CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration
Computer Science > Machine Learning arXiv:2610.12530 (cs) [Submitted on 8 Oct 2026] Title: CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration Authors: Qinhan Lyu, Jusheng Zhang, Zhixing Zhang, Hao Liu, Zijie Liang, Jin Jiang, Keze Wang.
计算机科学 > 机器学习 arXiv:2610.12530 (cs) [提交于 2026 年 10 月 8 日] 标题：CacheScope：用于免训练扩散加速的精确计算分配。作者：Qinhan Lyu, Jusheng Zhang, Zhixing Zhang, Hao Liu, Zijie Liang, Jin Jiang, Keze Wang。
Abstract: Training-free feature reuse accelerates diffusion transformer inference by replacing full denoising computation at selected steps with cached approximations. Existing methods mainly improve the reconstruction of skipped features, but we find that once a stable first-order reconstructor is used, increasing reconstruction complexity brings limited gains, whereas the placement of exact-computation steps has a much larger effect on fidelity.
摘要：免训练特征重用通过在选定步骤中用缓存近似值替换完整的去噪计算，从而加速扩散 Transformer 的推理。现有方法主要致力于改进跳过特征的重构，但我们发现，一旦使用稳定的一阶重构器，增加重构复杂度带来的收益有限，而精确计算步骤的放置位置对保真度的影响要大得多。
More importantly, absolute shallow-probe variation is jointly affected by input-trajectory displacement and network response strength, and therefore does not directly reflect cache risk. We address this confounding by normalizing probe-output variation with the corresponding input displacement, yielding a directional gain per unit trajectory movement.
更重要的是，绝对浅层探测（shallow-probe）的变化受到输入轨迹位移和网络响应强度的共同影响，因此不能直接反映缓存风险。我们通过使用相应的输入位移对探测输出变化进行归一化来解决这种混淆，从而得出单位轨迹移动的定向增益。
Based on this signal, we propose CacheScope, which identifies high-risk reuse steps online and selectively falls back to full denoising computation while leaving the underlying cache reconstructor unchanged. CacheScope requires no additional training, gradients, auxiliary networks, or prompt-specific offline schedule optimization.
基于这一信号，我们提出了 CacheScope，它可以在线识别高风险的重用步骤，并有选择地回退到完整的去噪计算，同时保持底层的缓存重构器不变。CacheScope 不需要额外的训练、梯度、辅助网络或针对特定提示的离线调度优化。
Across FLUX.1-dev, Wan2.1-T2V-1.3B, and Open-Sora 1.2, CacheScope achieves up to 3.57x acceleration and reduces median LPIPS by up to 70.1% over the strongest probe-based baseline.
在 FLUX.1-dev、Wan2.1-T2V-1.3B 和 Open-Sora 1.2 上，CacheScope 实现了高达 3.57 倍的加速，并将中位数 LPIPS 较最强的基于探测的基准降低了高达 70.1%。