🚀 NaviCache: Test-Time Self-Calibration Caching for Video Generation (ICML 2026)

This is the official repository for NaviCache: Test-Time Self-Calibration Caching for Video Generation, accepted to ICML 2026.

NaviCache is a training-free, offline calibration-free, test-time self-calibration caching method for accelerating video diffusion models. It calibrates itself during inference, tracks feature evolution with a lightweight state-space estimator, and adaptively decides whether to skip or compute DiT blocks.

✨ Highlights

• No offline calibration: no calibration datasets, pre-processing, or per-model fitting.
• Test-time self-calibration: the cache updates its feature-change estimator during inference.
• Plug-and-play acceleration: lightweight integration for Wan2.1, HunyuanVideo, and Open-Sora.
• Adaptive computation allocation: skip/update decisions are controlled by an uncertainty-aware gate.
• Strong speed-quality trade-off: multiple presets are provided for fast, mid, and slow modes.

🔔 News

• [May 2026] NaviCache was accepted to ICML 2026.
• [June 2026] arXiv preprint is available: arXiv:2606.26795.
• [Coming Soon] Project page and demo videos.

📖 Citation

If you find NaviCache useful, please consider citing:

@inproceedings{lv2026navicache,
  author    = {Zheqi Lv and Zhibo Zhu and Jinke Wang and Qi Tian and Shengyu Zhang and Zhengyu Chen and Chengxi Zang and Zhou Zhao and Fei Wu},
  title     = {NaviCache: Test-Time Self-Calibration Caching for Video Generation},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  publisher = {PMLR}
}

🖼️ Figures

The following paper figures are included as PDF files under assets/. No PNG previews are required.

Figure	Description	Link
Figure 1	Prediction accuracy comparison and offline calibration-free/test-time self-calibration characteristics	PDF
Figure 2	Overview of the NaviCache framework	PDF
Figure 3	Video generation case study	PDF
Figure 4	Skip frequency and compute ratio across timesteps	PDF

🎬 Video Case Study

We provide the generated videos from the Wan2.1 case study for direct comparison.

Method	Video
Wan2.1	wan.mp4
TeaCache	wan_teacache.mp4
MagCache	wan_magcache.mp4
EasyCache	wan_easycache.mp4
NaviCache	wan_navicache.mp4

For easier preview and quick visual comparison, we also provide clickable GIF previews below. Click each GIF to open the corresponding full video.

Wan2.1	TeaCache	MagCache	EasyCache	NaviCache

🧩 Supported Models

Model	Task	NaviCache entry point	Example scripts
Wan2.1	Text-to-Video / Image-to-Video	NaviCache4Wan2.1/navicache_generate.py	scripts/wan/
HunyuanVideo	Text-to-Video	NaviCache4HunyuanVideo/navicache_sample_video.py	scripts/hunyuan/
Open-Sora 1.2	Text-to-Video / Evaluation	NaviCache4OpenSora/eval/navicache/experiments/opensora.py	scripts/opensora/

⚙️ Installation

git clone https://github.com/HelloZicky/NaviCache.git
cd NaviCache
pip install -r requirements.txt

For Wan2.1 and HunyuanVideo, please first install the corresponding official repository and download the required model weights. The helper scripts below copy the NaviCache entry script into the official repository automatically when executed from the official repository directory.

🚀 Quick Start with Scripts

The scripts/ directory contains three top-level launchers. Each top-level launcher calls model-specific run scripts in its corresponding subdirectory.

scripts/
├── navicache_wan.sh
├── navicache_hunyuan.sh
├── navicache_opensora.sh
├── wan/
│   ├── run_wan_t2v_1.3b.sh
│   └── run_wan_i2v_480p.sh
├── hunyuan/
│   ├── run_hunyuan_544p.sh
│   └── run_hunyuan_720p.sh
└── opensora/
    ├── run_opensora_generate.sh
    ├── run_opensora_vbench.sh
    └── run_opensora_common_metrics.sh

Wan2.1

Run from the official Wan2.1 repository directory:

git clone https://github.com/Wan-Video/Wan2.1.git
git clone https://github.com/HelloZicky/NaviCache.git

cd Wan2.1
bash ../NaviCache/scripts/navicache_wan.sh

To run a single Wan example instead of the top-level launcher:

cd Wan2.1
bash ../NaviCache/scripts/wan/run_wan_t2v_1.3b.sh
bash ../NaviCache/scripts/wan/run_wan_i2v_480p.sh

