FriendlySplat
FriendlySplat is an Apache 2.0 Gaussian Splatting toolkit from Zhejiang University's FAST Lab and Differential Robotics that combines training, pruning, meshing and segmentation in one codebase.
Platform
Python package (Python 3.10 or newer, PyTorch with CUDA), with Docker support
GPU
NVIDIA GPU with CUDA required. Training uses PyTorch and gsplat CUDA kernels, and the Docker image needs a host NVIDIA driver of 530.30 or newer
Pricing
Free
License
Apache 2.0
Compare licence terms ↗
Best for
Research 3DGS training with geometry priors, pruning, meshing and segmentation
Updated
July 2026
FriendlySplat is an open source toolkit that gathers recent 3D Gaussian Splatting research into a single training codebase. It is developed by researchers and contributors from Differential Robotics, the FAST Lab and Zhejiang University, with code in the FriendlySplat repository. FriendlySplat launched publicly in March 2026 under the Apache 2.0 licence.
Training runs from COLMAP-style datasets through the fs-train command, built on gsplat's CUDA kernels with densification strategies drawn from Improved-GS, AbsGS, taming-3dgs, 3DGS-MCMC and mini-splatting. Optional depth and normal priors from models such as MoGe act as weak geometric supervision, and dynamic and sky masks can be supplied alongside the images. Trained scenes export as PLY, compressed PLY or SOG.
Pruning draws on ideas and code from GNS, speedy-splat, GaussianSpa and LightGaussian. A TSDF-style pipeline extracts meshes from trained scenes, and a MaskClustering-style workflow lifts 2D segmentation masks into consistent 3D groups that can produce coarse segmentations or bounding boxes. A helper tool generates the COLMAP reconstruction through HLOC.
The fs-view viewer, built on viser and nerfview, opens the latest checkpoint or PLY in a browser on a local port and shows training metrics and camera frustums while optimization runs. FriendlySplat installs with pip on Python 3.10 or newer and needs PyTorch with CUDA, and its Docker setup requires an NVIDIA host driver of 530.30 or newer. The package is at version 0.1.0, and its authors describe the codebase as still evolving.
Frequently Asked Questions
What is FriendlySplat?
FriendlySplat is an open source 3D Gaussian Splatting toolkit from Zhejiang University's FAST Lab and Differential Robotics that combines training, geometry priors, pruning, mesh extraction and segmentation in one codebase.
What license does FriendlySplat use?
FriendlySplat is released under the Apache 2.0 licence, according to the LICENSE file in its GitHub repository, and is free to use.
What hardware does FriendlySplat need?
FriendlySplat needs an NVIDIA GPU with CUDA, because training runs on PyTorch and gsplat CUDA kernels. FriendlySplat requires Python 3.10 or newer, and its Docker image needs a host NVIDIA driver of 530.30 or newer.
Which formats can FriendlySplat export?
FriendlySplat exports trained scenes as PLY, compressed PLY or PlayCanvas SOG, and can also extract meshes through a TSDF-style pipeline.
What are the alternatives to FriendlySplat?
Alternatives to FriendlySplat include Nerfstudio, the Berkeley framework whose gsplat library FriendlySplat builds on, LichtFeld Studio for native desktop training, and Brush for cross-platform training.
Alternatives
Recent Headlines