facebookresearch/MHR

What it solves

MHR (Momentum Human Rig) provides a high-fidelity, parametric 3D human body model. It allows users to generate realistic human meshes by controlling specific parameters for body identity, full-body articulation (pose), and detailed facial expressions, overcoming the need to manually sculpt every human variation.

How it works

The system uses a set of parameters to define a human figure:

• Identity: 45 parameters that control the shape of the body, head, and hands.
• Pose: 204 parameters for joint angles and scaling to handle body movement.
• Facial Expression: 72 semantic blendshape weights based on FACS for detailed face animation.
• Neural Correctives: It employs neural network-based pose-dependent deformations to ensure the mesh looks natural during movement.
• Differentiable Pipeline: Built with PyTorch, the model supports differentiable inference and optimization, meaning it can be integrated into machine learning workflows to optimize parameters based on a loss function.

Who it’s for

This tool is designed for researchers and developers working in 3D human modeling, computer vision, and animation who need a flexible, differentiable human rig for generating or recovering human poses and shapes.

Highlights

• Multiple LODs: Offers 7 levels of detail (LOD 0-6) to balance performance and visual quality.
• PyTorch Integration: Supports GPU/CPU acceleration and standard autograd for optimization.
• TorchScript Support: Provides a lightweight version of the model that requires no external codebase or assets for basic inference.
• Interoperability: Includes tools for converting between MHR and other common human models like SMPL and SMPL-X.