phillipi/pix2pix

Project overview

phillipi/pix2pix provides a Lua Torch implementation, training and testing scripts, pre-trained models, and datasets for paired image-to-image translation using conditional adversarial networks. The authors recommend using the PyTorch version for active development.

Project type

Image & Vision

Use cases

Image Processing

Deployment

Refer to project documentation

License

License pending

Best for

• Researchers and developers with command-line and coding experience who need to reproduce or study the original CVPR 2017 pix2pix results using Lua Torch.

Key capabilities

• Provides a Torch implementation and scripts for learning to translate input images to output images using conditional adversarial networks.
• Supports training the image translation model on paired datasets with configurable directions (AtoB or BtoA).
• Tests the trained model on validation or test splits, saving result images and a webpage to view them.
• Offers pre-trained models for tasks like facades generation, cityscapes translation, edges2shoes, edges2handbags, and day2night.
• Provides scripts to download various datasets like facades, cityscapes, maps, edges2shoes, edges2handbags, and night2day.
• Provides a Python script to combine pairs of images {A,B} into a single image file for training.
• Supports training for colorization by automatically converting RGB images into Lab color space.
• Provides Python and Matlab scripts to extract coarse edges from photos using HED edge detector.

Limitations and risks

• Uses Lua Torch, but the authors recommend the PyTorch version for active development.
• CPU mode and CUDA without CuDNN may work with minimal modification, but remain untested.
• The pre-trained FCN model is not supposed to work on Cityscapes in the original resolution of 1024x2048.
• Lua Torch is an older framework with a deprecated ecosystem compared to modern PyTorch.

Getting started

• Setup is rated as hard. Users must install Lua Torch, nngraph, display, and potentially CUDA/CuDNN from the torch distro. After cloning the repo and downloading a dataset, users can execute the training and testing scripts via specific command-line arguments.

Evidence and sources

• GitHub project description: Image-to-image translation with conditional adversarial nets
• README: Torch implementation for learning a mapping from input images to output images
• README: The test results will be saved to an html file here: `./results/facades_generation/latest_net_G_val/index.html`.
• README: DATA_ROOT=./datasets/facades name=facades_generation which_direction=BtoA th train.lua
• README: DATA_ROOT=/path/to/data/ name=expt_name which_direction=AtoB th train.lua

Reviewed at: 2026-07-18T02:54:29Z