Mobile-GS: Real-time Gaussian Splatting for Mobile Devices

If you like our project, please give us a star ⭐ on GitHub for latest update.

😮 Highlights

You are allowed to use this repository in commercial usage for free.

🚩 Updates

Welcome to watch 👀 this repository for the latest updates.

✅ [2026.3.13] : Release project page.

✅ [2026.3.13] : Code Release.

Setup

For installation: On current Blackwell/CUDA 13 systems, use uv, Python 3.12, and the latest PyTorch nightly cu130 wheels.

git clone git@github.com:xiaobiaodu/Mobile-GS.git

cd Mobile-GS
uv venv --python 3.12
source .venv/bin/activate

uv pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130
TORCH_CUDA_ARCH_LIST=12.0 uv pip install --no-build-isolation -r requirements.txt

--no-build-isolation is required because the local CUDA extension submodules build against the already-installed PyTorch package.

For a single-command bring-up:

./scripts/setup_uv_py312_cu130.sh

To verify imports, CUDA extensions, and every visible GPU afterwards:

.venv/bin/python scripts/verify_env.py

Install TMC (GPCC), and add tmc3 to your environment variable or manually specify its location in the code (lines 243 and 258, this script is sourced from HAC++).

For RAPIDS package troubleshooting, please refer to the CUML Installation Guide.

GPU selection

The training scripts currently run one process on one CUDA device. They now respect LOCAL_RANK, so you can choose a device with:

CUDA_VISIBLE_DEVICES=1 python train.py ...

or from a launcher that sets LOCAL_RANK.

We used Mip-NeRF 360, Tanks & Temples, and Deep Blending.

Running

Pre-training (Mini-Splatting)

#for outdoor scenes (e.g., Mip-NeRF 360 outdoor and T&T scenes)
python pretrain.py -s <path to COLMAP>  -m  <model path> --eval --imp_metric outdoor --sh_degree 3   --iterations 30000
#for indoor scenes (e.g., Mip-NeRF 360 indoor and DB scenes)
python pretrain.py -s <path to COLMAP>  -m  <model path> --eval --imp_metric indoor --sh_degree 3   --iterations 30000

Fine-tune

python train.py -s  <path to COLMAP> -m <model path>  --eval --start_checkpoint  <model path>/chkpnt30000.pth 

# To improve rendering perofmrance, you can use multi-view training from MVGS. It may cause longer training time and memory.
python train.py -s  <path to COLMAP> -m <model path>  --eval --start_checkpoint  <model path>/chkpnt30000.pth   --mv  3

WebGPU Viewer

This repo now includes a self-contained static browser viewer in docs/ that is suitable for GitHub Pages. It exports a Mobile-GS point cloud into compact browser assets and renders sorted billboard splats with:

• WebGPU in Chromium-class browsers when available
• WebGPU in Firefox when enabled and hardware allows it
• WebGL2 fallback in WebKit/Safari-class browsers

Current scope:

• Uses real Mobile-GS outputs (x/y/z, SH DC color, opacity, isotropic billboard radius)
• Ships a trained sample scene at docs/assets/scenes/lego-mini.{json,bin}
• Is designed for real-time inspection in the browser, not CUDA-rasterizer parity

Example flow:

# 1. Train or pretrain a scene so you have a point_cloud.ply
python pretrain.py -s data/nerf_synthetic/lego_mini -m output/webgpu_lego_mini_pretrain --eval --imp_metric indoor --sh_degree 3 --iterations 1000 -r 8 --save_iterations 1000 --checkpoint_iterations 1000 --test_iterations 1000 --port 6010

# 2. Export the trained PLY into the Pages viewer assets
python scripts/export_web_scene.py \
  --ply output/webgpu_lego_mini_pretrain/point_cloud/iteration_1000/point_cloud.ply \
  --cameras output/webgpu_lego_mini_pretrain/cameras.json \
  --output-dir docs/assets/scenes \
  --slug lego-mini \
  --title "Mobile-GS Lego Mini"

# 3. Serve docs/ locally
python3 -m http.server 8008 -d docs

# 4. Open the viewer
# http://127.0.0.1:8008/

Browser verification:

npm install
npx playwright install chromium firefox webkit
npm run test:viewer
npm run test:viewer:subpath

The verification script saves screenshots to artifacts/browser-tests/ and currently exercises Chromium, Firefox, and WebKit locally. On Linux, Chromium and Firefox are launched inside Xvfb so the GPU-backed render paths can still be exercised from this machine. test:viewer:subpath simulates GitHub project pages under /Mobile-GS/.

For GitHub Pages:

• point Pages at the docs/ directory
• or enable the checked-in GitHub Actions workflow at .github/workflows/pages.yml
• keep scene assets under docs/assets/scenes/
• the checked-in docs/index.html will load the bundled lego-mini scene automatically

Evaluation

python render.py -s <path to COLMAP> -m <model path> --decode
python metrics.py -m <model path> 

--decode

Rendering with the compressed file (comp.xz), otherwise using the ply file. The results are the same regardless of this option.

👍 Acknowledgement

This work is built on many amazing research works and open-source projects, thanks a lot to all the authors for sharing!

• Mini-Splatting
• OMG
• MVGS

BibTeX

@misc{du2026mobile-gs,
      title={Mobile-GS: Real-time Gaussian Splatting for Mobile Devices}, 
      author={Xiaobiao Du and Yida Wang and Kun Zhan and Xin Yu},
      year={2026},
      eprint={2603.11531},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.11531}, 
}