It uses Conditional GAN(Generative adversarial networks) to convert a front face image into a more primitive representation of the face.