You’ve probably got one or two apps on your smartphone capable of turning your selfies into impressionistic-style paintings. Well, researchers at UC Berkeley in California have developed an AI capable of not only doing what Prisma does but also the opposite: transforming a painting into a photograph. Say hello to CycleGAN.
How does CycleGAN work?
The software is able to generate photos from paintings by “translating” unpaired images, i.e. pictures that share an underlying relationship but do not otherwise directly correspond to one another. They may be different renderings of the same scene, for example. “We apply our method on a wide range of applications, including style transfer, object transfiguration, attribute transfer and photo enhancement.”
“In this paper, we present a system that can learn to do the same: capturing special characteristics of one image collection and figuring out how these characteristics could be translated into the other image collection, all in the absence of any paired training examples,” write Jun-Van Zhu, Taesung Park, Phillip Isola and Alexei A. Efros of the Berkeley AI Research (Bair) laboratory.
CycleGAN and a paper detailing its potential were uploaded to GitHub. The maths behind this is… complicated. And it’s a Thursday evening, so we can’t really be bothered to delve into too much detail.