contrastive-unpaired-translation

Why we included this project

This is the PyTorch reference for CUT, the ECCV 2020 unpaired image-to-image translation method that swaps CycleGAN's cycle-consistency losses for a patchwise contrastive (PatchNCE) objective. Dropping those losses makes training faster and easier on GPU memory, which matters when you're translating between large photo collections that have no paired examples. There's also a single-image mode where each 'domain' is just one picture, handy for stylizing a single source image. If you work with GANs, the code is straightforward to run for style transfer, domain adaptation, or relabeling datasets. Pretrained weights and launcher scripts for common datasets let you reproduce the paper's results and reuse the generator as a base for your own pipeline.