A stubbornly faithful TensorFlow port of pix2pix

What it does This repo implements Isola et al.’s pix2pix in TensorFlow, learning to map input images to output images—think architectural labels to photos, or sketches to shoes. It consumes side-by-side image pairs to train a conditional GAN, then generates an HTML gallery comparing inputs, outputs, and targets. A dedicated colorization mode handles single images by splitting them into lightness and color channels.

The interesting bit Unlike later rewrites, this port was built directly from the original Torch code rather than from an existing TensorFlow version, aiming for near-photocopy fidelity. The author notes that GPU training speed with cuDNN matched the Torch implementation, which matters more than it sounds if you are trying to reproduce a 2016 paper beat-for-beat.

Key highlights

• Explicitly a faithful port: the author states it “does not add anything” to the original work.
• Targets TensorFlow 1.4.1 and includes a Docker image for environment isolation.
• Supports paired-image translation (AtoB or BtoA) and a --lab_colorization mode for grayscale-to-color work.
• Test runs emit an index.html file that lines up inputs, outputs, and targets for easy visual inspection.
• Pre-trained models and dataset download scripts are provided for facades, cityscapes, maps, edges2shoes, and edges2handbags.

Caveats

• Requires TensorFlow 1.4.1, so expect archaeology when integrating with modern stacks.
• Three original variants remain unimplemented: defineG_encoder_decoder, defineG_unet_128, and defineD_pixelGAN.
• The codebase is strictly a reproduction, so do not expect modern conveniences or architectural tweaks.

Verdict Use this if you need a reference-quality TensorFlow reproduction of classic pix2pix for research or education. Skip it if you want a maintained modern implementation or need the missing generator and discriminator variants.