Stacked Generative Adversarial Networks for Image Generation based on U-Net discriminator

Abstract

Although Generative Adversarial Networks (GANs) are powerful generative models and have shown remarkable success in various tasks recently but suffers from generating high-quality images. In this paper, we proposed a U-Net-based discriminator structure on the network structure of the two-stage Stacked Generative Adversarial Networks (StackGAN++), aiming to generate high-resolution images with actual shapes and textures. To gain more insight from limited datasets, we focused on improving the discriminator's ability to discriminate the real from the fake. The discriminator based on U-Net architecture allows providing details per-pixels and global feedback to the generator to maintain the global coherence of synthetic images and the realistic of local shape and textures. In addition, for the problem that the training effect on a small number of sample datasets is not ide-al, we further improve the quality of the generated samples by transfer learning of model parameters. Compared with the StackGAN++ baseline, experiments show that we have significantly improved the IS and FID evaluation indicators of the ImageNet subset.