Google’s modular GAN rig: one codebase, eight papers, many metrics

What it does

compare_gan bundles TensorFlow implementations of common GAN building blocks—losses, penalties, normalization schemes, architectures, and evaluation metrics—into a single, Gin-configurable codebase. Researchers mix and match components from a shared menu instead of reimplementing BigGAN or ResNet from scratch, then run experiments on GPUs, TPUs, or CPUs. It functions like a laboratory instrument for controlled generative-model studies.

The interesting bit

The framework treats GAN research as a combinatorial experiment: keep the data pipeline and training loop constant while swapping losses, normalization, or architecture. That discipline is how Google produced papers like Are GANs Created Equal? and The GAN Landscape—by controlling variables in one shared rig rather than comparing numbers across disparate repositories.

Key highlights

• Bundles losses (non-saturating, least-squares, WGAN), penalties (gradient penalty), normalization (spectral, batch, layer), and architectures (BigGAN, ResNet, DCGAN)
• Includes evaluation metrics: FID, Inception Score, precision-recall, and KID
• Configurable via Gin, so experiments are defined declaratively rather than by editing source code
• Runs on GPU, TPU, or CPU
• Backed at least eight published papers, including NeurIPS 2018 and ICML 2019 work

Caveats

• Evaluation on TPUs is not supported; you will need a separate GPU VM for that
• Training and evaluation simultaneously can exhaust memory on a single local GPU
• ImageNet and CelebA-HQ require manual dataset download and preparation

Verdict

Researchers running GAN ablation studies or reproducing Google’s generative papers should start here. If you need a plug-and-play image generator or a PyTorch-native tool, this is not your repo.