pix2pix

Why we included this project

pix2pix comes from the 2017 CVPR paper that made conditional GANs a standard tool for translating between two image domains. Instead of generating from scratch, it learns a mapping from one kind of image to another, for example turning edge maps into building facades or sketches into photos, by training a generator and discriminator against each other on paired examples. That makes it a natural reference for teams experimenting with domain transfer, style translation, or paired image synthesis. The authors report that decent results can come quickly from small datasets, with the facade demo trained on only 400 images in about two hours on a single GPU, so it is a practical prototype to start from. One caveat: this implementation targets the legacy Torch ecosystem, so for real project work most people will want the actively maintained PyTorch reimplementation the authors also publish.