xinyu1205/recognize-anything

Recognize Anything is a family of open source image recognition models built to tag what appears in a photo automatically, without a person having to label anything by hand first. The project actually contains three related models released over time, and the README presents them together as one evolving line of work rather than three separate tools. The newest is RAM++, which can recognize both common everyday categories and unusual, open ended ones it was never explicitly trained on. Before that came the original RAM, an image tagging model accepted at a CVPR 2024 workshop that recognizes thousands of common categories with high accuracy while remaining cheap to reproduce since it trains on an open, annotation free dataset. The earliest model, Tag2Text, accepted at ICLR 2024, goes a step further than plain tagging: it uses the tags it detects to guide automatic captioning, producing more detailed and controllable image descriptions, and can also use those tags to help match images to text during search. According to the README's own comparisons, RAM++ outperforms other leading foundation recognition models on both common and rare tag categories, as well as on phrases describing interactions between people and objects. The tag vocabulary itself grew substantially across versions, from around 3,400 categories in Tag2Text to more than 6,400 in RAM. The models have also been combined with object localization tools, Grounding DINO and SAM, in a related project called Grounded-SAM, to build a pipeline that can recognize, detect, and precisely outline objects in an image together. People can try the models directly through a Hugging Face web demo or a Colab notebook without installing anything locally. This project is aimed at computer vision researchers and engineers who want a ready made, strong tagging or captioning model to build on, rather than casual users looking for a simple photo labeling app.