Face swapping transforms a source identity into a target image while preserving other important attributes. 3D face reconstruction is an efficient method for face swapping. However, the transformed images by 3D face reconstruction characterize with distortion of high-frequency components and abnormal discontinuity of hue value.In this paper, a face swapping framework based on fine texture restoration is proposed, referred to as GRFS, to address this issue on 3D face reconstruction. GRFS involves two texture restoring modules as follows: mouth restoration network restores mouth details of 3D face reconstruction results; generative local restoration network restores the distortion of high-frequency components and smooths the gaps of hue value. Extensive experiments demonstrate that GRFS performs better than traditional 3D face reconstruction based face swapping methods and balances the quality and efficiency better than other mainstream face swapping frameworks.

face swap; 3D face reconstruction; texture restoration; generative adversarial network; self-supervised learning