Divergence-Suppressing Couplings Improve Rectified Flow

A new paper on arXiv proposes divergence-suppressing couplings for Rectified Flow, an offline correction method that attenuates divergent components in learned velocity fields during coupling generation. The authors identify that trajectory entanglement in Rectified Flow is often linked to regions of nonzero divergence, where local expansion or contraction distorts trajectories. The correction is applied once per coupling pair and amortized over training, allowing deployment at identical wall-clock cost to standard Rectified Flow. Empirical results show consistent improvements on 2D synthetic data.