Learning identity-independent expression representation for identity-preserving facial expression transfer

Abstract

Facial Expression Transfer synthesizes a video with a given static portrait that imitates the driving video. To effectively capture the characteristics of the face with given face images, recent studies adopt the deformation operation, where the source images are deformed toward the target structures guided by driving frames. However, we observe that those methods frequently suffer from identity distortion for the source subjects since they transfer not only the expressions but also the personal shapes from the driving subjects. To resolve this issue, we propose Identity Preserving Portrait Animator (IPPA), which learns the identity-independent expression space and predicts the target structure based on the expression representation and the source image's structure. We evaluate IPPA on the two benchmark datasets such as Voxceleb and TalkingHead-1KH, and confirm that IPPA can generate high-quality videos imitating the driving expressions while preserving the identity of the source subjects.