EasyCache

Category: Sampler-level acceleration

Overview

EasyCache introduces a lightweight, runtime-adaptive caching mechanism that dynamically reuses previously computed transformation vectors, avoiding redundant computations during inference.

Parameters

Parameter	Type	Default	Range	Description
`sampler_name`	enum	euler	see supported list	Base sampler algorithm
`ec_threshold`	float	0.025	0.0 ~ 1.0	Accumulated error threshold; lower = higher quality
`ret_steps`	int	5	0 ~ 100	Number of initial steps forced to compute (warmup protection)

Parameter Tuning Guide

ec_threshold: Controls the aggressiveness of caching. 0.01~0.03 for high quality, 0.05~0.1 for more speedup. Too high may cause blurriness or artifacts.
ret_steps: Warmup period where no caching is applied (early steps have high variation). Recommended: 3~8 for 20-50 step schedules.

Official Recommended Parameters by Model

Model	ec_threshold	ret_steps	Note
HunyuanVideo	0.025	5	Official default, conservative/high-quality
Wan2.1	0.05	10	Official argparse default, more aggressive
Wan2.2	0.05	10	Same as Wan2.1

The node defaults (ec_threshold=0.025, ret_steps=5) follow the HunyuanVideo configuration. For Wan series models, consider increasing both values for better speedup.

Supported Samplers

euler, heun, heunpp2, dpm_2, lms, dpm_fast, dpm_adaptive, dpmpp_2m, ipndm, ipndm_v, deis, res_multistep, gradient_estimation, ddim, euler_ancestral, dpm_2_ancestral

Reference

Paper: Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching
GitHub: https://github.com/H-EmbodVis/EasyCache

Citation

@article{zhou2025easycache,
  title={Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching},
  author={Zhou, Xin and Liang, Dingkang and Chen, Kaijin and Feng, Tianrui and Chen, Xiwu and Lin, Hongkai and Ding, Yikang and Tan, Feiyang and Zhao, Hengshuang and Bai, Xiang},
  journal={arXiv preprint arXiv:2507.02860},
  year={2025}
}

2.3 KiB Raw Permalink Blame History