phillipi/pix2pix

Implements a U-Net generator paired with a PatchGAN discriminator to learn bidirectional mappings between image domains, enabling tasks like semantic label-to-photo synthesis and edge-to-object generation. Built in Torch with CUDA support, it trains efficiently on modest datasets (e.g., 400 images in ~2 hours) and includes pre-trained models for facades, Cityscapes, shoes, and handbags. Provides dataset utilities for edge detection via HED and Lab colorspace preprocessing, plus evaluation scripts for semantic segmentation metrics on Cityscapes.

10,628 stars. No commits in the last 6 months.