junyanz/CycleGAN

Implements unpaired image-to-image translation using cycle-consistent adversarial networks, enabling domain translation without paired training data by enforcing that images translated from domain A→B→A reconstruct the original. Built in Torch with adversarial loss between generator and discriminator networks, includes pix2pix and BiGAN/ALI variants, and provides pre-trained models for diverse applications from artistic style transfer to semantic segmentation.

12,843 stars. No commits in the last 6 months.