Emergent Outlier View Rejection in Visual Geometry Grounded Transformers

Jisang Han^1,2* · Sunghwan Hong^3* · Jaewoo Jung¹ · Wooseok Jang¹ · Honggyu An¹ · Qianqian Wang⁴ · Seungryong Kim^1† · Chen Feng^2†

¹KAIST AI, ²New York University, ³ETH AI Center, ETH Zurich, ⁴UC Berkeley

arXiv | Paper (High quality) | Project Page

We reveal that Visual Geometry Grounded Transformers (VGGT) has a built-in ability to detect outliers, which we leverage to perform outlier-view rejection without any fine-tuning.

What to expect:

• Demo inference code
• Evaluation code
• Visualization code

Installation

Our code is developed based on pytorch 2.5.1, CUDA 12.1 and python 3.10.

We recommend using conda for installation:

conda create -n robust_vggt python=3.10
conda activate robust_vggt

pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

Running Demo

To run the robust reconstruction demo with outlier rejection:

python robust_vggt.py --image-dir examples/trevi

python robust_vggt.py --image-dir examples/notredame --rej-thresh 0.3

Citation

@article{han2025emergent,
  title={Emergent Outlier View Rejection in Visual Geometry Grounded Transformers},
  author={Han, Jisang and Hong, Sunghwan and Jung, Jaewoo and Jang, Wooseok and An, Honggyu and Wang, Qianqian and Kim, Seungryong and Feng, Chen},
  journal={arXiv preprint arXiv:2512.04012},
  year={2025}
}

Acknowledgement

We thank the authors of VGGT for their excellent work and code, which served as the foundation for this project.