Segment Anything 2 (SAM 2)

Meta's Segment Anything Model 2 extends promptable segmentation from still images to video by treating an image as a single-frame video, so one model handles both modalities. Users point at objects with clicks, bounding boxes, or masks, refine results interactively, and track multiple objects through a clip; a transformer architecture with streaming memory keeps per-frame inference fast enough for real-time use. The SAM 2.1 checkpoints come in four sizes, from a 38.9 million parameter tiny model running at about 91 frames per second to a 224 million parameter large model at about 39, with results reported on the SA-V, MOSE, and LVOS benchmarks. Meta also released SA-V, a large video segmentation dataset built with a model-in-the-loop data engine, along with training code for fine-tuning on custom data. The model and code are open source under Apache 2.0, downloadable directly or through Hugging Face, and are widely used for video object tracking, dataset labeling, and as a segmentation backbone in larger vision pipelines.