Sampling from flow matching models relies on numerical ODE or SDE integration, so the cost of inference scales with the number of function evaluations (NFEs). Choosing the right budget is a practical challenge: too few evaluations degrades sample quality, while too many wastes compute. A new paper on arXiv introduces FlowBudget, a training-free approach to estimating the sampling budget for these models.
FlowBudget is designed for continuous flow-based generators, which include flow matching models. The method aims to predict how many function evaluations are needed for a given sampling task without requiring additional training. This could help practitioners set inference budgets more efficiently, avoiding both under- and over-sampling.
The abstract is truncated in the source, so details on the estimation mechanism and experimental results are not available here. Based on the announced scope, the contribution is a practical tool for reducing inference cost in flow-based generative models. The paper is listed as arXiv:2610.20636v1.