Arbitrary Motion Style Transfer via Contrastive Learning

In recent years, significant progress has been made in arbitrary motion style transfer. However, many existing methods primarily focus on end-to-end processes and overlook the differences and similarities among generated motions, thus failing to fully leverage style information. To address this limitation, we employ contrastive learning to bring generated motions closer to input motions and further away from others. This enables the model to construct better motion representations and optimize the quality of motion stylization. Additionally, existing methods often struggle to recognize different styles. To address this, we employ contrastive learning on random temporal motion clips of the same motion, allowing the model to compare details between the same and different styles at a finer temporal scale. Through extensive experiments, we demonstrate that our proposed method generates robust feature representations, resulting in a stronger stylization of generated motions and a more accurate preservation of content.