Less is Enough: Training-Free Video Diffusion Acceleration via Runtime-Adaptive Caching

Xin Zhou^1*, Dingkang Liang^1*, Kaijin Chen¹, Tianrui Feng¹, Xiwu Chen², Hongkai Lin¹,
Yikang Ding², Feiyang Tan², Hengshuang Zhao³, Xiang Bai^1†

¹ Huazhong University of Science and Technology, ² MEGVII Technology, ³ The University of Hong Kong

(*) Equal contribution. (†) Corresponding author.

🎬 Visual Comparisons

Video synchronization issues may occur due to network load, for improved visualization, see the project page

Prompt: "Grassland at dusk, wild horses galloping, golden light flickering across manes."(HunyuanVideo)

Baseline	Ours (2.28x)	TeaCache (1.68x)	PAB (1.19x)

Prompt: "A top-down view of a barista creating latte art, skillfully pouring milk to form the letters 'TPAMI' on coffee."(Wan2.1-14B)

Baseline	Ours (2.63x)	TeaCache (1.46x)	PAB (2.10x)

Compatibility with SVG

SVG with EasyCache on HunyuanVideo can achieve more than 3x speedup.

SVG.with.EasyCache.mp4

📰 News

• If you like our project, please give us a star ⭐ on GitHub for the latest update.
• [2025/08/01] 🔥 EasyCache for Wan2.2 I2V is released.
• [2025/07/31] 🔥 EasyCache for Wan2.2 T2V-A14B is released.
• [2025/07/30] 🔥 EasyCache for Wan2.2 TI2V-5B is released. Thanks @AChowdhury1211.
• [2025/07/06] 🔥 EasyCache for Wan2.1 I2V is released.
• [2025/07/05] 🔥 EasyCache for Wan2.1 T2V is released.
• [2025/07/04] 🎉 Release the paper of EasyCache.
• [2025/07/03] 🔥 EasyCache for Sparse-VideoGen on HunyuanVideo is released.
• [2025/07/02] 🔥 EasyCache for HunyuanVideo is released.

Abstract

Video generation models have demonstrated remarkable performance, yet their broader adoption remains constrained by slow inference speeds and substantial computational costs, primarily due to the iterative nature of the denoising process. Addressing this bottleneck is essential for democratizing advanced video synthesis technologies and enabling their integration into real-world applications. This work proposes EasyCache, a training-free acceleration framework for video diffusion models. EasyCache introduces a lightweight, runtime-adaptive caching mechanism that dynamically reuses previously computed transformation vectors, avoiding redundant computations during inference. Unlike prior approaches, EasyCache requires no offline profiling, pre-computation, or extensive parameter tuning. We conduct comprehensive studies on various large-scale video generation models, including OpenSora, Wan2.1, and HunyuanVideo. Our method achieves leading acceleration performance, reducing inference time by up to 2.1-3.3× compared to the original baselines while maintaining high visual fidelity with a significant up to 36% PSNR improvement compared to the previous SOTA method. This improvement makes our EasyCache a efficient and highly accessible solution for high-quality video generation in both research and practical applications.

🚀 Main Performance

We validated the performance of EasyCache on leading video generation models and compared it with other state-of-the-art training-free acceleration methods.

Comparison with SOTA Methods

Tested on Vbench prompts with NVIDIA A800.

Performance on HunyuanVideo:

Method	Latency (s)↓	Speedup ↑	PSNR ↑	SSIM ↑	LPIPS ↓
HunyuanVideo (Baseline)	1124.30	1.00x	-	-	-
PAB	958.23	1.17x	18.58	0.7023	0.3827
TeaCache	674.04	1.67x	23.85	0.8185	0.1730
SVG	802.70	1.40x	26.57	0.8596	0.1368
EasyCache (Ours)	507.97	2.21x	32.66	0.9313	0.0533

Performance on Wan2.1-1.3B:

Method	Latency (s)↓	Speedup ↑	PSNR ↑	SSIM ↑	LPIPS ↓
Wan2.1 (Baseline)	175.35	1.00x	-	-	-
PAB	102.03	1.72x	18.84	0.6484	0.3010
TeaCache	87.77	2.00x	22.57	0.8057	0.1277
EasyCache (Ours)	69.11	2.54x	25.24	0.8337	0.0952

Compatibility with Other Acceleration Techniques

EasyCache is orthogonal to other acceleration techniques, such as the efficient attention mechanism SVG, and can be combined with them for even greater performance gains.

Combined Performance on HunyuanVideo (720p):Tested on NVIDIA H20 GPUs.

Method	Latency (s)↓	Speedup ↑	PSNR (dB) ↑
Baseline	6594s	1.00x	-
SVG	3474s	1.90x	27.56
SVG (w/ TeaCache)	2071s	3.18x	22.65
SVG (w/ Ours)	1981s	3.33x	27.26

🛠️ Usage

Detailed instructions for each supported model are provided in their respective directories. We are continuously working to extend support to more models.

🎯 To Do

• Support HunyuanVideo
• Support Sparse-VideoGen on HunyuanVideo
• Support Wan2.1 T2V-1.3B/14B
• Support Wan2.1 I2V-14B
• Support Wan2.2 TI2V-5B
• Support Wan2.2 T2V-A14B
• Support Wan2.2 I2V

🌹 Acknowledgements

We would like to thank the contributors to the Wan2.1, HunyuanVideo, Wan2.2, OpenSora, and SVG repositories, for their open research and exploration.

• ComfyUI-WanVideoWrapper supports EasyCache4Wan2.1/EasyCache4Wan2.2. Thanks @kijai.
• ComfyUI supports EasyCache for several model includes Kontext, Qwen Image Edit, Wan2.2 14B. Thanks @Kosinkadink.

📖 Citation

If you find this repository useful in your research, please consider giving a star ⭐ and a 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}
}