Generative Adversarial Networks Based on Human-Computor Interaction

Abstract

In order to improve the generation quality and personalization of generative adversarial network, this paper proposes an open generative adversarial network (OpenGAN) based on human-computer interaction, which adds human subjective evaluation into the training. A subjective penalty function is added to the original generator loss and the smoothing network layer is designed to reduce the impact of loss mutation in the interaction. Our results show that the IS value on ADE20K, Cityscape and other datasets increases by 61% on average, while KID and LPIPS decrease by 32% and 44%, respectively.