AnchorCache: Empirical Span-Error Composition for Anchor-Schedule Synthesis in Video Rectified-Flow Inference

Cache-based acceleration reduces the cost of video diffusion and rectified-flow inference by reusing model outputs across denoising steps. Under a fixed budget of true model evaluations, a central question is where those evaluations should be placed. We introduce AnchorCache, a training-free method that models the final-output error caused by zero-order velocity-reuse spans. Controlled probes reveal three recurring empirical patterns in our setting: span error grows sublinearly with reuse length, its entry cost is larger early in the trajectory, and errors from separated spans are often weakly aligned. These observations motivate a squared span-cost surrogate that decomposes over reuse spans and can be minimized by a one-dimensional dynamic program. After at most three calibration-only local moves, the resulting anchor schedule is frozen and used for every prompt without runtime gating. On Wan2.1-T2V-1.3B with 12 model evaluations in a 30-step sampler, AnchorCache obtains a mean video rel-L2 of 0.3266 on an independent 100-prompt split, compared with 0.3842 for BudCache, 0.4005 for MagCache, and 0.4167 for TeaCache. A controlled latency audit measures a 2.44× speedup over dense 30-step inference. The squared-cost model is an empirical, topology-dependent approximation rather than a deterministic error bound.