Position reversal without external classifier-free guidance

Establish whether the observed reversal in the effect of longer prediction horizons across query positions persists under unguided or class-conditional sampling without external classifier-free-guidance extrapolation.

Background

The principal position scan uses classifier-free guidance scale 2 and finds that longer horizons worsen FID near the source while improving FID at the midpoint. The appendix reports class-conditional CFG-1 results in which midpoint effects are small and mixed, so the paper does not establish the same position reversal without external guidance.