Everyone's talking about diffusion, but if you actually want to understand the family tree that produced Sora and modern generative video, you need to spend some time with normalizing flows — the class of generative models that made tractable, exact-likelihood density estimation possible in deep learning. This review pairs two foundational papers: Rezende and Mohamed's original Variational Inference with Normalizing Flows, and Kobyzev, Prince, and Brubaker's later comprehensive survey of the field's construction techniques and current methods.

The walkthrough covers how normalizing flows transform a simple base distribution into an arbitrarily complex one through a sequence of invertible mappings, why that construction gives you both efficient sampling and exact density evaluation in one shot, and how finite versus infinitesimal flows fit into the taxonomy of approximate posterior families. It also traces the through-line into current generative AI: why flow-based ideas keep resurfacing in speech synthesis, image generation, and video models even as diffusion dominates the headlines, and where the open research questions still sit. If you're serious about the theory underneath modern generative systems, this is a piece of the foundation you don't want to skip.