Rewriting Style Features of a Pre-Trained GAN Based on a Single Image

Generative adversarial networks (GANs) have been recognized for their major contributions to image generation. Novice users can easily create diverse, realistic and high-quality images with GAN models. However, GAN training often requires a substantial amount of data, the lack of which presents challenges to image generation. To overcome this problem, this paper proposes a novel method that leverages GAN capabilities and advanced style transfer techniques to rewrite a pretrained generative model. Users only need to input a single style image to edit the generator and synthesize images that include the target style features while preserving the original content.