HunyuanVideo

Run from the official HunyuanVideo repository directory:

git clone https://github.com/Tencent-Hunyuan/HunyuanVideo.git
git clone https://github.com/HelloZicky/NaviCache.git

cd HunyuanVideo
bash ../NaviCache/scripts/navicache_hunyuan.sh

To run a single HunyuanVideo example instead of the top-level launcher:

cd HunyuanVideo
bash ../NaviCache/scripts/hunyuan/run_hunyuan_544p.sh
bash ../NaviCache/scripts/hunyuan/run_hunyuan_720p.sh

Open-Sora Evaluation

Run from the NaviCache repository root:

cd NaviCache
bash scripts/navicache_opensora.sh

To run one Open-Sora stage at a time:

cd NaviCache
bash scripts/opensora/run_opensora_generate.sh
bash scripts/opensora/run_opensora_vbench.sh
bash scripts/opensora/run_opensora_common_metrics.sh

🔧 Recommended Hyperparameters

Model	Mode	NAVICACHE_THRESH	NAVICACHE_ALIGN_STEPS
Wan2.1	fast	0.07	10
Wan2.1	mid	0.05	10
Wan2.1	slow	0.04	10
HunyuanVideo	fast	0.040	5
HunyuanVideo	mid	0.035	5
HunyuanVideo	slow	0.025	5
Open-Sora 1.2	fast	0.55	5
Open-Sora 1.2	mid	0.35	5
Open-Sora 1.2	slow	0.15	5

All shell scripts expose common options through environment variables. For example:

cd Wan2.1
NAVICACHE_THRESH=0.07 NAVICACHE_ALIGN_STEPS=10 bash ../NaviCache/scripts/wan/run_wan_t2v_1.3b.sh

cd HunyuanVideo
PROMPT="A cinematic shot of a corgi running through a snowy forest." bash ../NaviCache/scripts/hunyuan/run_hunyuan_544p.sh

📊 Results

NaviCache provides a strong speed-quality trade-off across multiple video diffusion backbones.

Model	Setting	Latency	Speedup	Notes
Wan2.1-1.3B	NaviCache-fast	96.40 s	2.23×	Fastest Wan2.1 setting
Wan2.1-1.3B	NaviCache-mid	106.97 s	2.01×	Balanced speed and fidelity
Wan2.1-1.3B	NaviCache-slow	115.86 s	1.86×	Highest visual retention among Wan presets
HunyuanVideo	NaviCache-fast	928.45 s	2.55×	Strong acceleration on long videos
HunyuanVideo	NaviCache-mid	1089.43 s	2.17×	Better PSNR/LPIPS than EasyCache at similar latency
Open-Sora 1.2	NaviCache-mid	35.29 s	1.60×	Balanced Open-Sora setting

Please refer to the paper for the full comparison with PAB, TeaCache, MagCache, and EasyCache.

🧠 Method Overview

NaviCache reformulates feature caching in video diffusion models as a test-time state estimation problem.

1. Initial Alignment: run several full-computation steps to initialize the feature-change ratio and uncertainty.
2. Test-Time Self-Calibration: track the relationship between input feature changes and output feature changes during inference.
3. Uncertainty-Aware Skipping: skip computation when the accumulated predicted error is below a fidelity threshold, and perform a full update when the threshold is exceeded.

📁 Repository Structure

NaviCache/
├── NaviCache4HunyuanVideo/        # NaviCache script for HunyuanVideo
├── NaviCache4OpenSora/            # Open-Sora NaviCache evaluation and VideoSys modules
├── NaviCache4Wan2.1/              # NaviCache script for Wan2.1
├── assets/                        # Paper figure PDFs
├── scripts/                       # Runnable helper scripts
├── requirements.txt
├── LICENSE
└── README.md

📖 Citation

If you find NaviCache useful, please consider citing:

@inproceedings{lv2026navicache,
  author    = {Zheqi Lv and Zhibo Zhu and Jinke Wang and Qi Tian and Shengyu Zhang and Zhengyu Chen and Chengxi Zang and Zhou Zhao and Fei Wu},
  title     = {NaviCache: Test-Time Self-Calibration Caching for Video Generation},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026},
  publisher = {PMLR}
}

🙏 Acknowledgements

We thank the contributors of Wan2.1, HunyuanVideo, Open-Sora, VideoSys, TeaCache, EasyCache, MagCache, and PAB for their excellent open-source work and inspiring research.

📄 License

This project is released under the Apache License 2.0.

⚠️ Responsible Use

NaviCache accelerates video generation and may lower the cost of producing synthetic media. We encourage responsible usage and support the development of detection, watermarking, and provenance-tracking tools.