Audio-Driven Stylized Gesture Generation with Flow-Based Model

The chapter introduces an innovative method for synthesizing stylized audio-driven gestures using a flow-based model. It addresses the challenge of the one-to-many relationship between audio and gestures by leveraging normalizing flows to generate diverse and realistic gestures. The proposed model jointly trains a generative flow with a global encoder, capturing both local and global features of gestures. The authors also design a gesture perceptual loss to improve the quality of generated results. The approach is evaluated on two benchmark datasets, demonstrating superior performance compared to state-of-the-art methods. The chapter concludes by highlighting the potential of the flow-based model for style transfer and manipulation of gesture styles, offering promising avenues for future research.