Reproduce “Exploring Structured Semantic Priors Underlying Diffusion Score for 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
Mingjia Li, Shuang Li, Tongrui Su, Longhui Yuan, Jian Liang, Wei Li
Researchers at the Beijing Institute of Technology and Beihang University developed DUSA, a test-time adaptation (TTA) method that extracts structured semantic priors from diffusion models using a single-timestep score function. DUSA achieves substantial performance gains on image classification and semantic segmentation under domain shift, outperforming prior diffusion-based TTA methods with improved computational efficiency.
Key insights
- The score function of a data distribution implicitly encodes structured semantic priors, expressible as conditional probabilities, which can be directly extracted from diffusion models.
- Effective semantic priors for TTA can be derived from a diffusion model's noise predictions at a single intermediate timestep, making diffusion-based adaptation significantly more efficient.
- Diffusion models, even when adapted unconditionally (DUSA-U), inherently possess strong semantic understanding that can be effectively transferred to task models for robustness against domain shift.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.