FriendlySplat Launches an Open Toolkit for Gaussian Splatting

A growing number of research teams and developers are beginning to package the rapidly expanding gaussian splatting ecosystem into more complete toolkits. The latest entry into that category is FriendlySplat, an open source platform designed to bring a wide range of modern 3D Gaussian Splatting workflows into a single, approachable framework.

FriendlySplat positions itself as a full training and inspection environment for radiance field reconstruction. The platform organizes the entire process of building a Gaussian scene into a modular pipeline. Those include geometry priors, pose refinement, pruning strategies, segmentation tools, and post processing steps, all integrated into the same workflow.

One of the more notable capabilities is the platform’s support for geometry supervision during training. FriendlySplat can incorporate depth and normal estimates from models such as MoGe and Pi3, using them as weak priors to guide the optimization process. This approach can help stabilize reconstructions in areas where photometric supervision alone struggles, such as textureless surfaces or large planar regions. Once a scene is trained, the system can also extract meshes through a TSDF-style pipeline, allowing the radiance field representation to be converted into more traditional geometry for downstream use.

Model compression is another area where the toolkit places a strong emphasis. FriendlySplat integrates both hard pruning techniques like PUP-GS and soft pruning approaches such as GNS. Together these allow users to aggressively reduce splat counts while attempting to preserve visual fidelity.

Beyond training and compression, the system also explores ways of structuring scenes semantically. Using MaskClustering-style workflows, FriendlySplat can lift 2D segmentation masks into consistent 3D groupings within the splat scene. The resulting clusters can be used to generate coarse segmentations or bounding boxes, opening the door to tasks such as scene understanding, object level editing, or dataset preparation for downstream learning systems.

FriendlySplat includes a real-time viewer built on Nerfstudio‘s Viser that displays the evolving scene alongside training metrics and camera frustums. Instead of waiting for training to complete before diagnosing issues, users can monitor convergence, inspect geometry, and identify problematic viewpoints while optimization is still running.

As individual Gaussian splatting techniques mature, developers are increasingly bundling them into unified environments that handle everything from reconstruction and editing to compression and analysis. FriendlySplat joins a growing set of tools attempting to make these advanced workflows easier to experiment with, while still exposing the underlying components researchers and developers rely on. FriendlySplat comes with an Apache 2.0 license and can be accessed here.