sam-3d-body-dinov3: Full-body 3D mesh recovery from a single photo

A model that estimates a person's full-body 3D mesh from a single image. It comprehensively captures pose from the torso down to the feet and fingers, maintaining consistent accuracy even under occlusion, difficult poses, and刁nan viewpoints. Developed by Meta Superintelligence Labs, SAM 3D Body (3DB) is characterized by its strong generalization performance under in-the-wild conditions.

Unlike existing methods, it uses Momentum Human Rig (MHR), a new parametric mesh representation that separates skeletal structure from surface shape. This improves both accuracy and interpretability. Based on an encoder-decoder architecture, it accepts auxiliary prompts such as 2D keypoints or masks, allowing users to directly control the inference process.

The output includes 3D mesh vertices in the camera coordinate system, 3D and 2D pose keypoints, camera motion parameters, and focal length. It also returns pose parameters for the torso, hands, and body shape, making it easy to use for downstream processing. It is trained on large-scale, high-quality data collected through a multi-stage annotation pipeline, covering a variety of poses and viewpoints.

It is suitable for developers who require precise 3D human modeling for tasks such as 3D human pose estimation, motion capture replacement, or augmented reality (AR) content creation. It offers a prompt-based interface similar to the SAM family of models, making it easy to integrate into existing workflows. Access to the model requires agreement to Meta's Privacy Policy, and it is distributed under the SAM License.