TopoRig: Topology-Agnostic Facial Rigging via Multi-Source Supervision

arXiv cs.GR details TopoRig, a topology-agnostic facial rigging framework that predicts FACS-conditioned deformations directly on the input mesh instead of transferring a rig onto a fixed template. That matters for character teams because it aims to preserve the original topology while reducing the correspondence errors and surface artifacts that often show up when rigs are pushed across very different meshes.

The system combines several supervision sources: accurate but template-biased common-topology rigs, noisier transferred rigs from more diverse meshes, and image-based cues for controls that geometric transfer handles poorly. It also blends local surface geometry, landmark-relative semantic features, global shape context, and FACS controls to drive per-vertex displacement, which is a useful reminder that facial rigging quality often comes down to how well semantics and geometry are fused.

The authors say they trained on 3,496 generated identities using 45 non-gaze controls from the 53-control ICT FaceKit vocabulary. On held-out identities and unseen topologies, TopoRig reportedly reproduces the reference expression space more faithfully than prior neural facial-rigging methods, with ablations showing that semantic landmark features and complementary supervision improve generalization.

For developers, the practical angle is clear: if this approach holds up outside the paper, it could reduce the amount of manual cleanup needed when shipping facial animation across a wide range of character meshes. The exact production cost and runtime integration details still need...