Hierarchical Neural Skinning Deformation with Self-supervised Training for Character Animation

We propose a self-supervised neural deformation method for character skin deformation animations, incorporating musculoskeletal and adipose tissue deformations driven by joint movements. This hierarchical approach combines two self-supervised neural emulators: the musculoskeletal neural emulator, based on a biomaterial musculoskeletal model, and the soft-tissue neural emulator, capturing soft secondary dynamics. The neural emulators achieve self-supervised learning through a loss function derived from physical principles, eliminating the need for ground-truth datasets. The musculoskeletal neural emulator enhances deformation effects, including musculoskeletal bulging, contraction, and dynamic responses, and generates inner-layer animations. Building on this, the soft-tissue neural emulator adds enhanced deformation effects, such as secondary motion, driven by the inner-layer animations. During inference, the method efficiently simulates realistic skin deformation animations for arbitrary character structures and joint motion sequences. The resulting animations demonstrate visual fidelity comparable to, or surpassing, state-of-the-art data-driven methods in the industry.