Training Library

gsplat

gsplat is an open-source CUDA-accelerated differentiable rasterization library with Python bindings, maintained around the Nerfstudio ecosystem.

Stage: TrainingResearch codeCustom trainersLarge-scene experiments

Open official linkOfficial Documentation->

What It Does

gsplat is lower level than a desktop app. It gives you the rasterization, projection, batching, sparse gradients, depth rendering, and related primitives needed to build or modify Gaussian Splatting methods.

If you want a ready-to-run product workflow, start with Nerfstudio or Postshot. If you want to implement a method, profile a renderer, or train at scale, gsplat is one of the most important libraries to know.

How To Use It In 3DGS

Install gsplat in a PyTorch/CUDA environment.
Use its examples to fit COLMAP captures or single images before writing custom logic.
Build your trainer around the rasterization and densification APIs.
Export trained splats to PLY or another format for viewing and conversion.

Things To Watch

It assumes comfort with Python, PyTorch, CUDA environments, and research code.
API details can evolve quickly, so pin versions for reproducible experiments.
It is not a full capture-to-share application by itself.

Official And Related Links

Official Documentation GitHub Repository Fit a COLMAP Capture

API Capabilities To Notice

Beyond basic rasterization, gsplat exposes depth rendering, batch rasterization, N-D feature rendering, sparse gradients, compression, and distributed rasterization paths.
Its docs include practical examples for COLMAP captures, single-image fitting, large-scene rendering, and NCore v4 captures.
If you are replacing `diff-gaussian-rasterization`, read the migration docs rather than swapping imports blindly.

Rasterization API Migration Notes