The idea in plain language.
Learn which direction makes a noisy point more like the training data.
How it works
The model estimates the gradient of the log data density, known as the score. Training uses several noise levels to make that estimation more manageable. Generation starts from noise and repeatedly follows learned score directions while adding controlled randomness. Decreasing the noise level over the process allows the sample to develop increasingly detailed structure.
What to keep in mind
Sampling quality and cost depend on the learned scores and the numerical procedure. A convincing generated sample is not evidence that its depicted content is factual.
Source: Generative Modeling by Estimating Gradients of the Data Distribution. 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.