Improved Training of Spectral Normalization Generative Adversarial Networks

In order to stabilize the training of generative adversarial networks, several recent works advocate spectral normalization in the discriminator. However, the method ignores the influence of the generator, and the quality of the images generated in practice is unstable. We propose L2 norm regularization in the generator based on the spectral normalization, which can solve the above shortcomings. Our method directly makes the generated data close to real data in Euclidean space, and indirectly helps the spectral normalization achieve tighter Lipschitz constraint during the training of generative adversarial networks. Our experiments on CIFAR-10 and STL-10 dataset confirm that our method can not only stable the quality of the images generated by spectral normalization, but also improve the quality of generated images.

I. Introduction

In recent years, Generative Adversarial Networks (GANs) [1] have been studied widely and applied to many fields successfully, such as, text to image generation [2], human faces generation [3], and image to image translation [4], which is an implicit generative model. The GAN consists of a generator network and a discriminator network. The generator network fits the true distribution so as to generate fake data that can deceive the discriminator network. The discriminator network determines whether the input data samples from the true distribution or the generator distribution. GANs find Nash equilibrium when the generator network completely captures the distribution of real data and the discriminator network cannot discern the source of the input data.