SAM 3D Body is a promptable model for single-image full-body 3D human mesh recovery that estimates pose and shape of the human body, feet, and hands from a single photograph.
The tool addresses the challenge of reconstructing accurate 3D human mesh from single images in diverse real-world conditions. It uses an encoder-decoder architecture and accepts auxiliary prompts such as 2D keypoints and masks to guide inference, similar to the SAM family of models. The model represents human anatomy using the Momentum Human Rig, a parametric mesh representation that separates skeletal structure from surface shape to improve accuracy and interpretability. Training employed a multi-stage annotation pipeline combining differentiable optimization, multi-view geometry, and dense keypoint detection.
Developers working on human pose estimation, 3D reconstruction, or body modeling applications should consider this tool. It is designed for single-image inference in varied in-the-wild conditions and supports user-guided prompting, making it suitable for applications requiring flexible, interactive mesh recovery. The repository provides trained model checkpoints, datasets, and example notebooks to facilitate adoption.
The project maintains active engagement with its user base, with nearly all open issues originating from external adopters rather than the core team, indicating substantial real-world usage. Maintainers respond to new issues and pull requests within a day. Documentation improvements dominate the issue tracker, reflecting ongoing effort to support users implementing the model.
