Why it matters

Score matching underlies diffusion theory. Understanding shapes fundamentals.

Advertisement

The architecture

Nabla log p(x): score function.

Denoising score matching: perturb + predict noise.

Equivalent to diffusion training.

Score matching structureScore functionnabla log p(x)Perturb xadd noisePredict scorelearned by MSESong + Ermon 2019 unified score-based and diffusion perspectives
Score matching.
Advertisement

How it works end to end

Denoising score matching (DSM).

Score = -noise / sigma up to constant.

Yang Song's noise-conditional score networks (NCSN).