For some time now I did not know exactly why we can compute the likelihood of a sample with a diffusion model. In this note, I discuss how the ODE nature of a diffusion model makes exact likelihood evaluation possible.

1

In this note, I some of the most basic content about Stein's method, which can be naturally adopted to solve computational and statistical challenges in practical machine learning.

2

In this note I mainly supplementing the skipped derivation details of the paper [How to Train Your Energy-Based Models] by Song and Kingma.

3