⚡ ZEUS: Zero-shot Efficient Unified Sparsity

Yixiao Wang*, Ting Jiang*, Zishan Shao*, Hancheng Ye, Jingwei Sun, Mingyuan Ma, Jianyi Zhang, Yiran Chen, Hai Li

Project Page

Duke University, * Equal contribution.

What does it minimally take to accelerate generative models—without training?

*Fig. 1. Overview of ZEUS pipeline.*

⚙️ Environment

⚡ZEUS can be directly adapted into any 🤗Huggingface Diffuser workflows. Start a new environment with:

conda create -n zeus python=3.10
conda activate zeus
pip install -r requirements.txt

🚀 Quickstarts

We provide the following demos to test ZEUS. Simply run:

python sd_demo.py 

python xl_demo.py 

python flux_demo.py 

python wan2_demo.py 

python cogvideo_demo.py 

---

with --solver {dpm|euler}， --prompt, and --seed

🧩 Zeus Patch: TL;DR

from zeus import patch

patch.apply_patch(pipe,
  acc_range=(10, 45), # when to apply ZEUS
  interp_mode="psi",
  caching_mode="reuse_interp", # default: ZEUS pattrn

  denominator=3, # sparsity ratio
  modular=(0,1, ),

  lagrange_int=4, 
  lagrange_step=24,
  lagrange_term=4
)

📕 Citation

If you find this work useful, please cite our paper:

@misc{zeus2025,
  title        = {ZEUS: Zero-shot Efficient Unified Sparsity for Generative Models},
  author       = {Yixiao Wang and Ting Jiang and Zishan Shao and Hancheng Ye and Jingwei Sun and Mingyuan Ma and Jianyi Zhang and Yiran Chen and Hai Li},
  year         = {2025},
  howpublished = {https://yixiao-wang-stats.github.io/zeus/},
  note         = {Code and project page available at {https://github.com/Ting-Justin-Jiang/ZEUS}}
}

---

🍾 Acknowledgement

ZEUS codebase is build upon the excellent work of SADA, Huggingface Diffuser and ToMeSD