学术研究★★★★arXiv · 2026-07-09
Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling
A recent study shows that score matching does not guarantee numerical stability in diffusion sampling. The researchers construct a smooth score field with arbitrarily small forward-marginal error, but its reverse-time process may still lose numerical stability. This finding has significant implications for the development of stable diffusion sampling algorithms.
📌 Key points
- Score matching does not guarantee numerical stability
- A smooth score field can lead to numerical instability
- Euler-Maruyama discretization may lose numerical stability
本页为 gitzw.com 基于公开来源的 AI 中文解读,非原文转载。