A team of researchers from UCLA and Brookhaven National Laboratory has proposed a new way to constrain flow-matching generative models without retraining them or continuously steering their internal dynamics. The method, called MintFlow, makes a deliberately small correction to an intermediate point in a model’s generation trajectory, then lets the original pretrained flow carry that corrected state to its final output.

The approach targets a persistent problem in controlled generation: a model may need to obey an observation, preserve part of an image, or respect a physical law, but forcing that condition can push the result away from the distribution the model originally learned. In their paper, submitted to arXiv on September 30, the authors frame that conflict as a search for the smallest intervention that can meet a terminal constraint.

MintFlow first generates a normal trajectory and measures how far its endpoint is from the required condition. It then uses an adjoint sensitivity calculation, run backward through the trajectory, to estimate how a change at an earlier time would affect the final constraint. That avoids explicitly building the full high-dimensional Jacobian of the flow map. For a selected intervention time, the method computes a minimum-norm correction and reintegrates the unchanged pretrained flow from that point.

A single midpoint adjustment guides a flowing trajectory to a constrained endpoint.
The method uses backward sensitivity to choose a minimum-norm correction at an intermediate time.

Timing is part of the design rather than a fixed setting. MintFlow evaluates several candidate intervention points and balances two considerations: how large a correction each point requires and how much trajectory remains to amplify that change. If one correction does not reduce the residual below a chosen tolerance, the procedure can apply later iterative refinements.

The authors tested the method on noise-free image inpainting, super-resolution and deblurring using 64-by-64 images from AFHQ-Cat and FFHQ. Against five training-free baselines using the same frozen Rectified Flow++ models and source noise, they report that MintFlow produced the lowest constraint error and FID across every evaluated inverse task, while also leading the reported PSNR, SSIM and LPIPS results. Those findings come from the paper’s own experiments and have not been independently validated.

On text-guided image editing with PIE-Bench and InstaFlow, MintFlow recorded the lowest constraint error and the highest prompt-adherence score among the compared methods. Freedom retained slightly higher similarity to the reference image, which the authors interpret as a different tradeoff: stronger preservation but weaker compliance with the requested edit. In the reported setup, MintFlow also had the shortest inference time, at 8.79 seconds, though it tied FlowChef for the lowest listed peak memory.

Image reconstruction and a physical heat field are linked by a constrained generative flow.
The paper tests the approach on image restoration, text-guided editing, and two physical-system benchmarks.

The paper also evaluates generated solutions to a periodic heat equation and a one-dimensional reaction-diffusion equation. MintFlow had the lowest reported point estimates for distribution distance and simulation error on both systems. Projection-based competitors achieved smaller explicit constraint residuals, however, so the authors do not claim uniform superiority. They describe MintFlow as occupying a different compromise point: much stronger constraint satisfaction than the unconstrained and several guided baselines, without the large distributional and dynamical errors seen in the strongest projection methods.

There are important boundaries around the result. The core correction relies on a first-order linearization and assumes differentiable constraints; the authors say highly nonlinear cases may require repeated corrections or higher-order approximations. Their ablation on the heat dataset also showed noticeable degradation below 100 Euler sampling steps, and changing the intervention-time regularization shifted the balance between constraint satisfaction and distribution preservation.

If the findings hold across larger models and harder tasks, the broader appeal is modularity. MintFlow leaves the pretrained vector field intact and requires no constraint-specific retraining, making it potentially useful when the rule changes from one request to the next. The authors identify non-differentiable constraints, lower computational overhead and extension to other dynamics-based generators as open directions rather than solved capabilities.