⚡️ FastVGGT: Training-Free Acceleration of Visual Geometry Transformer

Media Analytics & Computing Laboratory; AUTOLAB

You Shen, Zhipeng Zhang, Yansong Qu, Liujuan Cao

• [Sep 10, 2025] Added COLMAP outputs.
• [Sep 8, 2025] Added custom dataset evaluation.
• [Sep 3, 2025] Paper release.
• [Sep 2, 2025] Code release.

FastVGGT observes strong similarity in attention maps and leverages it to design a training-free acceleration method for long-sequence 3D reconstruction, achieving up to 4× faster inference without sacrificing accuracy.

First, create a virtual environment using Conda, clone this repository to your local machine, and install the required dependencies.

conda create -n fastvggt python=3.10
conda activate fastvggt
git clone [email protected]:mystorm16/FastVGGT.git
cd FastVGGT
pip install -r requirements.txt

Next, prepare the ScanNet dataset: http://www.scan-net.org/ScanNet/

Then, download the VGGT checkpoint (we use the checkpoint link provided in https://github.com/facebookresearch/vggt/tree/evaluation/evaluation):

wget https://huggingface.co/facebook/VGGT_tracker_fixed/resolve/main/model_tracker_fixed_e20.pt

Finally, configure the dataset path and VGGT checkpoint path. For example:

    parser.add_argument(
        "--data_dir", type=Path, default="/data/scannetv2/process_scannet"
    )
    parser.add_argument(
        "--gt_ply_dir",
        type=Path,
        default="/data/scannetv2/OpenDataLab___ScanNet_v2/raw/scans",
    )
    parser.add_argument(
        "--ckpt_path",
        type=str,
        default="./ckpt/model_tracker_fixed_e20.pt",
    )

Note: A large number of input_frames may significantly slow down saving the visualization results. Please try using a smaller number first.

python eval/eval_scannet.py --input_frame 30 --vis_attn_map --merging 0

We observe that many token-level attention maps are highly similar in each block, motivating our optimization of the Global Attention module.

Please organize the data according to the following directory:

<data_path>/
├── images/       
│   ├── 000000.jpg
│   ├── 000001.jpg
│   └── ...
├── pose/                # Optional: Camera poses
│   ├── 000000.txt 
│   ├── 000001.txt
│   └── ...
└── gt_ply/              # Optional: GT point cloud
    └── scene_xxx.ply   

• Required: images/
• Additionally required when --enable_evaluation is enabled: pose/ and gt_ply/

Inference only:

python eval/eval_custom.py \
  --data_path /path/to/your_dataset \
  --output_path ./eval_results_custom \
  --plot

Inference + Evaluation (requires pose/ and gt_ply/):

python eval/eval_custom.py \
  --data_path /path/to/your_dataset \
  --enable_evaluation \
  --output_path ./eval_results_custom \
  --plot

If you want the results in COLMAP’s format:

python eval/eval_custom_colmap.py \
  --data_path /path/to/your_dataset \
  --output_path ./eval_results_custom_colmap \

Evaluate FastVGGT on the ScanNet dataset with 1,000 input images. The --merging parameter specifies the block index at which the merging strategy is applied:

python eval/eval_scannet.py --input_frame 1000 --merging 0

Evaluate Baseline VGGT on the ScanNet dataset with 1,000 input images:

python eval/eval_scannet.py --input_frame 1000

Evaluate across two datasets, sampling keyframes every 10 frames:

python eval/eval_7andN.py --kf 10

• Thanks to these great repositories: VGGT, Dust3r, Fast3R, CUT3R, MV-DUSt3R+, StreamVGGT, VGGT-Long, ToMeSD and many other inspiring works in the community.

• Special thanks to Jianyuan Wang for his valuable discussions and suggestions on this work.

See the LICENSE file for details about the license under which this code is made available.

If you find this project helpful, please consider citing the following paper:

@article{shen2025fastvggt,
  title={FastVGGT: Training-Free Acceleration of Visual Geometry Transformer},
  author={Shen, You and Zhang, Zhipeng and Qu, Yansong and Cao, Liujuan},
  journal={arXiv preprint arXiv:2509.02560},
  year={2025}
}