Recently, neural-based style transfer has become extremely popular thanks to the seminal work of Gatys et al.  and its numerous publicly available implementations like DeepArt and Prisma. Selim et al.  extended this technique to provide better results when stylizing head portraits. In their system, additional spatial constraints improve the resemblance between the stylized portrait and its real counterpart. They align the style image to the target photo and compute a set of gain maps to modify the response of the neural network in order to suppress the local diferences in appearance.