Learning-based simulation of loose-fit garments

Abstract

In this paper, we propose a learning-based garment simulation algorithm that uses two consecutive networks to predict the deformation of loose-fit garments, preserving high details. The deformation is computed sequentially through low-frequency and high-frequency modules. First, the low-frequency module predicts the overall shape of the garments from the given body motion. Next, the high-frequency module estimates the high-resolution garments with detailed wrinkles by inferring the dynamics of the clothing, referred to the result of the previous module, the local information of the current garment mesh and some reference information. In addition, we improved the stability of rollout in inference time by mitigating the accumulation of errors over time using the scheduled sampling training method. The comparison shows that our method can estimate realistic and detailed garment meshes.