Reproduce “Two Simple Principles for Diffusion-Based Test-Time Adaptation” 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

Kaiyu Song, Hanjiang Lai, Yan Pan, Kun Yue, Jian Yin

Researchers from Sun Yat-Sen University formalized two core principles for diffusion-based Test-Time Adaptation and developed PDDA, a method that achieves state-of-the-art performance by guiding image purification with semantic preservation and minimal modification. This approach yielded notable improvements, including a 16.3% accuracy increase on CIFAR-10C compared to prior diffusion-based methods.

Key insights

  • Effective diffusion-based Test-Time Adaptation can be achieved through two simple, guiding principles: preserving semantic content and making minimal alterations.
  • The proposed patch-aware feature aggregation across all UNet layers efficiently captures comprehensive semantic information for guidance.
  • A hidden conflict between the "semantic keeper" and "modification keeper" requires resolution via gradient projection to avoid unstable generation and enhance accuracy.

Summary by alphaXiv.

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