Differentiable cloth simulators for inverse physics of dressed avatars

Develop differentiable physics-based cloth simulation methods that can accurately handle complex body–cloth collisions and are suitable for inverse estimation of garment density, membrane stiffness, and bending stiffness from multi-view videos of dressed human avatars, enabling gradient-based optimization of physical parameters within PhysAvatar’s pipeline.

Background

PhysAvatar estimates garment physical parameters by integrating a cloth simulator into a gradient-based optimization loop. Accurate gradient information is essential for this inverse physics step, particularly in the presence of complex body–cloth interactions modeled by a parametric human body (SMPL-X).

The authors note that existing differentiable simulators (e.g., DiffCloth) cannot be directly applied due to limitations with complex body colliders. While the robust C-IPC solver is used for high-fidelity simulation, it lacks accurate analytical gradients, necessitating finite-difference approximations. This motivates the need for suitable differentiable simulators to support the method’s objectives.