Reproduce “Correcting Source Mismatch in Flow Matching with Radial-Angular Transport” 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
Fouad Oubari, Mathilde Mougeot
Radial–Angular Flow Matching (RAFM) improves generative modeling for non-Gaussian data by explicitly matching the data's radial distribution at the source and employing radius-preserving spherical transport paths. This framework achieved competitive or superior performance compared to existing non-Gaussian models, often with substantially reduced training times, such as outperforming MSGM on PIV (d=256) across all metrics while training significantly faster.
Key insights
- A radial-angular Kullback–Leibler (KL) divergence decomposition formally quantifies the source mismatch, showing that matching the data's radial law removes a significant radial penalty.
- Decoupling radial and angular transport allows the model to focus primarily on angular alignment, as the radial component is addressed by source matching and path design.
- Inference-time tangential projection of the learned vector field is crucial for maintaining radial fidelity, especially in high-dimensional and challenging data regimes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.