Diffusion models
Destroying data is easy; add a little noise, a thousand times. A diffusion model learns to undo one small step of that destruction, and generates by undoing all of them, starting from pure noise. The forward process in closed form, DDPM's noise-prediction loss derived, noise schedules and sampling steps, DDIM, classifier-free guidance, and latent diffusion, which is how text-to-image systems work.
65 min
After this lesson you can
- Write the forward noising process and derive the closed form of
- Explain why the reverse step is approximately Gaussian, and state the posterior
- Derive how the ELBO for a diffusion model reduces to predicting the added noise, and what drops
- Implement DDPM training and ancestral sampling, and explain what DDIM changes
- Explain classifier-free guidance, and how latent diffusion with cross-attention turns text into images
In this lesson
- The forward process: destroying data, one step at a time
- The reverse process: undoing one small step
- The loss: predict the noise
- Sampling
- Noise schedules
- Steering: conditioning and guidance
- How text-to-image systems work: latent diffusion
- The code
- In the papers
This lesson is part of Descent Pro
Descent Pro unlocks every lesson on the route, from base camp to the summit. Every plan starts with a free trial.