SACANet: end-to-end self-attention-based network for 3D clothing animation

In real-time applications such as virtual try-on and video games, achieving realistic character clothing animation is an active research area. Existing learning-based methods aim to simulate realistic clothing effects while meeting real-time requirements. However, due to the diverse sources of clothing deformation, factors such as human body shape, pose, and clothing style all influence the realism of clothing deformation. Existing clothing animation methods often encounter issues such as body-garment interpenetrations and insufficient clothing wrinkles. In this paper, we propose a new end-to-end self-attention-based two-level network architecture for clothing animation, called SACANet, to address these issues. Specifically, we employ a physics-constrained loss function to enable our model to automatically resolve body-garment interpenetrations without the need for post-processing steps. Additionally, within our two-level network, we use an effective self-attention mechanism, namely cross-covariance attention, to enhance our model's ability to learn geometric features and to capture more realistic clothing deformations. The base deformation network learns the global deformation of the clothing by extracting global geometric features from the input data. The detail deformation network learns the local wrinkle details of the clothing by fusing features extracted by specific pose models related to human body shape and clothing style. Through a series of experiments, we comprehensively evaluate and compare the results. These experiments demonstrate the superiority of our method in simulating realistic clothing animation. https://github.com/JUNOHINATA/SACANet.