Reproduce “Carré du champ flow matching: better quality-generalisation tradeoff in generative 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

Jacob Bamberger, Iolo Jones, Dennis Duncan, Michael M. Bronstein, Pierre Vandergheynst, Adam Gosztolai

Carr´e du champ Flow Matching (CDC-FM) enhances generative model performance by integrating geometry-aware noise into the flow matching process. This approach effectively improves the quality-generalisation tradeoff by reducing training data memorisation, demonstrating superior performance in data-scarce or geometrically complex scenarios across diverse datasets, including LiDAR and single-cell genomics.

Key insights

  • Explicitly regularizing the generative path by aligning it with the local data manifold geometry can effectively prevent the collapse onto individual training points, thereby mitigating memorisation.
  • The Carr´e du Champ field (`bΓ(x)`) provides a principled way to incorporate local smoothness information as an inductive bias, inducing geometry-aware anisotropic diffusion into the continuity equation.
  • By promoting flow normal to the data manifold, CDC-FM enables the generation of genuinely novel samples that accurately represent the underlying data distribution, improving generalisation without sacrificing quality.

Summary by alphaXiv.

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