Reproduce “Improved Consistency Regularization for GANs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhengli Zhao, Sameer Singh, Honglak Lee, Zizhao Zhang, Augustus Odena, Han Zhang

Zhao et al. (UCI and Google Research) developed Improved Consistency Regularization (ICR) for Generative Adversarial Networks (GANs), which addresses augmentation artifacts and regularizes the latent space. The method achieves new state-of-the-art Fréchet Inception Distance (FID) scores, including 9.21 for conditional CIFAR-10 and 5.38 for conditional ImageNet-2012 synthesis using BigGAN.

Key insights

  • Unbalanced consistency regularization (only on real images) can mislead the generator into interpreting augmentation features as part of the true data distribution, leading to artifacts in generated outputs.
  • Extending consistency regularization to generated samples effectively eliminates augmentation artifacts, allowing for more aggressive data augmentation strategies.
  • Regularizing the latent space by simultaneously encouraging discriminator insensitivity and generator sensitivity to small latent perturbations is effective for improving both training stability and sample diversity in GANs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.