The idea in plain language.
A model of the score can guide the reversal of a noising process.
How it works
A stochastic differential equation gradually transforms data into noise. The corresponding reverse-time equation uses the score of the noisy distribution to transform noise back into samples. The paper connects this framework to numerical solvers and an associated deterministic differential equation. That common view supports different sampling procedures and applications such as reconstruction from partial measurements.
What to keep in mind
The learned score is approximate, and numerical choices affect quality and cost. Performance on the reported image tasks does not establish equal results across all data types.
Source: Score-Based Generative Modeling through Stochastic Differential Equations. The original manuscript contains the methods, experiments, figures, and references. An arXiv posting date may follow an earlier conference publication. Read the linked record for version history.