• [2026.01] 🚀📄 Published as a conference paper at SID’s Display Week 2026 !!

XStreamVGGT: Extremely Memory-Efficient Streaming Vision Geometry Grounded Transformer with KV Cache Compression

Zunhai Su^*, Weihao Ye^*, Hansen Feng, Keyu Fan, Jing Zhang, Dahai Yu, Zhengwu Liu, Ngai Wong

^* Equal contribution.

We propose XStreamVGGT, a tuning-free and extremely memory-efficient streaming vision geometry transformer that compresses KV cache through joint token pruning and distribution-aware quantization. By removing redundant tokens and quantizing remaining KV representations, XStreamVGGT achieves up to 4.42× memory reduction and 5.48× inference speedup, while maintaining mostly negligible performance degradation. This enables scalable and long-horizon streaming 3D reconstruction in real-world applications.

We recommend using Conda to set up the environment:

conda env create -f StreamVGGT_environment.yml
conda activate streamvggt

Download the pretrained StreamVGGT model weights from:

• https://huggingface.co/lch01/StreamVGGT

After downloading, place the checkpoint file under:

XStreamVGGT/ckpt/

Please refer to the official instructions of the following repositories to prepare the evaluation datasets:

• MonST3R
• Spann3R

The supported datasets include:

• Sintel
• Bonn
• KITTI
• NYU-v2
• ScanNet
• 7Scenes
• Neural-RGBD

The overall folder structure should be organized as follows:

XStreamVGGT
├── ckpt/
│   └── checkpoints.pth
├── config/
│   ├── ...
├── data/
│   ├── eval/
│   │   ├── 7scenes
│   │   ├── bonn
│   │   ├── kitti
│   │   ├── neural_rgbd
│   │   ├── nyu-v2
│   │   ├── scannetv2
│   │   └── sintel
│   ├── train/
│   │   ├── processed_arkitscenes
│   │   ├── ...
└── src/
    ├── ...

To evaluate XStreamVGGT with KV cache pruning enabled:

CUDA_VISIBLE_DEVICES=0 \
KV_POOL_SIZE=16 \
KV_CACHE_SIZE=2048 \
bash eval/video_depth/run.sh

To evaluate the version with KV cache quantization, please switch to the corresponding branch first:

git checkout prune_and_quantize

Then run:

CUDA_VISIBLE_DEVICES=0 \
KV_QUANT_MODE=KCVT \
KV_POOL_SIZE=16 \
KV_CACHE_SIZE=2048 \
bash eval/video_depth/run.sh

This codebase is built upon StreamVGGT and related streaming 3D reconstruction frameworks. We thank the authors for their open-source contributions.