What happened
Meta’s SAM 3D Body repository describes 3DB as one part of SAM 3D, alongside a separate model for object reconstruction. SAM 3D Body estimates a person’s body, feet and hands from one image using the Momentum Human Rig, or MHR, a parametric mesh representation intended to separate skeletal structure from surface shape. The model uses an encoder-decoder architecture and accepts optional prompts such as 2D keypoints and masks. A hand decoder can further refine the hands.
The repository says the model was trained using high-quality annotations produced through a multi-stage pipeline involving differentiable optimization, multi-view geometry, dense keypoint detection and a data engine. Meta says this data engine collected and annotated examples covering common and rare poses across a wide range of viewpoints. The repository links to the SAM 3D Body dataset on Hugging Face and says users should follow its download and processing instructions. The project materials describe the dataset as containing around five million samples.
For inference, Meta provides installation instructions, example notebooks and a demo script. The documented workflow downloads assets from Hugging Face and runs the script over an image folder, with an output folder for results. The example uses a default ViTdet detector and MoGe2 field-of-view model; an alternate command selects SAM 3 as detector. The Python example loads the estimator from the Hugging Face repository, processes an image and writes a visualization. Access to the model checkpoints must be requested, according to the installation instructions.
The README reports results for two checkpoints released on November 19, 2025: DINOv3-H+ with 840 million parameters and ViT-H with 631 million. Its table lists, respectively, MPJPE values of 54.8 and 54.8 on 3DPW, and 61.7 and 62.9 on EMDB; PVE of 60.3 and 61.7 on RICH; PCK@.05 of 86.5 and 86.8 on COCO, and 68.0 and 68.9 on LSPET; and PA-MPJPE of 5.5 for both on Freihand. The code and checkpoints are licensed under the SAM License. The repository does not state a price.
What the archive adds
New Runtime’s earlier coverage on November 9, 2026 described Tencent’s DSpine draft-model port for Qwen3.8-27B, distinct from a community GGUF model (GGUF port of Tencent’s DSpine draft model for Qwen3.8-27B). That was an adjacent model-distribution story, not prior coverage of human mesh reconstruction. SAM 3D Body opens a new application thread; the shared theme is access to model artifacts beyond a hosted interface.
Trend signal
This release does not yet fit a pattern New Runtime tracks. It publishes an application-specific model and data workflow, but the available facts do not demonstrate a broader shift in agent interoperability, runtime security or model interfaces.
Why it matters
For teams working with human imagery, a single-image reconstruction model can make 3D body estimates available without building a full capture setup. The prompt support also gives a potential integration point: an upstream detector or annotation system can provide keypoints or masks to steer inference, while the mesh output can feed visualization or downstream geometry tools.
The practical value will depend on more than the benchmark table. Teams need to assess whether their images resemble the evaluated datasets, how well the model handles occlusion and unusual viewpoints in their actual workflow, and whether the SAM License permits their intended use. Checkpoint access is a real onboarding dependency, and a dataset of this scale still requires its own download and processing work. The release is therefore a useful starting point for experiments, not proof that the model is ready for a production pipeline.
Limits and uncertainty
Meta’s repository reports benchmark scores, but the excerpt does not define their units or provide enough protocol detail to compare them independently. Its qualitative claims about generalization, accuracy and handling of occlusions are Meta’s own. Real-world performance on a particular camera setup, demographic mix, clothing, motion blur or image domain remains unmeasured here. The excerpt does not establish whether checkpoint access is granted to every applicant, or spell out commercial-use terms beyond naming the SAM License. Community ComfyUI and GGUF ports are not described in the repository excerpt and are not confirmed here.
Skill to build now
Build a small evaluation set before integrating a reconstruction model: collect representative images with varied poses, viewpoints and occlusions, record the expected downstream use, and compare generated meshes against human review. For SAM 3D Body specifically, first follow Meta’s checkpoint-access and installation instructions, then test the keypoint or mask prompts on cases where the default inference produces uncertain results. Keep licensing review separate from technical evaluation.
Reader takeaway
SAM 3D Body makes Meta’s single-image human mesh recovery workflow testable through code, checkpoints, notebooks and a released dataset. The meaningful next step is not to infer readiness from the release or scores: secure access, confirm license fit and evaluate outputs on the images your product will actually handle.