A Low Spectral Bias Generative Adversarial Model for Image Generation

We propose a systematic analysis of the neglected spectral bias in the frequency domain in this paper. Traditional generative adversarial networks (GANs) try to fulfill the details of images by designing specific network architectures or losses, focusing on generating visually qualitative images. The convolution theorem shows that image processing in the frequency domain is parallelizable and performs better and faster than that in the spatial domain. However, there is little work about discussing the bias of frequency features between the generated images and the real ones. In this paper, we first empirically demonstrate the general distribution bias across datasets and GANs with different sampling methods. Then, we explain the causes of the spectral bias through the deduction that reconsiders the sampling process of the GAN generator. Based on these studies, we provide a low-spectral-bias hybrid generative model to reduce the spectral bias and improve the quality of the generated images.