Meta

Meta Segment Anything Model 2 (SAM 2) is a segmentation model designed to identify and isolate any object within images and videos with speed and accuracy. It unifies object segmentation across both static images and dynamic video frames into a single model, eliminating the need for separate tools or workflows. Users can provide input via clicks, bounding boxes, or masks to select a target object on any image or video frame. Once selected, SAM 2 tracks the object across subsequent frames, allowing users to refine predictions by adding additional prompts such as points or boxes to correct errors or adjust boundaries. This capability makes it particularly useful for applications like video editing, where one might want to isolate a moving subject (e.g., a person, animal, or vehicle) across an entire clip without manually annotating each frame. The model also supports multi-object selection: users can choose several distinct objects in the same scene, and each object will be segmented and tracked independently. For image segmentation, SAM 2 can produce precise masks for complex shapes, including objects with fuzzy edges or partial occlusions. The model is released alongside a demo that lets users test it directly in a browser, as well as downloadable weights for local deployment. A research paper details the architecture and training methodology, which builds on the original SAM (Segment Anything Model) but introduces temporal consistency mechanisms for video. Key technical features include a promptable segmentation head that processes user inputs in real time, and a memory module that retains information across video frames to maintain object identities and handle appearance changes such as lighting shifts, motion blur, or partial disappearance. SAM 2 achieves this withou