As graciously explained to me by one of the authors, Ajay Jain. (Any errors are mine!)
Notes
-
The taxonomy on slide 6 isn't the only way to categorize the models. Some of the approaches could fall under multiple categories -- for example, IAF is a type of normalizing flow and the "IAF" in the diagram refers more to IAF-VAE.
- It's probably most correct to put Denoising Diffusion Probabilistic Models under the "Amortized Variational Inference" node.
- Slide 19 is for GIF visualizations of Langevin dynamics which would have been hard to include in the LaTeX beamer presentation. You can see the animations in this talk.