RESEARCH · RESEARCH · #1522
MintFlow: minimal-intervention constrained sampling for flow-matching models (arXiv:2610.02260v1)
This paper introduces MintFlow, a training-free constrained-sampling framework for flow-matching generative models that computes a closed-form minimal perturbation to an intermediate flow state (via an adjoint formulation) so the pretrained flow subsequently satisfies target constraints; it also adaptively chooses the intervention time and reports better preservation of the pretrained distribution while achieving competitive constraint satisfaction on vision and physical-modeling tasks.
KEY POINTS
- This paper introduces MintFlow, a training-free constrained-sampling framework for flow-matching generative models that computes a closed-form minimal perturbation to an intermediate flow state (via an adjoint formulation) so the pretrained flow subsequently satisfies target constraints; it also adaptively chooses the intervention time and reports better preservation of the pretrained distribution while achieving competitive constraint satisfaction on vision and physical-modeling tasks.
- MintFlow lets users enforce measurement- or physics-style constraints without retraining and with minimal deviation from the pretrained generative distribution, improving practical applicability of flow-based generative models.
- MintFlow: Minimal Trajectory Intervention for Constrained Flow Matching
WHY IT MATTERS
MintFlow lets users enforce measurement- or physics-style constraints without retraining and with minimal deviation from the pretrained generative distribution, improving practical applicability of flow-based generative models.