Sketch to Photo Conversion using Cycle-Consistent Adversarial Networks

Abstract

It is proposed to use the cycle-consistent adversarial network as a way to convert images of sketches to images of photos. The network learns to perform the mapping from the domain of sketches to the domain of photos and due to its architecture also learns the inverse mapping from the domain of photos to the domain of sketches. The network converts sketches to photos by reducing a weighted sum of the validity, reconstruction and identity losses. The advantage of using a cycle-consistent adversarial network over other network architectures is that it is not mandatory to have aligned image pairs as its training set and works in an unpaired setting.