Equation → implementation → evidence

Diffusion & Flow Lab

Explore probability paths, scores, transport and generative models. The analytic panels calculate exact mixture quantities in your browser; trained galleries expose the published experiments and their limitations.

Repository guide · Learning curriculum · Review and comparisons
Live analytic mixture

Forward Gaussian corruption

Eight components, radius 3, component standard deviation 0.25. Change the path and time; fixed paired particles show Xₜ = αX₀ + σε. Forward time runs from data to noise. Arrows use the exact marginal score.

Live equations and code mapping

Score and posterior mean

The noisy observation used below is (1, −0.5). Tweedie gives a conditional mean, which can fall between mixture modes. It is neither a fresh conditional sample nor a guaranteed denoised ground truth.

Definitions, time conventions and runtime code