Few Data Diversification in Training Generative Adversarial Networks

Abstract

The first GANs have initially produced sharp images in relatively small resolution and with limited variations, and unstable training. Later works proposed new GAN models capable of generating sharp images in high resolution and with a high level of variation. However, these models use unlimited and highly diversified image sets. We discuss here the use of these models with real-world image sets, since they are composed of limited sample size sets.