FlashVGGT: Efficient and Scalable Visual Geometry Transformers with Compressed Descriptor Attention

CVPR 2026

Zipeng Wang · Dan Xu

https://github.com/user-attachments/assets/3347dbe0-f3c0-48d3-9611-1516b59fbf94

TL;DR: Accelerate VGGT by spatially resampling keys and values for global attention.

Updates

• [05/02/2026] Evaluation code is released.
• [05/02/2026] Training code (both single-forward and streaming settings) is released.
• [05/02/2026] Code and checkpoints for FlashVGGT are released.

Overview

Instead of applying dense global attention across all tokens, FlashVGGT compresses spatial information from each frame into a compact set of descriptor tokens. Global attention is then computed as cross-attention between the full set of image tokens and this smaller descriptor set, significantly reducing computational overhead. Moreover, the compactness of the descriptors enables online inference over long sequences via a chunk-recursive mechanism that reuses cached descriptors from previous chunks.

Installation

Environment Setup

First, you should clone the repository and create an anaconda environment.

git clone https://github.com/wzpscott/FlashVGGT.git
cd FlashVGGT
conda create -n flashvggt python=3.10 -y
conda activate flashvggt

Then, You can use the following command to install the dependencies.

pip install -r requirements.txt

You can also install FlashVGGT as a package.

pip install -e . --no-deps

Checkpoints

You can download the checkpoints for single-forward and streaming variants of FlashVGGT from the HuggingFace. You should download the checkpoints to the ckpts folder.

# Create the checkpoints directory
mkdir -p ckpts

# Download the standard model
huggingface-cli download ZipW/FlashVGGT flashvggt.pt --local-dir ckpts

# Download the streaming model
huggingface-cli download ZipW/FlashVGGT flashvggt_stream.pt --local-dir ckpts

Quick Start

We provide a demo script demo_o3d.py to visualize the 3D reconstruction results as point clouds using Open3D. The output is a .ply file that can be easily visualized with most 3D viewers.

Usage Examples

Standard FlashVGGT Inference:

To run the standard FlashVGGT model on a folder of images:

python demo_o3d.py \
    --model FlashVGGT \
    --image_folder ./examples/garden/ \
    --output_dir outputs/

Streaming FlashVGGT Inference:

To run the streaming variant (FlashVGGTStream) which is optimized for longer sequences:

python demo_o3d.py \
    --model FlashVGGTStream \
    --image_folder ./examples/garden/ \
    --chunksize 10 \
    --output_dir outputs/

Key Arguments

• --model: Choose between FlashVGGT (single-forward) and FlashVGGTStream (streaming inference). Default is FlashVGGT.
• --image_folder: Path to the directory containing input images. Default is ./examples/garden/.
• --output_dir: Directory where the generated .ply point cloud will be saved. Default is outputs/.
• --chunksize: Frame chunk size for FlashVGGTStream streaming inference. Default is 10.
• --max_points: Maximum number of points to include in the output point cloud. Default is 1000000.
• --conf_threshold: Percentage of low-confidence points to filter out (0-100). Default is 40.0.
• --kv_downfactor: KV downfactor for attention compression. Default is 4.
• --keyframe_every: Keyframe interval for the standard FlashVGGT model. Default is 200.

Training

The training code for FlashVGGT (both single-forward and streaming settings) is available in the training branch. Please refer to the Training README for detailed instructions on installation, dataset preparation, and training commands.

Evaluation

The evaluation code is based on MonST3R and CUT3R.

You can use the following command to evaluate the model.

python eval.py --config-name dense_recon num_frames=100 save_name=dense_recon_100
python eval.py --config-name dense_recon num_frames=500 save_name=dense_recon_500
python eval.py --config-name dense_recon num_frames=1000 save_name=dense_recon_1000

The evaluation results are saved in the eval/logs/dense_recon folder.

Acknowledgements

Our code is based on the following awesome repositories:

• VGGT
• FastVGGT
• StreamVGGT
• CUT3R
• TTT3R

We thank the authors for releasing their code!

Citation

If you find our work useful, please cite:

@inproceedings{wang2025flashvggt,
  title={FlashVGGT: Efficient and Scalable Visual Geometry Transformers with Compressed Descriptor Attention},
  author={Wang, Zipeng and Xu, Dan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}