Entropy-Based Scheduler Boosts Flow Model Sampling Efficiency

Researchers have developed a training-free entropic inference-time scheduler for flow-based generative models, improving sample quality under a small inference budget. The method uses a conditional-marginal entropy-rate objective to discretize the probability path, separating bridge geometry from marginal flow evolution. For Gaussian Brownian bridges, the entropy rate is closed-form and U-shaped, motivating nonuniform grids that concentrate evaluations near boundaries. On trained 2D bridge/flow models, the scheduler improved 10-step ODE-Heun MMD by 18.1% and SDE-Heun by 22.7% over linear discretization. On EDM/CIFAR-10, the entropic time schedule also showed gains. The work is published on arXiv (2605.16126) and addresses a key challenge in flow matching and Schrödinger bridges, where inference grids are typically heuristic.