Reproduce “Cascaded Diffusion Models for High Fidelity Image Generation” 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

Jonathan Ho, Chitwan Saharia, William Chan, David J. Fleet, Mohammad Norouzi, Tim Salimans

This paper presents Cascaded Diffusion Models by Google, which achieve state-of-the-art image generation quality on class-conditional ImageNet at 256x256 resolution, with an FID of 4.88, without using auxiliary classifiers. The method introduces conditioning augmentation to resolve compounding errors in multi-stage generation, resulting in superior Classification Accuracy Scores of 63.02% Top-1.

Key insights

  • Conditioning augmentation is essential for the effectiveness of cascaded generative models, successfully mitigating the compounding error that otherwise degrades sample quality.
  • The optimal level of conditioning noise is non-monotonic; both too little (leading to train-test mismatch) and too much (over-corrupting input) negatively impact performance.
  • Different forms of conditioning augmentation (Gaussian noise vs. Gaussian blur) are effective for different upsampling resolutions, with blur being more beneficial for higher resolution stages.

Summary by alphaXiv.

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