DivGAN: A diversity enforcing generative adversarial network for mode collapse reduction

Abstract

Generative Adversarial Networks (GANs) are one of the most efficient generative models to generate data. They have made breakthroughs in many computer vision tasks. However, the generic GAN suffers from a mode collapse problem. To alleviate this problem, the present work proposes a new GAN framework called diversified GAN (DivGAN). DivGAN can be incorporated into any existing GAN. It includes a new network called DivNet that aims at enforcing the GANs to produce diverse data. One advantage of the proposed network is that it does not alter the architectures of the GANs and hence can be easily incorporated into them. Extensive experiments on synthetic and real datasets show that the proposed framework significantly contributes to mode collapse reduction and performs better than recent state-of-the-art GANs.