Reducing Differences Between Real and Realistic Samples to Improve GANs

Abstract

Generative Adversarial Nets (GANs) receive much attention and show great superiority in generating realistic images. However, GANs suffer from mode collapse. To address this problem, we introduce sample differences penalization (SDP) as a regularization term to the objective function of GANs. SDP is an easy-to-implement method that aims to reduce the score differences and the feature differences between the realistic generated samples and their nearest real samples. By introducing SDP, the discriminator presents reasonable outputs to the close pairs. The theoretical analyses demonstrate that SDP can help mitigate the gradient at real samples to some extent, which contributes to a more stable training process. Extensive experiments on real-world datasets including CIFAR-10, CIFAR-100, and Tiny ImageNet demonstrate that our GAN-SDP has a more stable training process and leads to a better performance than existing related methods in Frechet Inception Distance (FID) metric.