What it does

GAN Compression takes bloated conditional GANs—pix2pix, CycleGAN, GauGAN, MUNIT—and trims them down for actual deployment. The repo claims 9–21× computation reduction and 4.6–33× model size shrinkage while keeping visual fidelity intact. It also ships a TVM-tuned interactive demo that hits 8 FPS on a Jetson Nano, which is the difference between “works in the lab” and “works in someone’s hand.”

The interesting bit

The trick is a “once-for-all” student generator trained via distillation with weight sharing across all possible channel counts. After that one training run, you extract sub-generators at different widths and evaluate them without retraining—then pick your sweet spot between speed and quality. It’s basically a buffet where you decide the portion size after cooking the whole meal.

Key highlights

• Supports pix2pix, CycleGAN, GauGAN, and MUNIT (multimodal unsupervised translation)
• Includes pre-trained compressed models with shell scripts for testing and latency measurement
• Interactive demo with TVM optimization for edge GPUs (Jetson Nano at 8 FPS)
• Colab notebooks for CycleGAN and pix2pix to skip the setup slog
• T-PAMI version of the paper available, plus legacy model weights if retrained versions drift from published results

Caveats

• Linux and NVIDIA GPU strongly implied; CPU support exists but this is clearly GPU-targeted
• Dataset prep for Cityscapes and COCO-Stuff is manual and fiddly (license restrictions, extra model downloads for mIoU)
• README notes “a little differences” between retrained and paper models, which is honest but means you may need to fall back to legacy weights

Verdict

Worth a look if you’re trying to ship interactive image-to-image translation on hardware that isn’t a server rack. Skip it if you’re just training GANs for paper figures and don’t care about inference cost.