# CacheScope speeds up diffusion models without extra training

> The new CacheScope method cuts inference time up to 3.5× while lowering image quality loss, and it works on FLUX.1-dev, Wan2.1-T2V-1.3B and Open‑Sora 1.2.

Oossa · 2026-10-12 · https://oossa.com/en/cachescope-speeds-up-diffusion-models-without-extra-training

Researchers led by Qinhan Lyu introduced CacheScope, a technique that decides when to use full denoising computation and when to reuse cached approximations during diffusion model inference. It needs no extra training, gradients, or separate networks. On three popular diffusion models, it speeds up generation by up to 3.57 times and cuts the median LPIPS error by up to 70.1% compared with the strongest existing probe‑based method.

## The facts

- Paper posted on arXiv on 2026-10-12
- Up to 3.57× speedup and 70.1% LPIPS reduction on tested models

## Why it matters

Developers can generate images faster and with higher fidelity without spending resources on additional training.

## Sources & references

1. [CacheScope: Exact-Compute Allocation for Training-Free Diffusion Acceleration](https://arxiv.org/abs/2610.12530) – arXiv cs.LG, 2026-10-12

Last updated: 2026-10-12