Week 07 — Diffusion Models and Image Generation
Diffusion models for image generation: DDPM's denoising principle, latent diffusion (the engine behind Stable Diffusion), and how the conditional architectures of 2026 build on both.
Week 07 — Diffusion Models and Image Generation
Diffusion models for image generation: DDPM's denoising principle, latent diffusion (the engine behind Stable Diffusion), and how the conditional architectures of 2026 build on both.
Lecture
The forward and reverse diffusion processes · score-matching · DDPM, DDIM, ancestral sampling · classifier-free guidance · latent diffusion (Stable Diffusion) · ControlNet, LoRA for diffusion · image-to-image, video-from-image · evaluation (FID, CLIP-score, human eval).
Read before the lecture
- Ho, Jain, Abbeel, *Denoising Diffusion Probabilistic Models* (NeurIPS 2020)
- Rombach et al., *High-Resolution Image Synthesis with Latent Diffusion Models* (CVPR 2022, the Stable Diffusion paper)
Recitation — paper discussion
Song et al., *Score-Based Generative Modeling through Stochastic Differential Equations* (ICLR 2021) (paper)
Come ready to argue one side of each:
- Why is the SDE framing more general than the discrete-time formulation?
- What does the unification with score matching give you that DDPM alone doesn't?
Reference text for this week: chapter 07 of the bilingual notes — EN PDF · FR PDF.