RegHead: Non-Humanoid Head Blendshapes via Feed-Forward Registration

RegHead is a new pipeline for creating semantic blendshape sets for animatable non-humanoid head avatars. The practical goal is familiar to anyone who has wrestled with facial rigs: keep expressions interpretable, reusable, and retargetable across characters, even when the head shape is far from human.

The core problem is correspondence. Non-humanoid faces tend to have localized, asymmetric motion, and the usual route of optimizing each asset into a blendshape basis is slow and labor-intensive. RegHead addresses that with a feed-forward registration model that predicts anchor-based deformations from a neutral mesh, turning unregistered expression meshes into a corresponded blendshape set much faster than optimization-based methods.

To make the system work at scale, the team built a large dataset of non-humanoid identities paired with a shared expression vocabulary. That dataset was expanded from a small artist-rigged library using fine-tuned image editing, which is a notable production angle: the pipeline leans on artist-authored structure, then uses generative tooling to broaden coverage.

For game teams, the appeal is obvious. A low-dimensional semantic interface makes facial animation easier to author, debug, and retarget, especially when you want one tracking source to drive many creature heads. The project also demonstrates real-time retargeting from human face tracking signals to non-humanoid characters, preserving both head pose and localized facial motion. The paper is dated July 13, 2026, and the authors report higher-fidelity results than baselines while...