Face Reenactment Based on Motion Field Representation

Si Zheng, Junbin Chen, Zhijing Yang, Tianshui Chen, Yongyi Lu

Abstract: Face reenactment is a challenging problem that aims to transfer facial and head motions from a source actor to a target actor. However, existing GAN-based methods often struggle to adequately capture the diverse content within face reenactment videos, resulting in unrealistic transitions between frames. In this paper, we propose a novel facial reenactment framework based on motion field representation. Our approach effectively combines the target actor’s identity with the source actor’s motion, enabling separate modeling and learning of the portrait and background regions in video frames. As a result, we are able to generate highly realistic portrait images. Extensive experimental evaluations demonstrate that our algorithm outperforms many state-of-the-art facial reenactment methods, highlighting its superiority in this domain.