Reproduce “Demystifying Variational Diffusion Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Fabio De Sousa Ribeiro, Ben Glocker

Demystifying Variational Diffusion Models from Imperial College London reinterprets diffusion models as hierarchical latent variable models (HLVMs) through directed graphical modeling and variational inference. The work shows that common weighted diffusion objectives are equivalent to maximizing the Evidence Lower Bound (ELBO) under Gaussian noise-based data augmentation, bridging theoretical understanding with practical application.

Key insights

  • Diffusion models' fixed forward process with linear Gaussian transitions makes the crucial top-down posterior `q(z_s | z_t, x)` tractable via Gaussian conjugacy.
  • The continuous-time variational lower bound (VLB) remains invariant to the specific shape of the noise schedule function between its endpoints, impacting only optimization efficiency.
  • The "generative feedback" enabled by top-down inference in diffusion models mitigates issues like posterior collapse found in traditional HLVMs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.