SAM 3D Body

Meta extended the Segment Anything line into human mesh recovery with SAM 3D Body, which lifts a single RGB image into a full 3D human, estimating body, hand, and foot pose on the Momentum Human Rig, a parametric representation that decouples skeletal structure from surface shape for more accurate fits. Promptability is the differentiator: 2D keypoints and segmentation masks steer the estimate toward a particular person in a crowded frame or correct a bad fit, the same interaction style as the image SAM models. Two checkpoints are published, an 840M DINOv3-H+ encoder and a 631M ViT-H, gated behind an access request on Hugging Face, and the November 2025 release reports 54.8 MPJPE on the 3DPW benchmark, ahead of CameraHMR and HMR2.0b. Code, a demo script, a notebook, and a hosted playground demo accompany a released training dataset. Everything ships under Meta's SAM License rather than a standard open-source license, permitting broad use with some restrictions, and the model has settled quickly into pose tracking, motion capture, and avatar pipelines.