A ground-up technical course on diffusion models — the architecture behind modern image generators like Stable Diffusion. Starting from the intuition of adding and removing noise, it derives the forward process, the variational lower bound, the simplified training objective, the U-Net backbone, the DDPM training and sampling loops, fast deterministic samplers (DDIM), and finally conditioning and guidance.
Audience: technically comfortable; familiarity with deep learning recommended.
Prerequisites: Gaussian distributions, KL divergence, basic variational inference. The LLM course
is a natural companion.
Appendices
| A | Acronyms |
| B | Formulas & Symbol Glossary |