Key Takeaways
DDPMs achieve state-of-the-art image synthesis by learning to reverse a fixed Gaussian diffusion process.
The model defines a forward process that gradually adds Gaussian noise and a learned reverse process that denoises.
A novel parameterization of the reverse process reveals an equivalence to denoising score matching and annealed Langevin dynamics.
A simplified training objective (L_simple) that predicts the noise component (epsilon) leads to superior sample quality.
DDPMs exhibit progressive lossy compression, generating coarse features first and refining details over time.
The sampling procedure can be interpreted as a generalization of autoregressive decoding with a flexible 'bit ordering'.