SP-GAN: Cycle-Consistent Generative Adversarial Networks for Shadow Puppet Generation

Converting real face images into shadow puppet-style facial images is a challenging task. The problem lies in the asymmetry of the amount of information and geometric topology structures between real face images and shadow puppet facial images, which differs from previous style transfer tasks. In this paper, we propose Shadow Puppet-CycleGAN (SP-GAN), an asymmetric CycleGAN framework for shadow puppet stylization. Furthermore, we propose a contour fusion network module and a facial contour loss function to encourage the retention of basic personalized contours in the generated shadow puppet-style facial images. The experimental results show that our method generates high-quality shadow puppet-style facial images from real face images (i.e., retaining personalized facial information while embodying shadow puppet style) and outperforms existing methods.