GenieDrive

<div align="center">

GenieDrive: Towards Physics-Aware Driving World Model with 4D Occupancy Guided Video Generation

Zhenya Yang¹, Zhe Liu^1,†, Yuxiang Lu¹, Liping Hou², Chenxuan Miao¹, Siyi Peng², Bailan Feng², Xiang Bai³, Hengshuang Zhao^1,✉

<br> ¹ The University of Hong Kong, ² Huawei Noah's Ark Lab, ³ Huazhong University of Science and Technology <br> † Project leader, ✉ Corresponding author. <br>

📑 [arXiv], ⚙️ [project page], 🤗 [model weights]

<div align="center"> <img src="assets/teaser.jpg" width="100%"> <p><em>Overview of our GenieDrive</em></p> </div> </div>

📢 News

• [2025/12/15] We release GenieDrive paper on arXiv. 🔥

• 2025.12.15: DrivePI paper released! A novel spatial-aware 4D MLLM that serves as a unified Vision-Language-Action (VLA) framework that is also compatible with vision-action (VA) models. 🔥
• 2025.11.04: Our previous work UniLION has been released. Check out the codebase for unified autonomous driving model with Linear Group RNNs. 🚀
• 2024.09.26: Our work LION has been accepted by NeurIPS 2024. Visit the codebase for Linear Group RNN for 3D Object Detection. 🚀

📋 TODO List

• [ ] Release 4D occupancy forecasting code and model weights.
• [ ] Release multi-view video generator code and weights.

📈 Results

Our method achieves a remarkable increase in 4D Occupancy forecasting performance, with a 7.2% increase in mIoU and a 4% increase in IoU. Moreover, our tri-plane VAE compresses occupancy into a latent tri-plane that is only 58% the size used in previous methods, while still maintaining superior reconstruction performance. This compact latent representation also contributes to fast inference (41 FPS) and a minimal parameter count of only 3.47M (including the VAE and prediction module).

<div align="center"> <img src="assets/table_occ.png" width="85%"> <p><em>Performance of 4D Occupancy Forecasting</em></p> </div>

We train three driving video generation models that differ only in video length: S (8 frames, ~0.7 s), M (37 frames, ~3 s), and L (81 frames, ~7 s). Through rollout, the L model can further generate long multi-view driving videos of up to 241 frames (~20 s). GenieDrive consistently outperforms previous occupancy-based methods across all metrics, while also enabling much longer video generation.

<div align="center"> <img src="assets/table_video.png" width="65%"> <p><em>Performance of Multi-View Video Generation</em></p> </div>

📝 Citation

@article{yang2025geniedrive,
  author    = {Yang, Zhenya and Liu, Zhe and Lu, Yuxiang and Hou, Liping and Miao, Chenxuan and Peng, Siyi and Feng, Bailan and Bai, Xiang and Zhao, Hengshuang},
  title     = {GenieDrive: Towards Physics-Aware Driving World Model with 4D Occupancy Guided Video Generation},
  journal   = {arXiv:2512.12751},
  year      = {2025},
}

Acknowledgements

We thank these great works and open-source repositories: I2-World, UniScene, DynamicCity, MMDectection3D and VideoX-Fun.