3 — Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

CycleGAN is one of my fav recent papers https://t.co/Rsv3IshOQv

— Brannon Dorsey (@brannondorsey) August 16, 2017

Your tweet is very important to us and will be answered in the order in which it was received.

I recommend reading this post as well, which explains much of the computer science in a look-at-code sort of way.

This paper introduces CycleGANs — a system that pits NNs against each other to alternatingly ① generate an image $G(X)$ from a base-image set $X$ as a “best-guess”, and ② “grade” that image for how well it fits into a set $Y$, and then ③ compare how well that converted image $G(X)$ can be converted back into the starting image so that $F(G(X))$ most closely approximates original $X$ (the titular “cycle”).

As this is a codebase I’ve used in Real Life, the Results section was less interesting to me than the rationale and the underlying computer science (although the Results section of this paper is one of the most visually appealing of anything I’ve ever read).

In particular: Something that struck me from reading this original paper was the significance of the jump from paired- to unpaired translation, which seemed to me before as a logical ability for a GAN… That other methods like pix2pix don’t have this capability is something I completely failed to appreciate until now.

Had I not gone back to the literature and read this paper, I wouldn’t have ever realized that huge hole in my understanding of the importance of this work. Thanks @brannondorsey!