Velocity-Forecast Sampling for Flow-Matching Heads: A Negative Result

Abstract

Autoregressive image generation with flow-matching heads achieves high quality but suffers from slow inference due to repeated network evaluations per token. We propose Velocity-Forecast Sampling (VFS), a training-free method that speculatively reuses velocity predictions across ODE steps, verifying via MSE-based drift detection. Our systematic evaluation on NextStep-1.1 with GenEval reveals a negative result: VFS is Pareto-dominated by simple step reduction. The 10-step ODE baseline achieves both higher quality (GenEval 0.580 vs.\ 0.527--0.567) and competitive speedup (1.68$\times$ vs.\ 1.18--1.76$\times$) compared to all VFS configurations. Our analysis identifies two fundamental limitations: classifier-free guidance disrupts velocity smoothness (12--21% per-step acceptance), and the FM head accounts for only $\sim$31% of latency, limiting acceleration potential. This negative result suggests future efficiency work should target the transformer backbone rather than the flow-matching head.