SoftCache: Adaptive Step Size Caching for Diffusion Transformers (arXiv 2610.22369)

Forum topic · 小凯 · 2026-10-17

Summary

SoftCache is a training-free acceleration method for diffusion and flow models, proposed by Shimon Malnick, Nadav Orenstein, Ohad Fried, and Shai Avidan (arXiv:2610.22369). Existing dynamic caching approaches reuse recently computed activations to skip expensive forward passes, but treat the staleness signal as a binary switch: either refresh the network or reuse stale activations at the prescribed step size. SoftCache instead uses the signal's headroom—how far it sits below the refresh threshold—to predict the error induced by reuse, and lets this headroom determine how far each cached step can travel. When headroom is large, the sampler skips multiple scheduled nodes; as refresh approaches, step sizes gradually shrink back to the original schedule, allowing adaptive jumps that fall between scheduled nodes. SoftCache requires no training and no per-backbone setup. On FLUX image generation and HunyuanVideo and Wan2.1 video generation, it outperforms all baselines across all four fidelity metrics while achieving speedups of 3.5x to 6.9x.

SoftCache: Adaptive Step Size Caching for Diffusion Transformers

Field: Computer Vision Authors: Shimon Malnick, Nadav Orenstein, Ohad Fried, Shai Avidan Published: 2026-10-15 arXiv: 2610.22369

Paper Overview

Diffusion and flow models generate samples over iterative steps, each step being a full forward pass through a large network, which dominates generation cost. Training-free caching accelerates generation by reusing recently evaluated activations. Dynamic caching methods decide when to refresh by accumulating a cheap probe of the network input into a staleness signal, and perform hard caching: the signal acts as a binary switch between refresh and reuse.

The authors observe that the distance below the threshold—the headroom—predicts the error induced by reuse. They propose SoftCache, which performs *soft caching*: the headroom determines how far each cached step travels. With large headroom, the sampler skips multiple scheduled nodes; as refresh approaches, step sizes gradually shrink back to the original step size. These adaptive jumps may fall between scheduled nodes, allowing the sampler to deviate from the fixed grid.

SoftCache is training-free and adds no per-backbone setup. On FLUX image generation and HunyuanVideo / Wan2.1 video generation, it outperforms all baselines on all four fidelity metrics at speedups of 3.5x to 6.9x.

Original Abstract (excerpt)

> Diffusion and flow models generate a sample over iterative steps, each a full pass through a large network, and that pass dominates generation cost. Training-free caching accelerates them by reusing activations from recent evaluations instead of recomputing them. Dynamic caching methods decide when to refresh by accumulating a cheap probe of the network input into a staleness signal and spending a fresh evaluation once it crosses a threshold. They perform hard caching: the signal is a binary switch between refresh and reuse, and every step, fresh or cached, advances by the prescribed step size. We observe that the signal carries much more than this switch. How far it sits below the threshold, its headroom, predicts the error induced by reuse. We propose SoftCache, which performs soft cachi...

--- *Auto-collected on 2026-10-17*