QVGen:
Pushing the Limit of Quantized Video Generative Models

       

[ Conference Paper | Models | Dataset ]

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

(📧 denotes corresponding author.)

This is the official implementation of our paper QVGen. It 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.

🔥 News

• Jan 31, 2026: 🔥 We release our Python code and checkpoints for QVGen presented in our paper. Have a try!

• Jan 23, 2026: 🌟 Our paper has been accepted by ICLR 2026! 🎉 Cheers!

🎬 Visual Examples

Comparison of samples from 4-bit per-channel weight and per-token activation quantized CogVideoX-2B (upper) and Wan 14B (lower), across different methods.

📖 Overview

Overview pipeline of the proposed QVGen. (a) This framework integrates auxiliary modules $\Phi$ to improve training convergence. (b) To maintain performance while eliminating inference overhead induced by $\Phi$, we design a rank-decay schedule that progressively shrinks the entire $\Phi$ to $\varnothing$ through iteratively applying the following two strategies: (i) SVD to identify the low-impact components in $\Phi$; (ii) A rank-based regularization $\mathbf{\gamma}$ to decay the identified components to $\varnothing$.

✨ Quick Start

After cloning the repository, follow these steps to train and run inference. We use Wan 1.3B as the example (the same workflow applies to CogVideoX 2B).

Requirements

Install dependencies with pip install -r requirements.txt. For quantization, we use 8x H100/H800/A100/A800 GPUs; otherwise, adjust the scripts accordingly.

Prepare data and models

Download the pretrained Wan2.1 1.3B to models/Wan2.1-T2V-1.3B-Diffusers, then prepare the dataset:

# download and preprocess data
python prepare_dataset/download.py --output_directory dataset
sh script/data/prepare_dataset_wan_1-3b.sh

Note: please replace any path/to/your/... placeholders in the scripts with your local paths.

We also provide quantized checkpoints on Hugging Face:

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

Training

Below is an example training command. For more details, please refer to the paper.

sh script/train/w4a4_wan_1-3b.sh

Inference

Here is the corresponding inference command.

# You can also set the path to a pre-downloaded quantized checkpoint in the script.
sh script/inference/w4a4_wan_1-3b.sh

Evaluation

We recommend using our inference code and following the steps in VBench. For evaluation, you can also refer to our distributed inference scripts for video generation: inference/ddp_cogvideox_t2v.py and inference/ddp_wan_t2v.py.

💪 TODO

The codebase has not been fully cleaned for public release yet. We will continue to clean and refine it in subsequent updates.

• Training and inference code for large models (e.g., Wan 14B, CogVideoX 5B).
• More quantized checkpoints.
• More efficient training code.

🤝 Acknowledgments

Our code is built on the open-source finetrainers.

✏️ 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}, 
}