mit-han-lab/gan-compression

Health v2 · maintenance only

Flags: no_license

Adoption not part of the score

1115 stars · 149 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

A PyTorch research codebase implementing GAN Compression, a CVPR 2020 method that reduces the computation of conditional GANs like pix2pix, CycleGAN, MUNIT, and GauGAN by 9-29x via once-for-all distillation and architecture search. It includes pretrained models, training/test scripts, Colab notebooks, and an interactive demo targeting edge devices like Jetson Nano.

Use cases

• compress a pretrained pix2pix or CycleGAN generator for faster inference
• run image-to-image translation models on edge devices like Jetson Nano
• distill a smaller student GAN from a large teacher generator
• search for efficient sub-generators under a FID or compute budget
• reproduce the GAN Compression paper results
• speed up GauGAN for interactive synthesis

When to choose

• you need to shrink conditional GAN models for real-time or embedded deployment
• you want a research baseline for GAN distillation and once-for-all network compression
• you work with pix2pix, CycleGAN, MUNIT, or GauGAN in PyTorch

When to avoid

• you need production-supported, well-maintained software with a permissive license
• you want to compress non-GAN models like classifiers or LLMs
• you need Windows or macOS support, since it targets Linux with GPUs

Facets

library · maturity maintenance

machine-learning image-processing deep-learning machine-learning computer-vision image-processing python gan-compression model-compression knowledge-distillation neural-architecture-search pix2pix cyclegan pytorch research-code cvpr-2020 research linux gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="mit-han-lab/gan-compression")