Tachiegan: Generative Adversarial Networks for Tachie Style Transfer

Abstract

Tachie painting is an emerging digital portrait art form that shows a character in a standing pose. Automatic generation of a Tachie picture from a real photo would facilitate many creation tasks. However, it is non-trivial to represent Tachie's artistic styles and establish a delicate mapping from the real-world image domain to the Tachie domain. Existing approaches generally suffer from inaccurate style transformation and severe structure distortion when applied to Tachie style transfer. In this paper, we propose the first approach for Tachie stylization of portrait photographs. Based on the unsupervised CycleGAN framework, we design two novel loss functions to emphasize lines and tones in the Tachie style. Furthermore, we design a character-enhanced stylization framework by introducing an auxiliary body mask to better preserve the global body structure. Experiment results demonstrate the robustness and better generation capability of our method in Tachie stylization from photos in a wide range of poses, even trained on a small dataset.