MeanFlow is a generative model that predicts interval-average velocities, a design choice that suits few-step sampling. In conventional usage, the interval the model is asked to predict is tied directly to the duration of the state update being performed. That coupling limits flexibility: changing the update step size forces a corresponding change in the query interval.

The paper proposes separating these two quantities at inference time. By decoupling the model query interval from the state-update duration, the same fixed MeanFlow model can be used with different update schedules. This is purely an inference-time change, requiring no retraining or modification of the model weights.

The authors argue that this decoupling opens up new ways to control the sampling trajectory. Because the model remains fixed, the approach can be applied to existing MeanFlow checkpoints. The abstract does not report experimental results, so the practical gains are not quantified here, but the conceptual shift is clear: treat the query interval as an independent control knob rather than a fixed consequence of the update step.