Position in the series

hierarchical latent variable -> DDPM variational bound

Outline

  1. Forward noising process as fixed inference path.
  2. Reverse denoising process as generative model.
  3. DDPM as a Markovian hierarchical latent variable model.
  4. Variational lower bound decomposition.
  5. Connection between KL terms and denoising prediction.
  6. From variational bound to simplified training objective.

Notes

Write here.