Where, Not How Much: Fixed versus Adaptive Noise Schedules for Flow-Matching SDE Samplers at Low NFE

Abstract Stochastic samplers for flow-matching image models inject noise with a strength (eta) that is usually a hand-picked constant inside a fixed window. We ask whether this strength can be set per step from free local signals available inside a multistep predictor–corrector solver, namely the trajectory curvature and the predictor–corrector error, at 6–15 network evaluations. We implement a budget-conserving allocator for SA-Solver that redistributes a fixed total noise budget over the steps, so that a gain cannot be explained by a different overall noise level, and compare it with constant-eta baselines on three flow-matching models (Anima, Flux 3 Klein, MOSS-Image) using geometry and texture measures in latent space; a fourth model (Krea 2) was tried informally and was inconclusive. Curvature-driven allocation had no stable effect on any model. Error-driven allocation that places more noise where the solver error is small gave a small gain on Flux 3 Klein (about 1–2 % of the texture value), which held on fresh seeds under criteria fixed in advance; a one-sided effect on MOSS-Image; and no detectable effect on Anima. A fixed per-step schedule calibrated on independent runs reproduced the Flux 3 gain, so online adaptation was not necessary: where the noise is injected along the trajectory matters, the local state does not. In a two-alternative human study (39 respondents), people detected a large difference in noise strength but no difference between the error-driven allocation and uniform noise. We release the code, per-image metrics and analysis scripts, and report the negative results in full.