Harahan/QVGen-CogVideoX-2B-W4A4

QVGen:
Pushing the Limit of Quantized Video Generative Models

     

[ Conference Paper | Models | Dataset | Code ]

Yushi Huang, Ruihao Gong📧, Jing Liu, Yifu Ding, Chengtao Lv, Haotong Qin, Jun Zhang📧

(📧 denotes corresponding author.)

📖 Overview

QVGen is the first to reach full-precision comparable quality under 4-bit settings and it significantly outperforms existing methods. For instance, our 3-bit CogVideoX-2B improves Dynamic Degree by +25.28 and Scene Consistency by +8.43 on VBench.

⚙️ Usage

See our official code base.

✨ Model Zoo

Model	#Bit
Wan 1.3B	W4A4
CogVideoX-2B	W4A4

✏️ Citation

If you find QVGen useful, please cite our paper:

@inproceedings{huang2026qvgenpushinglimitquantized,
  title={QVGen: Pushing the Limit of Quantized Video Generative Models}, 
  author={Yushi Huang and Ruihao Gong and Jing Liu and Yifu Ding and Chengtao Lv and Haotong Qin and Jun Zhang},
  booktitle={International Conference on Learning Representations},
  year={2026},
  url={https://arxiv.org/abs/2505.11497}, 
}