Dynamic scheduling of the sensitivity threshold in SenCache

Design and characterize timestep-dependent schedules for the sensitivity threshold ε in Sensitivity-Aware Caching (SenCache) for diffusion-model inference, identifying effective patterns that allocate the per-step error budget across denoising timesteps to further accelerate sampling while maintaining generation quality.

Background

SenCache proposes a sensitivity-aware criterion for caching denoiser outputs in diffusion and flow-matching models. The method uses a tolerance parameter ε as a local error budget to decide when to reuse cached outputs, and the experiments in the paper employ a fixed ε across timesteps.

The authors note that ε maps directly to a per-step error budget, and different timesteps contribute unequally to the final fidelity. They suggest that dynamically scheduling ε over time may allow larger error at less critical stages to gain more speed while preserving output quality, but they do not provide such schedules or an analysis of effective patterns.

References

Additionally, as the sensitivity threshold ε maps directly to an error budget at each denoising step, dynamically scheduling ε across timesteps could further accelerate inference while maintaining generation quality: different steps contribute unequally to final fidelity, so allowing larger error at less critical stages may be acceptable. In this paper, we used a fixed threshold; designing schedules and characterizing effective patterns is left for future work.

— SenCache: Accelerating Diffusion Model Inference via Sensitivity-Aware Caching  (2602.24208 - Haghighi et al., 27 Feb 2026) in Section: Discussions and Future Work

We have not systematically characterized the distribution of golden paths or the distances between them. Whether small schedule changes can connect them at a fixed cache ratio while preserving high output quality across prompts remains an open question.

— The Golden Path Hypothesis: Reusable Schedules in Diffusion Caching  (2609.39343 - Wang et al., 30 Sep 2026) in Section 8, Discussion and Conclusion, Limitations