ReverseDiffusionVariance
Enums · AiDotNet.Enums
Selects the fixed variance sigma_t^2 used by a Gaussian reverse-diffusion step.
For Beginners
When the model removes noise one step at a time, it adds back a little fresh randomness. This setting picks how much: the larger “beta” amount or the smaller “posterior” amount.
How It Works
Ho et al. (2020, “Denoising Diffusion Probabilistic Models”, §3.2) report that both
sigma_t^2 = beta_t and sigma_t^2 = beta_tilde_t (the true posterior variance) work well.
TabDDPM (Kotelnikov et al. 2023, reference code yandex-research/tab-ddpm,
gaussian_p_mean_variance) samples with sigma_t^2 = beta_t, which is the default here.
Fields
| Field | Summary |
|---|---|
ReverseDiffusionVariance FixedLarge |
sigma_t^2 = beta_t (“fixedlarge”), the TabDDPM reference default. |
ReverseDiffusionVariance FixedSmall |
sigma_t^2 = beta_tilde_t = beta_t (1 - alpha_bar_{t-1}) / (1 - alpha_bar_t) (“fixedsmall”). |
int value__ |