Open cxw-droid opened 4 months ago
Thank you for introducing this paper. Although I did not find the derivation of the SDE there, the paper looks interesting.
No worries! To clarify, that paper translates the notation from the original SDE paper into VDM's, the actual derivation is in the former as mentioned.
Best, Fabio
Hi, thank you for your paper demystifying the "Variational Diffusion Models". In the original paper Variational Diffusion Models version 1 and 2, there is an equation (Eq. 14) that describe a reverse SDE, where $$f(t) = \frac{d\textbf{log}\alpha_t}{dt}$$ and $g^2(t)=\frac{d\sigma^2(t)}{dt} - 2\frac{d\textbf{log}\alpha_t}{dt} \sigma^2(t)$ Would you like to show how the author get this equation? I cannot find it in your current paper. This equation was referenced in other papers such as Progressive Distillation for Fast Sampling of Diffusion Models, see p13