SAM 2 (Segment Anything Model 2)

SAM 2 generalizes SAM from still images to video. You prompt an object once, a click or box on one frame, and SAM 2 propagates its mask through the rest of the video, following the object as it moves, deforms, and is briefly occluded. It does this with a memory module that remembers the object across frames, and it still handles single images as a special case.

Like SAM, it is promptable and zero-shot, so it segments objects it was not specifically trained on, now across time. It was trained on the large SA-V video segmentation dataset and runs fast enough for interactive use.

You annotate an object on one frame and SAM 2 carries the mask forward, with the chance to correct it where it drifts. In FiftyOne, SAM 2 is available through the integrations, so you can propagate masks across video, store the results as tracked labels, and review them in the app.

SAM 2 attacks the most painful annotation cost in video: masking the same object frame after frame. The win is interactive, not just about accuracy. Because SAM 2 propagates from a single prompt, the annotator's job becomes spotting and fixing where the mask drifts, fast motion, occlusion, a similar object nearby, rather than drawing every frame. The leverage on video labeling is enormous, but the human review still concentrates exactly where tracking is hard, which is why correction tooling matters as much as the model.

SAM segments still images. SAM 2 adds video, propagating a mask across frames with memory.

Yes, its memory mechanism follows an object across frames from a single prompt.

Yes, it handles both images and video in one model.

Segment Anything Model (SAM), SAM 3, Click-to-segment, Video annotation, Object tracking.