ConicFlow: Constrained Inference-Time Optimal Control for Flow Matching
October 14, 2026
ConicFlow frames constrained sampling in flow matching models as an optimal control problem using a receding-horizon primal-dual algorithm. This approach allows for inference-time steering toward feasible samples that satisfy hard terminal constraints while minimizing deviation from pretrained trajectories.
HOW THIS AFFECTS YOU
●
researcherYou can implement precise, constraint-satisfying sampling in generative flow models using discretized optimal control.