Search papers, labs, and topics across Lattice.
This paper refines the theoretical convergence properties of Diffusion Flow Matching (DFM) by deriving improved Kullback-Leibler (KL) divergence and 2-Wasserstein distance bounds, focusing on the discretization error associated with Brownian motion-based DFMs. The authors achieve state-of-the-art scaling in terms of dimensional dependence under minimal conditions, which enhances the understanding of DFM's performance in high-dimensional settings. This work not only clarifies the convergence behavior of DFMs but also provides a robust framework for future generative modeling research.
Improved KL and Wasserstein convergence bounds for Diffusion Flow Matching reveal state-of-the-art performance in high dimensions under minimal assumptions.
Diffusion Flow Matching (DFM) has recently emerged as a versatile framework for generative modeling, yet its theoretical convergence properties remain only partially understood. In this work, we provide refined and novel convergence guarantees for Brownian motion based DFMs, focusing on the discretization error. Our analysis is conducted under the Kullback-Leibler (KL) divergence and the 2-Wasserstein distance. Under finite-moment conditions and a mild score integrability assumption, we derive KL convergence bounds with improved dimensional dependence compared to prior work, achieving, up to our knowledge, state-of-the-art scaling under minimal conditions. We further extend the analysis to the 2-Wasserstein distance: under an additional first-order score integrability assumption and a weak log-concavity condition, we obtain convergence guarantees with dimensional dependence consistent with the KL case.
