The data are Murphy's 8-bit swirl (1024 points, from the book's code), noised with the book's schedule: sigma2(t) = sigmoid(γ(t)) with γ(t) = −8 + 13t, so the noised point is zt = √(1 − sigma2) x + sigma ε. This is not the cosine or linear schedule of the earlier version of this page. The marginal qt(z) is a mixture of 1024 Gaussians, so its log density and score are exact. Left: the data. Middle: forward noising (or the reverse sampler). Right: the exact density of the noised mixture at that time, with the points on top.
With gamma(t) = −8 + 13t: sigma2(t) = sigmoid(gamma(t)) and alpha-bart = 1 − sigma2(t). At t = 0 the noise is 0.03 percent of the variance, at t = 1 it is 99.3 percent. The reverse sampler integrates the reverse-time SDE of this variance-preserving process, dz = [f(t) z − g(t)2 ∇ log qt(z)] dt + g(t) dW with f(t) = −6.5 sigma2(t) and g(t)2 = 13 sigma2(t), from t = 1 to t = 0 in 200 Euler steps, with 300 particles and the exact score. The book's notebook has the forward noising, the exact density and the score, but no sampler; the sampler here is built from the same exact score. Its random numbers come from the page's seeded generator, so the points differ from a Python run while the statistics agree.