Diffusion caching stops trusting raw probe drift: dividing by trajectory movement cuts LPIPS at the same compute

For training-free acceleration of diffusion transformers, the decisive variable shifts from how skipped features are reconstructed to which denoising steps receive the limited exact computation. Existing adaptive schedulers ranked steps by the raw variation of a shallow probe output, treating a large change as evidence that the step needs a full forward pass. With the reconstructor, probe depth, warm-up policy and total number of exact denoiser evaluations all fixed, different reuse placements still produced markedly different generation quality; at matched speedups CacheScope reached the lowest median LPIPS on Wan2.1, FLUX and Open-Sora 1.2.

Whether a step deserves exact computation depends on how strongly the network responds along the current trajectory direction, not on the absolute size of the output change. Raw probe variation entangles trajectory displacement with network response strength, so it is not directly comparable across denoising steps. Dividing probe-output variation by the corresponding input displacement yields a directional gain per unit trajectory movement; on Wan2.1 and FLUX this normalized signal lowered median LPIPS relative to the default DiCache signal and a difference-based signal, while the three schedulers executed nearly identical numbers of exact forwards.

The rule accumulates sensitivity online and falls back to a full denoising evaluation once the accumulated value crosses a threshold, then resets the accumulator. It leaves the underlying cache reconstructor unchanged and needs no training, gradients, auxiliary network or offline prompt-specific schedule search. The threshold is calibrated once on a 16-prompt development set to match the target exact-compute budget and then frozen for evaluation on a disjoint 32-prompt test set; the added cost is one scalar division and accumulation, with small measured latency differences.

The scheduling rule targets existing probe-based training-free caching pipelines and is evaluated on FLUX.1-dev, Wan2.1-T2V-1.3B and Open-Sora 1.2 under their respective generation protocols, with thresholds calibrated on a development set and then frozen. It only decides where additional exact evaluations are placed; it does not change the cache reconstructor or the cached history. Sensitivity is currently estimated from a single shallow probe along the realized denoising direction, and multi-probe and higher-order trajectory signals are listed as future work.

On Wan2.1, SSIM shows isolated trade-offs, so perceptual fidelity (LPIPS) is treated as the primary outcome rather than claiming uniform dominance on every metric. The offline aggregation analysis indicates that at this FLUX operating point much of the useful scheduling structure is shared across prompts, while prompt-specific online variation provides no statistically detectable additional benefit, a result specific to the evaluated backbone and operating point. The ordering of the probe-normalized signal at high skip ratios is backbone-dependent, and no universal ordering is claimed.