LiAuto-GeoX: Efficient Grounded Driving Transformer

Jiawei Lian^1,2,*, Haoyi Sun^2,*, Yang Wu^1,*, Lifu Mu^2,*, Siyuan Wang^1,2, Le Hui^3,4,†, Ning Mao^2,†,‡, Tao Wei², Pan Zhou², Kun Zhan², Jian Yang^1,†

¹🎓 Nanjing University of Science and Technology    ²🏢 Li Auto Inc.    ³🎓 Northwestern Polytechnical University   
⁴🎓 Department of Computing, The Hong Kong Polytechnic University

^*Equal Contribution    ^†✉️ Corresponding Author    ^‡Project Leader

• [✓] Paper Release
• [✓] LiAuto-GeoX Weight
• [✓] Inference Instructions
• GeoX-Large
• Data Processing
• Training Pipeline

Before using the models, please request access to the checkpoints once they are released.
All released models will be evaluated under the same protocol as reported in the paper.

Install the required dependencies:

pip install -r requirements.txt

Single Frame Example - Basic inference with RGB images:

CUDA_VISIBLE_DEVICES=2 python inference.py \
    --image_folder /path/to/your/images \
    --port 8082

RGB + Sky Mask Example - Filter out sky regions for cleaner reconstruction:

CUDA_VISIBLE_DEVICES=2 python inference.py \
    --image_folder /path/to/your/images \
    --port 8082 \
    --mask_sky

RGB + Pose Example - Use ground truth camera poses for better accuracy:

CUDA_VISIBLE_DEVICES=2 python inference.py \
    --image_folder /path/to/your/images \
    --camera_folder /path/to/your/cameras \
    --port 8083

After running inference, open your browser and navigate to http://localhost:PORT (replace PORT with your specified port) to visualize the 3D reconstruction results interactively.

Additional Options:

• --conf_threshold: Adjust the confidence threshold (default: 10.0) to filter low-confidence points. Lower values show more points, higher values show fewer but more confident points.
• --mask_black_bg: Filter out black background pixels
• --mask_white_bg: Filter out white background pixels
• --save_glb: Export the reconstruction as a GLB file

Thanks to these great repositories: DINOv2, CUT3R, VGGT, DA3, PI3, DVGT, OmniVGGT, FastVGGT, LiteVGGT, SparseWorld-TC, and many other inspiring works in the community.

If you find LiAuto-GeoX useful for your work, please cite:

@article{lian2026geox,
  author    = {Lian, Jiawei and Sun, Haoyi and Wu, Yang and Mu, Lifu and Wang, Siyuan and Wei, Tao and Hui, Le and Mao, Ning and Zhou, Pan and Zhan, Kun and Yang, Jian},
  title     = {LiAuto-GeoX: Efficient Grounded Driving Transformer},
  journal   = {arXiv:2606.05774},
  year      = {2026},
}