HybriDeformer: A hybrid deformation method for arbitrary 3D avatar controlling

Abstract

In this paper, we address the task of 3D avatar control, focusing on adjusting the shapes of characters. Most existing approaches generally fall into two different paradigms. Parameter deformation methods control shapes using several parameters, resulting in fast and accurate shape deformations, but they are not suitable for arbitrary 3D avatars. In contrast, non-parameter deformation methods can be applied to any 3D avatar but require substantial labor costs or supervision. To this end, we propose a hybrid deformation method termed HybriDeformer, which combines the strengths of both paradigms. The HybriDeformer includes a disentangled parameter deformer (DP) and a non-parameter deformer (NP). Specifically, the DP deformer allows users to freely deform arbitrary 3D avatars, while the NP deformer employs an optimized deformation strategy to drive each part of the 3D avatar to the desired shape, preserving smoothness as much as possible. Additionally, we present a new avatar dataset named ASTD for arbitrary 3D controlling, which can also be used for 3D avatar style transfer. Extensive experimental results on the arbitrary 3D avatar controlling task demonstrate that our method can be directly used by layman users to achieve high-quality deformation results. Our code and project are available at https://sites.google.com/view/hybrideformer.