G²VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning

<a href="https://gordonhu608.github.io">Wenbo Hu</a><sup>1,2*</sup>,
<a href="https://openreview.net/profile?id=~Jingli_Lin1">Jingli Lin</a><sup>1,3*</sup>,
<a href="https://openreview.net/profile?id=~Yilin_Long2">Yilin Long</a><sup>1,4*</sup>,
<a href="https://openreview.net/profile?id=~Yunlong_Ran1">Yunlong Ran</a><sup>1,5</sup>,
<a href="https://jianglh-whu.github.io/">Lihan Jiang</a><sup>1,6</sup>,
<a href="https://github.com/yyfz">Yifan Wang</a><sup>1,3</sup>,
<a href="https://zcmax.github.io/">Chenming Zhu</a><sup>1,7</sup>,
<a href="https://runsenxu.com/">Runsen Xu</a><sup>1,8</sup>,
<a href="https://tai-wang.github.io/">Tai Wang</a><sup>1&dagger;</sup>,
<a href="https://oceanpang.github.io/">Jiangmiao Pang</a><sup>1&dagger;</sup>

¹Shanghai AI Lab, ²UCLA, ³SJTU, ⁴FDU, ⁵ZJU, ⁶USTC, ⁷HKU, ⁸CUHK

^*Equal Contribution    ^†Corresponding Author

📑 Paper | 📖 arXiv | 🌐 Homepage | 🤗 Model

🏠 About

<img src="https://github.com/InternRobotics/G2VLM/raw/main/assets/teaser.png" alt="Dialogue_Teaser" width=100% >



We present <b>G<sup>2</sup>VLM</b>, a geometry grounded vision-language model proficient in both spatial 3D reconstruction and spatial
understanding tasks. For spatial reasoning questions, G<sup>2</sup>VLM can natively predict 3D geometry and employ interleaved reasoning for an answer.

📢 News

• [2026-04-18] 📝 We release our training code in the train folder. Please refer to our training instructions below.
• [2026-02-20] 🔥 Our work is accepted by CVPR 2026!
• [2026-01-14] We release the evaluation code for 3D reconstruction. Please refer to the eval readme here.
• [Coming!] 📝 We will release the checkpoint of G²VLM-SR, a strong spatial reasoning model. Stay tuned!
• [2025-11-27] 🔥 We release the example training data preprocessing code in the data folder.
• [2025-11-27] 🔥 We release the inference code and the checkpoint of G²VLM.
• [2025-11-27] 🔥 We release the paper of G²VLM.

Model

G²VLM is a unified model that integrates both a geometric perception expert for 3D reconstruction and a semantic perception expert for multimodal understanding and spatial reasoning tasks. All tokens can do shared multi-modal self attention in each transformer block.

🚀 Quick Start

1️⃣ Set up environment

git clone https://github.com/InternRobotics/G2VLM
cd G2VLM
conda create -n g2vlm python=3.10 -y
conda activate g2vlm

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

Optional: For training

pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu12torch2.5cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

2️⃣ Download pretrained checkpoint

from huggingface_hub import snapshot_download

save_dir = "models/G2VLM-2B-MoT"
repo_id = "InternRobotics/G2VLM-2B-MoT"
cache_dir = save_dir + "/cache"

snapshot_download(cache_dir=cache_dir,
  local_dir=save_dir,
  repo_id=repo_id,
  local_dir_use_symlinks=False,
  resume_download=True,
  allow_patterns=["*.json", "*.safetensors", "*.bin", "*.py", "*.md", "*.txt"],
)

3️⃣ Run Inference from Command Line

Try our example inference script. Here is the script for 3D recontruction.

# Run with default example images
python inference_recon.py

# Run on your own data (image folder)
python inference_recon.py --image_folder <path/to/your/images_dir>

Here is the script for spatial reasoning task. We encourage to try spatial reasoning with G²VLM-SR which will be released soon!

# Run with default example images and default question
python inference_chat.py

# Run on your own data (image folder) and question
python inference_chat.py --image_path <path/to/your/images> --question "user question"

Optional Arguments:

• --model_path: Path to a custom model checkpoint file.
• --image_folder: Path to the input image directory. (Default: examples/dl3dv)
• --image_path: Path to the image, if you want to specify the image. (Default: examples/25_0.jpg)
• --question: Input question. (Default: If the table (red point) is positioned at 2.6 meters, estimate the depth of the clothes (blue point).)
• --save_path: Path to save the output .ply point cloud. (Default: examples/result.ply)

🔥 Training

We provide pre-train and joint-train scripts in scripts folder

# For pre-train of visual geometry expert only 
bash scripts/pretrain.sh

# For joint-train of semantic perception expert and optionally (freeze or train) visual geometry expert
bash scripts/joint_train.sh

For finetuning or joint training, please edit the config file here data/configs/joint_train.yaml. The actual data config is in this file data/dataset_info.py where you can edit the path to your dataset.

🔗 Citation

If you find our work and this codebase helpful, please consider starring this repo 🌟 and cite:

@article{hu2025g2vlmgeometrygroundedvision,
      title={G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning}, 
      author={Wenbo Hu and Jingli Lin and Yilin Long and Yunlong Ran and Lihan Jiang and Yifan Wang and Chenming Zhu and Runsen Xu and Tai Wang and Jiangmiao Pang},
      year={2025},
      journal={arXiv preprint arXiv:2511.21688},
      url={https://arxiv.org/abs/2511.21688}, 
}

📄 License

G²VLM is licensed under the Apache 2.0.

👏 Acknowledgements

• Bagel: Our codebase is built upon Bagel.
• Pi3: We develop our visual geometric expert based on Pi3.
• VGGT: We thank VGGT for their efforts in visual geometry learning.