A Deep Emulator for Secondary Motion of 3D Characters

Figure 1 : a) Our method is trained on a primitive (sphere) dataset with deformation dynamics but generalizes across topology varying 3D characters. The uniform volumetric mesh surrounding the surface mesh is used for prediction, where the red vertices are set to be constraints. b) At inference time, the input is an artist-specified skinned mesh without dynamics. c) Our neural network predicts the dynamic mesh with secondary motion. d) The surface mesh is rendered with textures.