New flow matching objective improves generative models on manifolds

Researchers have developed a new flow matching objective for training generative models on manifolds, addressing issues where existing methods fail on compact spaces like the 1D flat torus. Their proposed method, built from an exponential family over endpoints, is compatible and ensures velocities remain feasible on compact manifolds. Empirically, this approach demonstrates improved negative log-likelihoods on geospatial benchmarks and outperforms prior work on human pose datasets. AI

IMPACT This research could lead to more accurate and efficient generative models for complex data distributions found in fields like geospatial analysis and human pose estimation.

RANK_REASON This is a research paper published on arXiv detailing a new method for generative models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv stat.ML →

• 1D flat torus
• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• Connected Papers
• CORE Recommender
• DagsHub
• exponential family
• Gotit.pub
• Hugging Face
• IArxiv Recommender
• Litmaps
• manifold
• Riemannian Flow Matching
• ScienceCast
• scite Smart Citations

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →