Generative models create images or robot motions from examples, and users often want to steer the output, say to make a robot avoid an unseen obstacle. gradually remove noise; flow matching is a newer cousin that is cheaper and more stable, but steering it has been unreliable.
Flow matching trains by pairing each noisy intermediate sample with the exact final data point it should become. Older steering methods move only the final point, breaking that pairing. Coupled Guidance corrects both sides together, preserving the training structure.
In tests on robot collision avoidance, dexterous manipulation, and image inpainting, deblurring, and super-resolution, it improved stability and sample quality with no extra computation and little tuning, making reactive control and image editing more practical.