LongVie2

LongVie 2: Multimodal Controllable Ultra-Long Video World Model

LongVie 2 is a multimodal controllable world model for generating ultra-long videos with depth and pointmap control signals, as presented in the paper LongVie 2: Multimodal Controllable Ultra-Long Video World Model. It is an end-to-end autoregressive framework trained to enhance controllability, long-term visual quality, and temporal consistency.

• 📝 Paper on Hugging Face
• 🌐 Project Page
• 💻 GitHub Repository

🚀 Quick Start

Installation

To get started with LongVie 2, follow the installation steps from the GitHub repository:

conda create -n longvie python=3.10 -y
conda activate longvie
conda install psutil
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
python -m pip install ninja
python -m pip install git+https://github.com/Dao-AILab/flash-attention.git@v2.7.2.post1
cd LongVie
pip install -e .

Download Weights

1. Download the base model Wan2.1-I2V-14B-480P:

python download_wan2.1.py

1. Download the LongVie2 weights and place them in ./model/LongVie/

Inference

Generate a 5s video clip (~8-9 mins on a single A100 GPU) using the following command:

bash sample_longvideo.sh

📄 Citation

If you find this work useful, please consider citing:

@misc{gao2025longvie2,
  title={LongVie 2: Multimodal Controllable Ultra-Long Video World Model}, 
  author={Jianxiong Gao and Zhaoxi Chen and Xian Liu and Junhao Zhuang and Chengming Xu and Jianfeng Feng and Yu Qiao and Yanwei Fu and Chenyang Si and Ziwei Liu},
  year={2025},
  eprint={2512.13604},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2512.13604}, 
}