A Comparison of Cartoon Portrait Generators Based on Generative Adversarial Networks

Abstract

Cartoon portraits are deformed figures that capture the appearance and characteristics of people, and are often used to express one’s image in applications such as social media, games, application profiles, and avatars. Current research regarding the translation of facial images into cartoon portraits focuses on translation methods that use unsupervised learning and methods for translating each part individually. However, studies that reflect the unique personality of professional illustrators have yet to be published. In this study, we examine a suitable network for reflecting the unique personality of a professional illustrator. Specifically, we will consider four networks: pix2pix, Cycle Generative Adversarial Network (CycleGAN), Paired CycleGAN, and Cyclepix. The main difference between these is the loss function. Pix2pix takes the error between the training data and the generated data. However, the main difference in CycleGAN is that it takes the error between the input data and the re-converted data obtained by further translating the generated data. Cyclepix takes both errors. Additionally, pix2pix and Paired CycleGAN require that the input of the discriminator be input data and generated data pairs. The difference between CycleGAN and Cyclepix is that only the input of the discriminator is generated data. Using the cycle consistency loss, considering only the input of the discriminator as generated data, and using the L1 Loss for supervised learning, the experimental results showed that the evaluation of CycleGAN and Cyclepix was high. This is useful for generating high-precision cartoon portraits.