Title: Low-Pass Flow Matching

Abstract: Standard Flow Matching methodologies predominantly utilize white noise, a selection that frequently conflicts with the spectral characteristics of natural data, which typically exhibit frequency decay. To bridge this gap, we present Low-Pass Flow Matching, a novel variation grounded in an operator-modulated interpolant. This approach generates a dynamic spectral bias that shifts from the initial source distribution toward a frequency-decaying profile as the trajectory nears the target data. We evaluate the efficacy of this technique on unconditional image generation benchmarks, such as the Galaxy10 scientific dataset. Our experimental results demonstrate that the method achieves superior performance, especially when integrated with adaptive ODE solvers. In these scenarios, the proposed approach maintains or enhances sample fidelity while significantly lowering the computational cost of sampling relative to conventional baseline methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC