pretrained checkpoints in three sizes (ViT-B/L/H) plus concept weight files let you pick an accuracy/speed trade-off and swap encoders as your workload changes. It is research-grade software with notebooks for inference and evaluation, so expect to adapt it to your own data rather than treat it as a turnkey service.