Abstract
Mode collapse is a critical problem in training generative adversarialnetworks. To alleviate mode collapse, several recent studies introduce newobjective functions, network architectures or alternative training schemes.However, their achievement is often the result of sacrificing the imagequality. In this paper, we propose a new algorithm, namely a manifold guidedgenerative adversarial network (MGGAN), which leverages a guidance network onexisting GAN architecture to induce generator learning all modes of datadistribution. Based on extensive evaluations, we show that our algorithmresolves mode collapse without losing image quality. In particular, wedemonstrate that our algorithm is easily extendable to various existing GANs.Experimental analysis justifies that the proposed algorithm is an effective andefficient tool for training GANs.