SAM2Matting: Generalized Image and Video Matting

Despite impressive advances in image matting, video mattingremains challenging due to the inherent gap between high-level tracking,which requires frame-wise understanding, and low-level matting, whichfocuses on extremely fine-grained details. Existing methods attempt thisusing costly domain-specific video matting datasets, which may limittheir out-of-domain generalization and leave tracking robustness vulner-able. We rethink this paradigm with SAM2Matting, a novel frameworkthat decouples the task by enhancing the foundational tracker of SAM2with a region-proposal bridge and dedicated matting heads. This en-ables uncompromised SAM2 to handle tracking while the matting com-ponents focus exclusively on resolving fine-grained intricate details. No-tably, despite being trained only on images, SAM2Matting establishesnew state-of-the-art performance on video matting, with robust general-ization across both human-centric and in-the-wild matting scenarios.