What it does This repo implements CycleGAN in TensorFlow, letting you train models that translate images from one domain to another—say, horses to zebras—without needing matched before-and-after pairs. You supply two folders of unrelated images; the network learns the mapping by cycling translations through a pair of adversarial generators and enforcing that a round-trip gets you roughly back where you started.

The interesting bit The “cycle consistency” constraint is the clever hinge: instead of hand-labeling pairs, the model penalizes itself if horse→zebra→horse doesn’t reconstruct the original horse. It’s a neat workaround for domains where paired data is impossible or tedious to collect.

Key highlights

• Ships with a download_dataset.sh script covering 11 preset datasets (horse2zebra, monet2photo, iphone2dslr_flower, etc.)
• Provides a pretrained horse↔zebra model via OneDrive link
• Training and test entry points are a single main.py with --phase and --which_direction flags
• TensorBoard logging baked in via ./logs
• Explicitly targets TensorFlow r1.1 (yes, r1.1—this is vintage 2017)

Caveats

• Dependencies are pinned to very old versions: TensorFlow r1.1, numpy 1.11.0, scipy 0.17.0. Expect friction on modern Python/TF stacks.
• The README is minimal: no mention of hardware requirements, training time, or expected output quality beyond the sample images shown.
• Several application sections (paintings, season transfer, photo enhancement) are commented out in the README, leaving only horse/zebra results on display.

Verdict Worth a look if you need a clean, minimal CycleGAN reference in TensorFlow 1.x or want to dissect the cycle-consistency mechanics without PyTorch abstractions. Skip it if you need production-ready code or a modern TF 2.x/Keras implementation—this is a period piece.