Diffusion models can skip some of their most expensive calculations by reusing recent results, but each skip risks making the generated output drift. A new method called SoftCache uses a cache’s existing measure of staleness to decide not only when to refresh, but how far the sampler can move before checking again.
Reuse the signal as a step-size guide
Image and video diffusion models gradually turn noise into a sample through repeated passes through a large neural network. A cache saves computation by reusing a recent network contribution rather than recomputing it at every step. Existing dynamic caches track how much the input has changed and refresh once that signal crosses a threshold.
SoftCache treats the distance below that refresh threshold as “headroom.” When the signal is small and the cached contribution is still likely to be useful, the sampler takes a longer step, skipping scheduled points and sometimes landing between them. As the signal approaches the threshold, its steps shorten; it refreshes when the threshold is reached. The method requires no model training and uses no network evaluations beyond the cache’s existing probe.
SoftCache more closely matched uncached outputs at similar speed
The authors tested SoftCache on FLUX.1-dev image generation, HunyuanVideo and Wan2.1 video generation, comparing it with cache methods at similar measured speed. All three models used 50 Euler sampling steps. Tests used 200 DrawBench prompts for FLUX, 946 VBench prompts for HunyuanVideo and a stratified 100-prompt subset for Wan2.1.
On HunyuanVideo at roughly 5.4 times the uncached speed, SoftCache recorded a 22.36 peak signal-to-noise ratio (PSNR) and 0.246 LPIPS, compared with 21.82 and 0.292 for SeaCache at 5.33 times, and 20.34 and 0.293 for DPCache at 5.42 times. Higher PSNR and lower LPIPS indicate closer resemblance to the uncached output generated from the same prompt and seed. The paper also reports SSIM and DreamSim, two additional similarity measures, and says SoftCache led the tested baselines on all four measures in each of three speed bands.
The benchmark does not measure user preference or deployed cost
These fidelity scores test how closely an accelerated sampler reproduces its own uncached output; they do not establish that people prefer the resulting images or videos. The experiments are author-run and reported in a version-one arXiv preprint, with no independent replication identified in the paper. The prompt sets and hardware also differ by model, so the reported speedups should not be treated as a cross-model cost comparison.
Sources
- Full paper: “SoftCache: Adaptive Step Size Caching for Diffusion Transformers,” arXiv, version 1
- Abstract and submission record, dated Oct. 15, 2026
- Illustrative image: MrAlanKoh, “Example of Forward Diffusion Models”, CC BY-SA 4.0. The image illustrates diffusion generally; it is not from the SoftCache study. No changes made.