Fine-Grained Face Swapping Via Regional GAN Inversion

Source / Target MegaFS [48] StyleFusion [15] HiRes [41] Ours Figure 1. Compared with the existing StyleGAN-based face swapping approaches [ 15, 41, 48] , our proposed method can achieve highfidelity results that show better identity keeping from the source while keeping the similar pose and expression as the target. Note that skin color preservation and proper occlusion handling are our advantages over others. All the facial images are at 1024⇥1024.