Identity preserving face synthesis using generative adversarial networks

Abstract

This paper provides a comparative analysis between two recent image-to-image translation models based on Generative Adversarial Networks. The first one is UNIT which consists of coupled GANs and variational autoencoders (VAEs) with shared-latent space, and the second one is Star-GAN which contains a single GAN model. Given training data from two different domains from dataset CelebA, these two models learn translation task in two directions. For evaluation, we conduct some experiments and provide a quantitative comparison using direct metric GAM (Generative Adversarial Metric) to quantify the ability of generalization and the ability of generating photorealistic photos. In addition, StarGAN and UNIT are tested in generating new head poses with preserving identity, where there is no prior information about their ability to translate faces from a source pose domain to a target pose domain. Furthermore, some modifications on reconstruction loss function are proposed to improve synthesis of new poses with identity preserving. The experimental results show the superiority of cross-model UNIT over multi-model StarGAN on generating age and eye glasses attributes, and the equivalent performance to synthesize other attributes. While with regard to synthesizing new poses, they show the superiority of multi-model StarGAN over cross-model UNIT and better performance on our loss functions.