MintFlow: Minimal Trajectory Intervention for Constrained Flow Matching

Flow matching models have become powerful tools for generative modeling, but they often struggle with constraints. Many real-world applications require samples to satisfy specific conditions—whether physical laws or observed measurements—while maintaining the quality of the pretrained distribution. The challenge lies in balancing constraint enforcement with preserving the original data distribution. Existing constrained samplers tend to push samples too far from the pretrained distribution when applying constraints, leading to poor quality outputs. MintFlow addresses this fundamental trade-off with a novel approach that requires no additional training.
The core innovation in MintFlow lies in its formulation of constraint enforcement as a minimal intervention on the pretrained flow trajectory. Instead of modifying the entire model or retraining, MintFlow identifies the smallest necessary perturbation to an intermediate flow state that ensures the final output satisfies the target constraint. This is achieved through an adjoint formulation that provides a closed-form expression for the required perturbation, eliminating the need for expensive iterative optimization. The method maintains the pretrained flow field unchanged while only slightly adjusting the flow state at strategically chosen intervention points.
One of the most elegant aspects of MintFlow is its adaptive intervention time selection. The framework determines the optimal point in the flow trajectory to apply the minimal perturbation, balancing two competing factors: the magnitude of the required perturbation and its amplification as the flow continues to evolve. This dynamic adjustment ensures that constraints are satisfied with minimal disruption to the pretrained distribution. The mathematical foundation behind this approach is rooted in optimal control theory, specifically leveraging adjoint methods to compute necessary conditions efficiently.
The practical implications of MintFlow extend to multiple domains. In generative vision tasks, the method can enforce visual constraints while maintaining image quality and diversity. For physical system modeling, it can ensure outputs respect physical laws without compromising the generative capabilities of pretrained models. The training-free nature of the approach makes it particularly valuable for applications where retraining is impractical or computationally prohibitive.
The evaluation results demonstrate MintFlow's competitive performance against state-of-the-art constrained methods. Across various benchmarks, the framework achieves comparable constraint satisfaction while significantly preserving the pretrained generative distribution. This represents a meaningful advancement in constrained sampling, addressing a long-standing challenge in the field. The closed-form solution provided by the adjoint formulation is particularly noteworthy, as it enables efficient computation without compromising solution quality.
From an implementation perspective, MintFlow can be integrated into existing flow matching frameworks with minimal modifications. The method requires only the addition of the minimal intervention step and the adjoint computation, making it accessible for researchers and practitioners working with pretrained models. This accessibility is crucial for widespread adoption, as it doesn't require significant architectural changes or computational overhead.
The theoretical contributions of MintFlow extend beyond practical applications. By framing constraint enforcement as a minimal trajectory intervention, the work provides new insights into the relationship between flow matching and optimal control. The adjoint formulation establishes a principled approach to constraint satisfaction in generative models, potentially inspiring new directions in both theoretical and applied research.
As generative models continue to advance, the ability to enforce constraints while maintaining distributional quality will become increasingly important. MintFlow represents a significant step forward in this direction, offering a mathematically grounded, computationally efficient solution to a fundamental challenge in the field. The framework's ability to preserve pretrained distributions while satisfying constraints opens new possibilities for applications in science, engineering, and creative domains where both fidelity and compliance are essential.