CycleGAN and Unpaired Translation

Ordinary image-translation models need paired training examples — the same scene shot in both styles, so the model can be shown exactly what a photo should look like as a painting. Paired data like this is rare or impossible to collect for most interesting problems: nobody has a photo and a Monet painting of the same haystack taken from the same angle. CycleGAN removes this requirement entirely, learning to translate between two unpaired collections — a folder of photos and an unrelated folder of paintings — using only the fact that both collections exist.

It trains two generators at once: one that turns photos into paintings, and one that turns paintings back into photos. The key trick is cycle consistency: take a real photo, translate it to a painting, then translate that painting back to a photo, and penalize the model if the round trip doesn't return something close to the original photo. This constraint alone rules out generators that produce a plausible-looking painting that has thrown away the photo's actual content (the shape of the haystack, its position), because such a painting would be impossible to translate back correctly. A separate discriminator network for each domain still checks that the style looks convincingly like a painting or a photo, exactly as in an ordinary GAN.

In markets research the same idea has been used to translate synthetic price paths between regimes — generating a plausible "stressed" version of a calm-market series, or vice versa — without needing a matched calm/stressed pair for the same underlying, since such pairs simply don't exist historically.