SAM 3D Body is a promptable model for single-image full-body 3D human mesh recovery developed by Meta Superintelligence Labs. The repository provides inference code, trained model checkpoints, datasets, and example notebooks demonstrating how to use the model for reconstructing 3D human body meshes from single images. The model represents a component of the broader SAM 3D initiative, which also includes SAM 3D Objects for 3D shape reconstruction.

The core functionality of SAM 3D Body centers on estimating human pose and shape from a single image with state-of-the-art performance across diverse in-the-wild conditions. The model uses an encoder-decoder architecture and supports auxiliary prompts including 2D keypoints and masks, enabling user-guided inference similar to other models in the SAM family. The body, feet, and hands are estimated based on the Momentum Human Rig, a parametric mesh representation that decouples skeletal structure from surface shape to improve accuracy and interpretability. The training pipeline employed high-quality annotations from a multi-stage process incorporating differentiable optimization, multi-view geometry, dense keypoint detection, and a data engine designed to collect and annotate data covering both common and rare poses across varied viewpoints.

Two model checkpoints were released on November 19, 2025. The DINOv3-H+ backbone variant contains 840 million parameters and achieves 54.8 MPJPE on 3DPW and 61.7 MPJPE on EMDB, while the ViT-H variant with 631 million parameters achieves the same 54.8 MPJPE on 3DPW but 62.9 MPJPE on EMDB. Both checkpoints are available on Hugging Face along with their configuration files. The SAM 3D Body dataset was also released on Hugging Face, with instructions provided for downloading and processing the data.

The repository is written in Python and classified across multiple computer vision domains including 3D segmentation, human body modeling, deep learning, image segmentation, shape estimation, neural networks, promptable segmentation, and 3D reconstruction. According to GitGenius activity tracking, the repository has shown median issue and pull request response latency of 9.7 hours across 66 tracked items, with a mean latency of 137.1 hours. Documentation has been the most actively tracked issue label with 2 occurrences. The most active contributors tracked by GitGenius are noahcao with 26 events, xyang35 with 14 events, and carlosedubarreto with 13 events. The repository shares overlapping contributors with major projects including Microsoft Visual Studio Code, Microsoft TypeScript, and the Rust programming language repository.

The codebase includes a demonstration notebook at notebook/demo_human.ipynb showing complete inference and visualization workflows. Users can load models directly from Hugging Face following the installation instructions in INSTALL.md. The repository also provides integration examples demonstrating how to combine SAM 3D Body results with SAM 3D Objects output, aligning both reconstructions in the same frame of reference. The model checkpoints and code are licensed under the SAM License, with contribution guidelines and code of conduct documentation provided for community participation.