Relative contributions of prediction error and CFG construction bias

Determine the relative roles of approximation error and classifier-free-guidance construction bias in producing the observed endpoint effects of varying the prediction horizon in pretrained few-step MeanFlow.

Background

The paper decomposes the discrepancy between the learned interval-average velocity and the velocity required for the physical update into horizon mismatch and interval-dependent prediction error. With external classifier-free guidance, the effective predictor can also contain guidance-construction bias. The experiments demonstrate that changing the prediction horizon affects endpoint FID, but they do not separate these possible mechanisms.