Denoising Diffusion Probabilistic Models
A forward process gradually turns data into noise; a network learns to reverse it one step at a time. The paper shows the variational objective simplifies to predicting the added noise, and that this simple loss produces high-quality image samples.
Key ideas
- Generation as iterative denoising
- The training objective reduces to a noise-prediction regression
- Sampling quality depends on the noise schedule and the number of steps
Why read it now
Diffusion is the dominant generative approach for images and audio, and its derivation ties together the VAE bound, score matching, and Langevin dynamics.
Question to keep in mind
Why does predicting the noise work better than predicting the clean image directly, given they are linearly related?