FriendlySplat

Train fast. Inspect live. Edit effortlessly.

TL;DR: FriendlySplat is a user-friendly, open-source Gaussian Splatting toolkit, integrating SOTA features

into a unified platform for training, pruning, meshing and segmentation.

Framework

FriendlySplat is designed as a modular 3DGS training loop: it parses a scene, initializes Gaussians, renders and optimizes them, and supports optional modules for geometry priors, post-processing, pruning, pose refinement, and runtime inspection. The same pipeline also supports checkpointing, splat export, evaluation, and downstream geometry workflows.

FriendlySplat framework diagram

Features

Geometry Supervision + Mesh Extraction

FriendlySplat leverages depth and normal priors from models like Moge and Pi3 as weak supervision, enabling robust convergence even in textureless regions. It also supports TSDF-style mesh extraction.

High-Ratio Pruning Support

FriendlySplat integrates both hard pruning (PUP-GS) and soft pruning (GNS) algorithms, supporting aggressive, high-ratio pruning to achieve significant model compression.

MaskClustering-Style 2D-to-3D Lift

Lifts 2D masks into consistent 3D structures via Instascene/MaskClustering, providing coarse segmentations and 3D bboxes for downstream tasks.

Live Viewer with Charts and Camera Frustums

A real-time viewer visualizes training metrics, rendering states, and camera frustums, enabling early diagnostics and instant progress tracking without waiting.

Acknowledgements

FriendlySplat is built with substantial help from the broader Gaussian Splatting community. We first thank gaussian-splatting and gsplat for efficient CUDA kernels and strong feature integration.

We also thank Improved-GS, AbsGS, taming-3dgs, 3dgs-mcmc, and mini-splatting for high-quality densification implementations and references.

For pruning-related ideas and code references, we thank GNS, speedy-splat, GaussianSpa, and LightGaussian.

We also thank PGSR, 2DGS, GGGS, dn-splatter, mvsanywhere, and 2DGS++ for their explorations of geometry regularization and high-quality code releases.

We further thank CityGaussian for valuable code references on urban-scale scene reconstruction, and InstaScene together with MaskClustering for 2D-to-3D lifting references.

Finally, special thanks to XiaoBin2001 for helpful suggestions throughout development.