IPFR: Identity-Preserving Face Reenactment with Enhanced Domain Adversarial Training and Multi-level Identity Priors

In the face reenactment task, identity preservation is challenging due to the leakage of driving identity and the complexity of source identity. In this paper, we propose an Identity-Preserving Face Reenactment (IPFR) framework with impressive expression and pose transfer. To address the leakage of driving identity, we develop an enhanced domain discriminator to eliminate the undesirable identity in the generated image. Considering the complexity of source identity, we inject multi-level source identity priors to keep the identity domain of generated image close to that of source. In detail, firstly, we utilize a 3D geometric prior from the 3D morphable face model (3DMM) to control face shape and reduce artifacts caused by occlusion in the module of generating motion field; secondly, we use an identity texture prior extracted by face recognition network to supervise the final stage, aiming to make the identity domain of the generated close to that of source. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art methods on image quality and identity preservation. Ablation studies are also conducted to further validate the effectiveness of our individual components.