MintFlow, a constrained-sampling method: a small nudge that lets pretrained generators obey real-world constraints

A pretrained generative model can paint plausible images, simulate fluid flow, or fill in missing sensor data. It cannot, on its own, guarantee that a generated tornado follows the conservation of energy, or that a reconstructed medical scan matches a known measurement. That gap between what a model "knows" and what a user actually needs has forced research teams into a familiar bind: retrain the model under the new rules, or run a costly post-hoc fix that warps the output.

A new MintFlow paper argues there is a third path. The method treats constraint satisfaction as a minimal intervention on a pretrained flow model's trajectory: find the single intermediate state whose nudged value, propagated forward through the original flow, lands exactly on the constraint, while disturbing the data distribution as little as possible.

The trick is that this nudge has a closed-form solution. Flow matching is a family of generative models that learn a time-dependent vector field mapping noise to data, and it has become a workhorse for image synthesis and physical simulation. Constrained sampling, the task MintFlow targets, asks the generator to produce samples that satisfy a specific rule, such as a known boundary condition or an observed measurement, on top of looking like the training data. To find the minimal perturbation that steers a sample toward such a rule, the authors derive an adjoint expression that reads the gradient of the constraint against the flow field directly. No iterative optimization, no inner loop, no retraining. One linear solve per constrained sample.

Existing constrained samplers tend to pick one side of this trade-off and lose the other. Methods that strictly enforce the constraint often push the output far from the pretrained distribution, producing samples that satisfy the rule but look or behave unrealistically. Methods that try to stay close to the data distribution often soften the constraint, leaving the user to decide how much rule-breaking to tolerate. MintFlow's pitch is that the nudge can be sized correctly: small enough to keep the output distribution intact, large enough to land the constraint exactly.

Two design choices make that pitch workable. First, the closed-form adjoint solve removes the per-sample optimization loop that earlier methods needed. Second, the framework adaptively picks when to intervene along the flow, trading off the magnitude of the nudge against its amplification by the remaining path. Intervene too early and a tiny correction can be magnified; intervene too late and the model has already drifted away from a region where the constraint is satisfiable. The preprint reports that this adaptive timing, combined with the closed-form direction, yields competitive constraint satisfaction and substantially better preservation of the pretrained distribution than state-of-the-art constrained baselines.

The authors evaluate the method on generative vision and on physical system modeling, the latter being the kind of benchmark where a constraint like a partial differential equation is non-negotiable. The reported gains are author-claimed. The work is an arXiv preprint, not a peer-reviewed paper, and the abstract names no deployment partners or production use.

Two preconditions also bound the claim. The adjoint derivation requires that the constraint and the flow field be differentiable and smooth enough for the gradient to make sense. Forcing a discontinuous rule, such as a hard integer threshold, through MintFlow would not be straightforward. The "minimal intervention" guarantee is mathematical within a model class, not a universal one.

A team that already runs a pretrained flow model and needs it to honor a known measurement can, in principle, wrap MintFlow around the inference loop and pay roughly one extra linear solve per sample, instead of retraining the model or running a black-box optimizer. Whether that advantage shows up outside benchmarks is the next question.

The full preprint is available on arXiv as MintFlow: Minimal Trajectory Intervention for Constrained Flow Matching.