Interpolation based Anime Face Style Transfer

This paper presents a new post-processing method for transferring human face to anime style with geometry factors. This is a branch of image-to-image problem, which is attracting the attention of researchers recently. Existing methods just deal with either texture transfer or geometry transfer. They still do not keep the signatures of human face. Moreover, previous methods do not allow users to control output image with some specific factors. This thing makes these approaches not user-friendly. To address these issues, we propose a system which leverages the advantages of Generative Adversarial Networks model and Thin Plate Spline algorithm for face morphing. The generative model is trained on selfie2anime dataset including anime and human face images. Another advantage of this approach is that it does not require paired images. We evaluate our system on 3400 images resized to 256x256 for each domain. Experimental results illustrate that our method is able to generate anime faces, which have geometry factor associating with human faces.