Pluralistic Face Image Completion Based on StyleGAN

Xiaofeng Shao, Zhenping Qiang, Fei Dai, Shiqi Su, Yuxuan Zuo, Libo He

Abstract: For masked images, although there are many methods that can be completed to a reasonable result, their output is only one. In fact, there should be multiple possible predictions for a masked face image. To address this limitation, in this paper, we propose a pluralistic face image completion method to generate more than one reasonable result. The generative adversarial network can generate images from random noise, and we believe that each image corresponds to a latent vector. We use the image generated by StyleGAN as the training set, train a ResNet module to extract the latent vector of the image, and then use this latent vector to generate the image so that the image can be reconstructed. For masked images, we do not consider the missing information area when we calculate the loss. We conduct experiments on the FFHQ dataset, and experimental results show that this method can effectively achieve pluralistic face image completion.