A new technique forces AI generative models to follow strict rules without making their output look fake.
Researchers describe MintFlow, a training-free method for flow matching models, the class of generative systems behind much of today's image synthesis and physical simulation work. Instead of retraining a model or aggressively steering its output to satisfy a constraint, such as a known measurement or a physical law, MintFlow calculates the smallest possible nudge to an in-progress generation step that still gets the final result to comply. That nudge is computed with a closed-form formula derived from adjoint methods, so there is no expensive trial-and-error optimization loop. The method also picks the best moment during generation to apply the intervention, trading off how big the nudge needs to be against how much the model's remaining steps will amplify it.
Constrained generation is the unglamorous but critical layer underneath things like inpainting a photo to match surrounding pixels or simulating a fluid flow that has to respect conservation laws. Prior constrained samplers tend to satisfy the constraint but at a cost: the output drifts away from what the pretrained model actually learned, producing samples that look plausible in isolation but no longer resemble the training distribution. MintFlow's selling point, per the paper's tests across vision and physical-system tasks, is holding that distribution steadier while still meeting the constraint.
That said, this is a single preprint with no peer review and no public code yet, so claims of beating "state-of-the-art constrained methods" deserve the usual wait-and-see treatment until others can reproduce the numbers.