Segment Anything 2

SAM 2 extends Meta's Segment Anything Model from images to video by treating an image as a one-frame video. Users prompt with clicks, boxes, or masks to select an object, and the model segments it and tracks the mask across frames using a transformer architecture with streaming memory, which processes video sequentially in real time instead of holding entire clips in memory. Four checkpoint sizes on a Hiera backbone run from tiny at 38.9M parameters to large at 224.4M. Alongside the model, Meta released SA-V, described as the largest video segmentation dataset to date, built with a model-in-the-loop data engine. The SAM 2.1 checkpoints improved accuracy further, with the large model reaching 79.5 J&F on the SA-V test set at about 39.5 FPS on an A100 GPU. Code and weights are open under the Apache 2.0 license, and the model is widely used for video object tracking, interactive annotation, robotics, and editing tools.