Improved Training of Generative Adversarial Networks Using Representative Features

Despite the success of generative adversarial networks(GANs) for image generation, the trade-off between image diversity and visual quality remains a significant issue. Conventional techniques focus on either visual quality or image diversity, with improvement in one area often resulting in degradation in the other. This paper achieves both aims simultaneously by improving the stability of training GANs. The key idea of the proposed approach is to implicitly regularize the discriminator using representative features. This representative feature is extracted from the data distribution using a pre-trained autoencoder, and then transferred to the discriminator to enforce slow gradient updates. Consequently, the entire training process is stabilized because the discriminator learning curve varied slowly. The proposed approach is verified to improve visual quality and diversity of state of the art GANs using extensive evaluations.