High-Fidelity and Freely Controllable Talking Head Video Generation

PECHead (a). Frontalization Input Yaw Pitch Roll Input Neutral Smile Laugh Input Smile Expressionless Roll (b). Free pose and expression editing Figure 1. Presented herein are representative results showcasing the effectiveness of our proposed method in the tasks of frontalization, as well as free pose and expression editing. (a) FOMM [46] often produces face distortion issues, while our model PECHead generates high-fidelity results. (b) Our proposed framework empowers the generation of talking head videos that offer free control over the head pose and expression. More results can be found in the supplementary materials.