TARig: Adaptive template-aware neural rigging for humanoid characters

Abstract

We present TARig, a template-aware neural rigging method, to automatically generates industry-standard and high-quality skeleton and skin weights for humanoid characters. The resulting skeleton consists of a widely-used humanoid template with 21 template joints and associate secondary joint sets which are generated adaptively for handling auxiliary components. To explicitly utilize the anatomic consistency among humanoid characters, we incorporate the manually annotated anatomic label of each template joint to fully supervise the rigging process in an end-to-end training scheme. Together with a shared graph neural network backbone, such semantic constraint ensures to generate precise joint positions and accurate skin weights simultaneously. Beyond the pre-defined topology of the template skeleton, we additionally learn a boneflow field to further determine the inner connections for each secondary joint set and avoid counter-anatomical skeleton construction. Extensive experiments demonstrate TARig generates high-quality rigging results and outperforms other state-of-the-art auto-rigging methods in terms of skeleton generation and skin weight estimation for humanoid characters with 8∼20 times speedup.