Week 12: Generative models II: diffusion

DATA 443/643 · Week 12

Module 4 — Generative models II: diffusion

Notes

• 12.1 The Forward Process(slides)
• 12.2 The Reverse Process(slides)
• 12.3 Latent Diffusion and Text Conditioning(slides)

Labs

• Lab 12: Noise a Point, Then Take One Reverse Step(PDF)

Videos

• Umar Jamil: How diffusion models work — explanation and code (forward process)
• Umar Jamil: Coding Stable Diffusion (watch 0:00–45:00: VAE, CLIP, UNet)

Additional readings

• Stanford CS231N 2025 L14: diffusion, LDM (D=8, C=16), classifier-free guidance
• Ho et al., Denoising diffusion probabilistic models
• Rombach et al., Latent diffusion / Stable Diffusion

Discussion

• Why sample x_t from q(x_t | x_0) instead of walking t Markov steps in the loss?
• If ε_θ equals the true ε, what should one reverse mean do in the 12.1 cartoon?
• s=1 versus s=12: which is “no text,” and what happens to variety?
• 256×256×3 versus 32×32×16: whose bug is a slow sampler if you still denoise pixels?