ffsplat is a powerful framework to convert and compress 3D Gaussian Splatting scenes.

NOTE: this code is pre-alpha, under heavy development, a community developer preview. Please expect more documentation in the coming weeks (July 2025).

• Github repository: https://github.com/w-m/ffsplat/

• Install CUDA 12.x
• Install uv

Run

uvx --from git+https://github.com/w-m/ffsplat@main ffsplat live --help

Then open a web browser to localhost:8080 (or the address that is printed in the viser box on stdout). On the first launch, it will take a few minutes to compile gsplat kernels before a rendered image shows up.

Check out the project:

git clone https://github.com/w-m/ffsplat.git
cd ffsplat

Set up the virtual environment with uv:

make install

Run the live encoding:

uv run ffsplat live \
    --input /data/gaussian_splatting/mini-splatting2/truck_sparse/point_cloud/iteration_18000/point_cloud.ply \
    --input-format=3DGS-INRIA.ply \
    --dataset-path /data/gaussian_splatting/tandt_db/tandt/truck/ \
    --verbose

Check the types and formatting of the code base:

make check

Find the documentation for the encoding and decoding description, and the yaml container format in CONTAINER_FORMAT.md.