Image-to-Image Translation Method for Game-Character Face Generation

Abstract

Traditional image-to-image translation methods effectively change the style; however, these methods have several limitations in shape changing. Particularly, current image-to-image translation technology is not effective for changing a real-world face image to the face of a virtual character. To solve this problem, we propose a novel unsupervised image-to-image translation method that is specialized in facial changes accompanied by radical shape changes. We apply two feature loss functions specialized for faces in an image-to-image translation technique based on the generative adversarial network framework. The experimental results show that the proposed method is superior to other recent image-to-image algorithms in case of face deformations.