Table of Contents
🌏 中文版
Lecture 4 of MIT 6.S191 2026 is Generative Modeling: From Latent Spaces to Diffusion. It Separates generative from discriminative tasks, organizes VAE, GAN, and diffusion objectives, and leads into Lab 2’s DB-VAE. This note uses only the official 2026 slides and video; it does not mix in similarly named material from 2025.
What to take away
- Explain how latent variables can represent factors of variation
- Distinguish reconstruction, adversarial, and denoising objectives
- Recognize that sample quality does not imply that data bias is gone
These goals have one thing in common: recognizing terminology is insufficient. You should be able to identify inputs, outputs, the learning signal, and the main constraint before moving on.
A VAE encodes an input as a distribution, samples a latent value, and decodes a reconstruction. A GAN pits a generator against a discriminator, while diffusion learns to reverse a noising process. Because their objectives differ, “better generation” must be defined in terms of fidelity, diversity, and downstream use.
How to watch
First scan the sections and diagrams in the official slides, then watch the official video. On a second pass, pause at equations and architecture diagrams and redraw them in your own notation. Afterward, close the material and write three central ideas plus one unresolved question.
An exercise for tonight
Draw a VAE with encoder, sampling, and decoder; label both loss terms before opening Lab 2 Part 2.
“Finished” means leaving a checkable diagram, calculation, program output, or short note—not merely reaching the end of the video. You should also be able to explain one failure mode to someone else.
Scope and limits
6.S191 is a high-intensity introduction, and this article is only a lecture guide. It does not replace the full recording, rigorous derivations, or instructor feedback. Use a semester course or primary papers when a topic needs theoretical depth.
References
Loading...