gsplat

What it is and what it does

gsplat is a Python library that implements fast, memory-efficient CUDA-accelerated rasterization of 3D Gaussians for neural rendering. It builds on the 3D Gaussian Splatting paper but optimizes the core rendering kernels and adds support for multiple camera models (pinhole, fisheye, LiDAR), sparse rendering, multi-GPU distributed training, and alternative optimization strategies like MCMC. The library is designed for training and rendering neural radiance fields and 3D scene reconstructions, with both dense and sparse rendering paths.

The package depends on PyTorch for tensor operations, numpy for numerical work, torch's ninja build system, and several utility libraries (jaxtyping, rich, typing_extensions). It requires a CUDA-capable GPU and compiles CUDA kernels either during installation or on first run. Recent updates (as of mid-2026) include spherical harmonics improvements, orthographic camera support, LiDAR rendering, and experimental low-latency inference rendering via HiGS.

Use it for:

• Train 3D Gaussian splatting models on COLMAP or NCore captures for real-time novel-view synthesis.
• Reconstruct large-scale 3D scenes with sparse rendering and multi-GPU training for memory efficiency.
• Render pre-trained Gaussian scenes at low latency using the experimental HiGS inference path.
• Optimize 3D Gaussian parameters using MCMC or standard densification strategies for scene fitting.
• Fit 2D images or 3D point clouds with differentiable Gaussian rasterization for inverse graphics tasks.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

gsplat is a CUDA-accelerated Python library for rasterizing 3D Gaussians, implementing differentiable rendering based on 3D Gaussian Splatting with support for multiple camera models and optimization strategies.

Yes, if you have a CUDA GPU and need fast, differentiable 3D Gaussian rendering. The library is actively maintained, has low install friction, and offers significant performance and memory improvements over the reference implementation. However, verify the license terms in the repository first, and confirm your PyTorch and CUDA versions are supported by pre-compiled wheels or compatible with JIT compilation.

Install

gsplat on PyPI

pip install gsplat

uv add gsplat

poetry add gsplat

Installing gsplat

Before you install

Low friction installation via pip with JIT CUDA compilation on first run, or pre-compiled wheels for specific PyTorch-CUDA combinations. Active maintenance with recent updates; requires PyTorch 2.7+ and a CUDA-capable GPU.

License in practice

License status is unclear—no SPDX identifier or raw license text is available in the package metadata. Verify the actual license terms in the repository before use in proprietary or restricted contexts.

Quickstart

pip install gsplat
import gsplat
# Rasterize 3D Gaussians using gsplat's CUDA kernels
# (exact API depends on your scene and camera setup)

Requires PyTorch 2.7+, a CUDA-capable NVIDIA GPU, and CUDA toolkit installed; CUDA code compiles on first run if using pip wheels.

Verify before relying

• Exact performance gains (e.g., '4x less GPU memory', '15% less time') relative to the official Gaussian Splatting implementation—claimed in description but not independently verified here.
• Whether pre-compiled wheels cover your specific PyTorch and CUDA version combination.
• Full scope of the experimental inference rendering path (HiGS) and its production readiness.

Package facts

Evidence: gsplat-1.5.3-py3-none-any.whl

Keywords: gaussian, splatting, cuda

Tags

More Artificial Intelligence packages

Further reading