Face Reenactment with Diffusion Model and Its Application to Video Compression

Wataru Iuchi, Kazuaki Harada, Hayato Yunoki, Koki Mukai, Shun Yoshida, Toshihiko Yamasaki

Abstract: In this study, we present a face video compression scheme featuring extremely low bit rates. In our proposed method, a face image is reconstructed from the previous frame by recursively using a diffusion model, thereby reducing the trade-off between person identification and facial expression generation while achieving smoothness between frames. Experimental results show that our proposed method outperforms baseline methods in terms of image quality.