Reproduce “Differentiable Augmentation for Data-Efficient GAN Training” 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

Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, Song Han

Researchers from IIIS, Tsinghua University, MIT, Adobe, and CMU developed Differentiable Augmentation (DiffAugment), applying consistent differentiable transformations to both real and generated images during GAN training. This method enables high-fidelity image generation with significantly reduced data, achieving competitive performance on 100-shot datasets and mitigating discriminator overfitting.

Key insights

  • Applying differentiable augmentation uniformly to both real and fake images during GAN training acts as a powerful regularizer, effectively preventing discriminator overfitting to specific training samples.
  • The differentiability of the augmentation allows the generator to receive proper gradients, enabling it to learn from augmented samples without producing artifacts or distorting the target data distribution.
  • Simple, fixed differentiable augmentation policies (e.g., color + translation + cutout) are sufficient to dramatically improve GAN data efficiency and training stability across diverse datasets and architectures.

Summary by alphaXiv.

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