Learning the Integral of a Diffusion Model | Hacker News

A Hacker News post links diffusion models to normalizing flows and reversible mappings for continuous distributions.

It explains how ODE-based models compute the log-determinant term needed for likelihood-based training.

The post also notes that diffusion models can be viewed as approximating this setup with a stochastic partial differential equation.

Hutchinson’s estimator is mentioned as a practical tool for handling the log-determinant computation.