Study links denoising to Stein identities for learning hidden structure in diffusion models
A new paper analyzes score learning from denoising pairs when data contain a hidden non-Gaussian component, showing that if covariance reveals the component, its subspace can be identified, with moment access and a resource hierarchy for learning hidden structure.
Read the full story at arXiv stat.ML ↗
Timeline · 1 report
- 2026-11-04 05:00 · arXiv stat.ML
Denoising Is Stein: Moment Access and a Resource Hierarchy for Learning Hidden Structure in Diffusion Models