A Two-Stage Unsupervised GAN Approach for Anime-to-Manga Translation

Abstract

This project aims to investigate how cartoon and comic images can be better explored within the context of unsupervised image-to-image translation. Specifically, we seek to study the translation of anime illustrations into their manga representations, given a manga book as a reference. Although current state-of-the-art image-to-image translation models can convert images between different domains, existing methods for translating illustrations to manga style are scarce. We propose to exploit the unique characteristics of anime and manga images, allowing for a preliminary output that can support the translation process in two stages. We believe this approach can reduce model complexity while generating high-fidelity outputs. Furthermore, we aim to impose minimal restrictions on the manga target domain, making the translation fully unsupervised. Finally, the proposed framework's output can be used to produce rich datasets composed of colored and synthetic manga images, which would support colorization methods that rely on large amounts of paired training data.