Reproduce “Weak-to-Strong Diffusion with Reflection” 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

Lichen Bai, Masashi Sugiyama, Zeke Xie

The 'Weak-to-Strong Diffusion with Reflection' framework introduces an inference enhancement for diffusion models that reduces the gap between learned and real data distributions. It uses a reflective sampling strategy to leverage the estimated difference between a strong and a weak model, consistently improving generation quality, human preference metrics, and prompt adherence across various model architectures and tasks.

Key insights

  • The estimable 'weak-to-strong difference' (difference in probability density gradients between a strong and weak model) can effectively approximate the inaccessible 'strong-to-ideal difference' (difference between the ideal data distribution and the strong model).
  • The reflective operation theoretically steers latent variables towards the true data distribution by perturbing them in the direction of this weak-to-strong difference.
  • W2SD provides a unifying theoretical understanding for several previously heuristic inference enhancement methods (e.g., Re-Sampling, Z-Sampling) by reinterpreting them as specific instances of its generalized framework.

Summary by alphaXiv.

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