A single photo does not reveal a person’s full 3D pose, especially when limbs are occluded or the viewpoint is unusual. This model’s key idea is to make mesh recovery promptable: users can add 2D keypoints or masks to steer an estimate, while the Momentum Human Rig separates skeletal pose from body shape for a more interpretable representation.
Key Capabilities
- It estimates body, hand, and foot pose together, so the output is more complete than a body-only skeleton.
- The MHR representation decouples skeletal structure and surface shape, making pose and shape parameters separately useful for downstream analysis.
- Optional 2D keypoint and mask prompts allow user guidance when the image alone leaves ambiguity.
- The released DINOv3-H+ checkpoint has 840 million parameters; the project also offers a ViT-H checkpoint, providing a choice of backbone.
Who It’s For—and the Trade-offs
Great fit if you need single-image human mesh estimates for research, pose analysis, or visual content workflows, particularly for varied real-world poses. Look elsewhere if you need multi-view reconstruction, real-time lightweight inference, or a model that can be downloaded without access approval: checkpoint access requires a request, and the large DINOv3-H+ backbone brings substantial compute demands.
