Back to index
Other18 lectures, around 30 hours with the reading·Free on YouTube with public course notes

Stanford CS236: Deep Generative Models

4.5

The course that makes diffusion models stop being magic. Heavily mathematical, unusually well organised, and the best free treatment of generative modelling as a coherent theory rather than a list of architectures.

What We Liked

  • Treats every generative family as an answer to the same question, which makes the whole field click
  • The score based diffusion lectures are the clearest public explanation of that material anywhere
  • Ermon is a primary contributor to score based methods, so the diffusion coverage is authoritative
  • Full lecture videos, slides and notes are public and free
  • Prepares you to read current generative modelling papers rather than just use the tools

What Could Be Better

  • The mathematical demand is genuinely high and the course does not soften it
  • Very little practical implementation, so you finish understanding models you have not built
  • The 2023 recordings predate several important developments in video and flow matching
  • Assignments and solutions are not fully available outside the enrolled course
  • No coverage of the engineering side, so nothing about training at scale or inference cost

Detailed review

There is a large gap between people who can use Stable Diffusion and people who can explain what it is doing, and CS236 is the bridge. Stefano Ermon's course takes generative modelling seriously as a subject with a theory, and the payoff is that by the end you can see autoregressive models, variational autoencoders, normalizing flows, GANs, energy based models and diffusion as different trade offs on the same underlying problem rather than as six unrelated tricks. That framing is the course's central contribution. The problem is always the same: you want to represent a probability distribution over high dimensional data, you want to be able to sample from it, and you want to be able to learn it from examples.

Every architecture is a different answer to which of tractable likelihood, easy sampling and expressive family you are willing to give up. Once you have that grid in your head, papers you previously found impenetrable become legible, because you can immediately place a new method by what it sacrifices. The diffusion material deserves particular attention. Ermon is one of the people whose work established the score based view of diffusion, and the lectures on score matching, Langevin dynamics and stochastic differential equations are the clearest public treatment of that material I have found.

If you have read the score based generative modelling papers and bounced off them, watching lectures thirteen through sixteen is likely to be the thing that fixes it. The mathematical demand is the main barrier and I will not soften it. You need probability at a level where conditional independence, KL divergence, change of variables and expectation manipulation are tools rather than obstacles. You need to be comfortable with the mechanics of variational inference by the time VAEs arrive.

Multivariate calculus is assumed throughout. People who try to take this course with a machine learning background but a shaky probability foundation tend to stall around lecture six. If that is you, spend a fortnight on the probability material first and come back. The complementary weakness is practice.

This is a theory course and it is honest about that. You will not build a diffusion model in it. There are no substantial guided implementations available to outside learners, and the enrolled assignments are not fully public. The result is that you can finish CS236 with an excellent understanding of models you have never trained, which is an odd position to be in.

My recommendation is to pair it with the Hugging Face diffusion course or with a from scratch implementation of a small diffusion model, so the theory has something to attach to. Currency is a minor issue. The widely available recordings are from the 2023 offering, which means flow matching, rectified flows and the newer video generation work are either absent or only briefly touched. The foundations are unaffected, and the whole point of a foundational course is that it makes the newer papers readable, so I would not weight this heavily.

There is also nothing about engineering. Training cost, distributed training, sampling speed, quantisation and everything else that determines whether a generative model is deployable falls outside the scope. Do not come here for that. My 4.5 reflects a course that does one difficult thing at a very high standard and is completely free.

If you work with image, audio or video generation and you have only ever operated the tools, this is the material that turns you into someone who can reason about them.

[ final ]

The verdict.

Take this if you want to understand why diffusion works rather than how to call a pipeline. It demands real probability, and it repays that demand better than almost any course I have looked at.