Deformation Blocks for Caricature Generation Using StyleGAN

Abstract

In this thesis, a novel approach is presented for realistic caricature generation from a given facial photo. Although caricature artists exaggerate their subject's face by considering the 3D shape as well as shading, only 2D image deformation has been applied in previous studies to exaggerate facial images. To resolve this limitation, we propose a novel framework using deformation blocks, which deforms images in high-dimensional disentangled feature space. Owing to the high capacity of the feature space, the proposed network can express stereoscopic deformations as well as natural shading effects corresponding to the deformation. In addition, using facial attribute dataset, the exaggeration was directed to be semantically meaningful. Experimental results show that the proposed method generates more realistic caricatures than those generated by the current state-of-the-art methods. Furthermore, we also demonstrate user control of our framework, such as exaggeration control, caricature expression manipulation, and caricature style selection.