QVGGT: Post-Training Quantized Visual Geometry Grounded Transformer

---

Zhizhen Pan1,2, Hesong Wang1, Huan Wang1*

1Westlake University    2Beijing University of Posts and Telecommunications

*Corresponding author

🔥 News

• [2026.07.03] Code and checkpoints released.
• [2026.05.30] Project page released.
• [2026.02.21] The paper has been accepted by CVPR 2026!

📖 Overview

QVGGT is a post-training quantization framework tailored for the Visual Geometry Grounded Transformer (VGGT).

🛠️ Installation

We recommend cloning QVGGT and vggt-hf into the same parent directory:

git clone https://github.com/DDsacu/QVGGT.git
git clone https://github.com/DDsacu/vggt-hf.git

The expected layout is:

workspace/
  QVGGT/
  vggt-hf/

0. Create and activate an environment

conda create -n qvggt python=3.11 -y
conda activate qvggt
pip install --upgrade pip

1. Install PyTorch

Install a CUDA-enabled PyTorch build that matches your local CUDA toolkit and GPU driver.

Please follow the official instructions from PyTorch:

• https://pytorch.org/get-started/locally/

After installation, verify:

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

2. Install this repository's Python dependencies

From the QVGGT_CVPR26 repository root:

cd QVGGT
pip install -r requirements.txt

This installs the common runtime dependencies for QVGGT inference.

3. Install vggt-hf

From the parent workspace directory:

cd ../vggt-hf
pip install -e .

This step is required because QVGGT loads the VGGT model definition from vggt-hf.

4. Build awq_inference_engine

From the repository root:

cd qvggt/runtime/awq_inference_engine
pip install -e .
cd ../../..

This step builds the CUDA extension required by the AWQ inference runtime. This will speed up our reasoning process!

⚡ Quick Start

From the QVGGT_CVPR26 repository root, run inference with a quantized checkpoint:

python scripts/run_inference.py \
  --quantized-checkpoint-path /path/to/model-v2.pt \
  --input-dir /path/to/input_images \
  --output-dir /path/to/output_dir

To additionally export a 3D visualization:

python scripts/run_inference.py \
  --quantized-checkpoint-path /path/to/model-v2.pt \
  --input-dir /path/to/input_images \
  --output-dir /path/to/output_dir \
  --export-glb

🎬 Demo

We also provide a Gradio demo that runs one complete inference pass and visualizes the result.

Bundled example input images are included in:

examples/cat/images

Launch the demo from the repository root:

python scripts/run_gradio_demo.py \
  --quantized-checkpoint-path /path/to/model-v2.pt \
  --port 7860

Then open:

http://127.0.0.1:7860

🛣️ TODO Roadmap

Planned next-stage public release items:

• quantization pipeline
• calibration data preparation
• evaluation on geometry benchmarks
• example checkpoints and download instructions
• demo assets and visualization examples

Acknowledgements

This codebase builds on VGGT and AWQ. Thanks for their awesome work!

Contact

For questions or collaboration requests, please feel free to open an issue or contact me at zhizhenpan9@gmail.com

Citation

If you find QVGGT useful, please consider citing:

@article{pan2026qvggt,
        title={QVGGT: Post-Training Quantized Visual Geometry Grounded Transformer},
        author={Pan, Zhizhen and Wang, Hesong and Wang, Huan},
        journal={arXiv preprint arXiv:2605.31124},
        year={2026}
      }