One-Sided Quantile Coupling for Flow Matching

2026-08-02

AI Summary

This work addresses the computational inefficiency of conventional flow matching methods, which rely on intra-batch sample pairing and incur at least quadratic complexity in batch size, thereby limiting optimization efficiency and generation quality. The authors propose a lightweight one-sided coupling mechanism that, for the first time, incorporates a projected rank structure into flow matching: by sampling only a data batch and mapping its ranks along random orthogonal directions to Gaussian quantiles, source samples are constructed directly without explicitly solving an optimal transport problem. This approach preserves a standard Gaussian prior while eliminating irreducible regression variance along slices, enforcing strictly linear ideal flow trajectories. Experiments demonstrate consistent and significant improvements over baselines across CIFAR-10, CelebA, FFHQ, and ImageNet-64, with FID scores reduced by up to 12.9%, comprehensively outperforming OT-CFM and confirming the method’s efficiency and scalability.

Abstract

Flow Matching trains continuous-time generative models by regressing the velocity field of a probability path between a simple source distribution and a target data distribution. The coupling that pairs source and target samples strongly affects optimization and sample quality, but structured couplings typically rely on mini-batch transport or assignment procedures whose cost grows at least quadratically in batch size. We propose Quantile Coupling Flow Matching (QC-FM), a lightweight one-sided coupling: rather than matching two pre-sampled batches, it samples only the data batch and constructs each paired source directly. Data ranks projected along a small number of random orthogonal directions are mapped to Gaussian quantiles, and the latent code is completed in the orthogonal complement by conditional Gaussian sampling. The construction is one-dimensional per slice, so the coupling requires no pairwise cost matrix and no assignment to solve. We show that, for each drawn frame, this coupling eliminates the irreducible regression variance along every selected slice and makes the ideal flow exactly straight there, while leaving the sampling prior unchanged: generation still starts from the standard Gaussian, and the training source deviates from it only through the copula of the slice codes, whose transport cost we bound. For training, we apply QC to an anchor subset and complete the remaining source slots with exact Gaussian samples, retaining the QC bias while preserving an explicit signal from the Baseline coupling. Across CIFAR-10, CelebA, FFHQ, and ImageNet-64, QC-FM improves over the Baseline under matched training budgets, reducing FID by up to 12.9%, and outperforms OT-CFM on all four datasets. These results suggest that preserving projected rank structure is a simple and scalable way to inject useful geometric bias into FM couplings without solving a mini-batch transport problem.

Problem

• Flow Matching
• coupling
• quantile
• transport cost
• generative modeling

Innovation

• Quantile Coupling
• Flow Matching
• One-sided Coupling
• Rank-preserving Transport
• Conditional Gaussian Sampling