Robust prediction of high-quality skinning weights for arbitrary 3D characters

Develop robust algorithms to predict high-quality skinning weights for arbitrary 3D character meshes that avoid leakage and deformation artifacts across diverse poses and topologies, enabling reliable geometry-space rigging and posing.

Background

Traditional geometry-space rigging relies on accurate skinning weights to deform meshes according to skeletal motion. Even state-of-the-art auto-rigging methods can suffer from skinning leakage and spiking artifacts, particularly for limbs in close proximity or topologically imperfect meshes.

The authors highlight this limitation and motivate their latent-space posing approach, while explicitly noting that learning to robustly predict high-quality skinning weights for arbitrary 3D characters remains unresolved.