Exploring Generative Adversarial Networks for Photo-to-Anime Image-to-Image Translation: A Comparative Study

Abstract

Generative Adversarial Networks (GAN) is a generative modeling approach with the ability to learn highly complex data. Particularly, they don’t require immediate pairing between the data in input and output domains. This property makes it ideal for image translation tasks. Image translation from photo image into anime style images using GAN is a fast and efficient way to generate art for creative industries. Recently, some algorithms such as U-GAT-IT, CycleGAN, AnimeGAN, and CartoonGAN emerged as few algorithms to accomplish this particular task. The purpose of this paper is to compare the performance of these algorithms in photo-to-anime styled image-to-image translation and discuss the results these algorithms in image-to-image translation task between photo image domain into anime image domain.