Exploring the capability of JoJoGan: AI-powered style transfer

Abstract

JoJoGAN is a Generative Adversarial Network that focuses on transferring a fixed style from a reference image into input images to produce high-quality stylised images. Unlike traditional methods which require substantial datasets, JoJoGAN leverages on StyleGAN and can produce higher-quality stylised images with only one reference image. This project aims to dive deep into JoJoGAN particularly on human faces as JoJoGAN is relatively unexplored. Repeated experiments on varying the inputs and parameters of the JoJoGAN models further support the strength of JoJoGAN but also highlight the weaknesses of JoJoGAN. JoJoGAN does well in producing quality stylised results. It is also quite flexible in terms of retaining the input image’s identity and expressions. However, there are some random inconsistencies in recognising facial features resulting in some errors while transferring the styles. Moreover, JoJoGAN is quite dependent on the quality of the input. The input images have to have resemblance to a human face in order to have accurate style transfer. Extreme faces such as animals may not produce expected outputs. This project provides insight into JoJoGAN’s practicality in digital art and applications. Future work could expand outside of human faces to fields such as animal faces, or even inanimate objects like cars and houses.