MintFlow
MintFlow is a training-free constrained sampling framework built on flow-matching generative models. It achieves constraint satisfaction by applying minimal interventions to pretrained flow trajectories, derives closed-form perturbation expressions using the adjoint formula, and adaptively selects intervention timing to balance constraint satisfaction against distributional fidelity.
Timeline (last 90 days)
MintFlow is a training-free constrained sampling framework that reformulates the constraint imposition process as minimal intervention on pre-trained flow trajectories.
MintFlow does not modify the flow field of the pretrained model; instead, it seeks the minimal perturbation to some intermediate flow state such that the perturbed state, when evolved under the original flow field, satisfies the target constraint.
MintFlow derives the required perturbation using an adjoint formula, obtaining a closed-form expression for that perturbation.
MintFlow's closed-form solution allows the perturbation to be computed directly, avoiding expensive iterative optimization and substantially reducing sampling cost.
MintFlow adaptively selects the intervention time to balance the magnitude of the required perturbation against the degree to which it is amplified by the remaining flow evolution process.
The paper evaluates MintFlow on tasks including generative vision and physical system modeling, and the results show that it achieves competitive constraint satisfaction while significantly better preserving the pretrained generative distribution compared to current state-of-the-art constraint methods.
MintFlow's experimental results indicate that it achieves a better Pareto frontier between constraint satisfaction and generation quality.
MintFlow is a newly released arXiv preprint paper.
All Facts (8)
MintFlow is a training-free constrained sampling framework that reformulates the constraint imposition process as minimal intervention on pre-trained flow trajectories.
50%UnverifiedMintFlow does not modify the flow field of the pretrained model; instead, it seeks the minimal perturbation to some intermediate flow state such that the perturbed state, when evolved under the original flow field, satisfies the target constraint.
50%UnverifiedMintFlow derives the required perturbation using an adjoint formula, obtaining a closed-form expression for that perturbation.
50%UnverifiedMintFlow's closed-form solution allows the perturbation to be computed directly, avoiding expensive iterative optimization and substantially reducing sampling cost.
50%UnverifiedMintFlow adaptively selects the intervention time to balance the magnitude of the required perturbation against the degree to which it is amplified by the remaining flow evolution process.
50%UnverifiedThe paper evaluates MintFlow on tasks including generative vision and physical system modeling, and the results show that it achieves competitive constraint satisfaction while significantly better preserving the pretrained generative distribution compared to current state-of-the-art constraint methods.
50%UnverifiedMintFlow's experimental results indicate that it achieves a better Pareto frontier between constraint satisfaction and generation quality.
50%UnverifiedMintFlow is a newly released arXiv preprint paper.
50%