Standard flow matching creates trajectory crossings that sharply increase the local Lipschitz constant of the velocity field. This makes the field difficult to learn and amplifies numerical errors during fast sampling. The authors propose CoFlow, which introduces a repulsive drift during training to push positive sample paths away from negative ones, lowering the Lipschitz constant and smoothing the field. On ImageNet 256x256, CoFlow reduces FID at 20 inference steps compared to standard flow matching without extra training cost.
Repelling crossing trajectories makes flow matching smoother and easier to sample
CoFlow modifies intermediate training paths using negative samples, reducing sharp velocity changes that impair few-step image generation.
Academic
Ziqi Jiang · Zhenqi He · Long Chen
The Hong Kong University of Science and Technology
Research Digest··3 min read
In one line
Contrastive repulsion of training trajectories produces smoother velocity fields for few-step flow matching.
Read it as
Jiang, He and Chen present CoFlow, a flow-matching method that repels nearby, conflicting trajectories during training rather than changing endpoint couplings or applying post-training distillation.
The authors reinterpret trajectory crossing in flow matching through local Lipschitz constants, which measure how sharply the target velocity changes between nearby points.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.