Meta's SAM 3D Body Turns a Single Photo Into a Riggable Human Mesh

Meta Superintelligence Labs has released SAM 3D Body, known internally as 3DB, on Hugging Face. The model reconstructs a posed, full-body 3D human mesh from one photograph and accepts 2D keypoints or segmentation masks as corrective prompts.

The release combines the model with Momentum Human Rig, or MHR, a parametric body representation that separates skeletal structure from surface shape. Developers can edit the resulting pose, retarget it for animation, or use it as a starting point for avatar, virtual fitting, visualization, and research pipelines.

A photo becomes a riggable body

Given a single-person crop from an ordinary image, 3DB predicts a watertight mesh, meaning a closed surface without gaps. The reconstruction includes the body, hands, and feet, with optional hand refinement when suitable hand crops are available. Meta designed the model to handle occlusion, unusual camera angles, and poses that challenge conventional human mesh recovery systems.

Optional keypoints and masks give developers control when the automatic reconstruction misplaces an occluded limb or ambiguous joint. The model converts these prompts into tokens and combines them with image features through cross-attention, allowing the supplied spatial hints to influence the recovered pose and shape without retraining.

MHR separates bones from shape

MHR supplies the parametric mesh underneath 3DB. A parametric mesh represents a body through adjustable values for pose, skeletal proportions, and surface shape instead of treating every vertex as an independent output.

Separating the skeleton from the body surface makes pose inspection, shape editing, animation, and retargeting more direct. That structure can reduce the conversion work required when reconstructed people move into avatar systems, animation rigs, augmented-reality tools, or biomechanics workflows.