MIT's 2026 flow matching, diffusion course debuts

This is less a launch than a public syllabus for the modern diffusion stack, and that's exactly why it matters. It gives learners a rare path from the SDE math all the way to working generators, without leaving implementation as an exercise for later.

• –The 2026 edition adds latent spaces, diffusion transformers, and discrete diffusion, matching where the field is headed.
• –The scope spans images, video, proteins, and language, so it frames diffusion as a general generative framework rather than a vision-only trick.
• –Three labs plus self-contained notes make it far more useful than a typical lecture archive for teams that want to learn by building.
• –MIT-hosted materials make it a strong reference for onboarding researchers or engineers into the space.