Unbalanced GANs: pre-training the generator of generative adversarial network using variational autoencoder

Abstract

In this paper, we propose Unbalanced GANs, which pre-trains the generator of the generative adversarial network using variational autoencoder. By pre-training the generator, we prevent the discriminator's fast convergence at early epochs and show the generator's stabilized training. Furthermore, we improve the quality of the generated images through balancing the generator and the discriminator at early epochs. We apply our method to other GANs and find that Unbalanced GANs reduce mode collapses. We also show that Unbalanced GANs outperform VAE-GAN and ordinary GANs in terms of stabilized learning, faster convergence, and better image quality at early epochs.