Paired photos in, hallucinated photos out, circa 2017

What it does

This is the original Torch/Lua implementation of the 2017 CVPR paper that popularized general-purpose image-to-image translation. You feed it paired images—say, a label map and a real photo of the same scene—and it learns to generate the photo from the map, or vice versa. It ships with scripts and pre-trained models for facades, cityscapes, edges-to-shoes, day-to-night, and a few other tasks.

The interesting bit

The authors themselves now point users to their PyTorch successor, which makes this repo something of a historical artifact. Still, it is the codebase that established the pix2pix framework: a conditional adversarial network trained on paired input-output images that turned matched-pair training into a general-purpose image translation pipeline.

Key highlights

• Original implementation of the CVPR 2017 paper by Isola et al.
• Trains on paired datasets; the README notes that 400 facade images yielded decent results in about two hours on a single Pascal Titan X.
• Includes pre-trained models for facades, Cityscapes, edges-to-shoes/handbags, and day-to-night translation.
• Comes with helper scripts for dataset formatting, colorization prep, edge extraction, and Cityscapes evaluation (the latter expects 256×256 inputs and upsamples them for scoring).
• Optional display server for plotting L1, generator, and discriminator error during training.

Caveats

• The authors explicitly recommend the PyTorch version for new work, noting it is under active development and produces comparable or better results.
• Requires Linux or macOS with an NVIDIA GPU and CUDA CuDNN; CPU mode is described as untested.
• Several evaluation and preprocessing dependencies lean on older toolchains like Caffe and Matlab.

Verdict

Grab this if you need to reproduce the exact 2017 paper or are maintaining legacy Lua/Torch pipelines. Everyone else should follow the authors’ own advice and head straight to the PyTorch implementation.