Home▸Machine Learning & Neural Networks▸CycleGAN: Translating Images Between Domains Without Paired Data

🔄 CycleGAN: Translating Images Without Matched Pairs

Explore how CycleGAN translates images between styles like photos and paintings without needing matched training pairs, using dual generators and cycle-consistency loss.

Machine Learning & Neural Networks3DModerate60 FPS

The simulation visualises CycleGAN's dual-generator, dual-discriminator architecture and shows how an image translated from domain A to B and back to A is compared against the original to compute cycle-consistency loss.

🔬 What It Demonstrates

🎮 How to Use

Select a domain pair to translate between, use the speed slider to control how fast the training loop animates, and use play/pause and rebuild to run or reset the cycle-consistency demonstration.

💡 Did You Know?

CycleGAN was published in 2017 by Zhu, Park, Isola, and Efros, and its horse-to-zebra transformation demo became one of the most widely recognised images in generative AI research.

⚙ Under the hood

Watch two generators translate images back and forth between two domains, held consistent by cycle-consistency loss even without matched training pairs.

cycleganimage translationunpaired datacycle consistencymachine-learninggenerative models

3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install