A research team has built a way to dramatically speed up Flow Matching image generators without retraining them.
Flow Matching models are a leading class of generative AI, but they rely on iterative math — solving ordinary differential equations step by step — that makes generation slow. Existing shortcuts force a tradeoff: skip the training overhead and quality tanks at low step counts; pay the training cost and you lose the ability to drop the fix into any existing pipeline. The Bi-Anchor Interpolation Solver, or BA-solver, sidesteps that tradeoff by adding a small companion network — roughly 1-2% the size of the main model — that learns to predict velocities at nearby points in the generation trajectory. The main model stays frozen; the small network fills in the gaps. On ImageNet at 256x256 resolution, the approach matches the quality of a 100-plus step Euler solver in just 10 steps, and holds up at as few as 5.
The gap between research-grade quality and deployable speed has been a persistent cost driver for anyone running image generation at scale. A plug-and-play fix that requires almost no training — and that works with existing pipelines including image editing — is the kind of thing infrastructure teams actually care about. The math here is not novel on its own, but the specific two-anchor scheme for approximating intermediate velocities is a practical engineering contribution.
For context, distillation methods like consistency models also aim at low-step generation but require significant retraining — the tradeoff BA-solver explicitly avoids. Whether the gains hold on video or audio generation models, which face far longer trajectories, is the obvious next question the paper does not answer.