nerfstudio-project/gsplat

What it solves

gsplat is a CUDA-accelerated library designed to make the rasterization of 3D Gaussians more efficient. It addresses the high memory and computational costs associated with rendering radiance fields, providing a faster and more memory-efficient alternative to the original 3D Gaussian Splatting implementation.

How it works

The library provides Python bindings for high-performance CUDA kernels. It implements various rasterization paths, including a dense rendering path and an experimental inference-only path (HiGS) that uses macro-tile fused rasterization and fp16 scene packing for low-latency rendering. It supports a wide range of sensor models (pinhole, FTheta, fisheye, and LiDAR) and integrates advanced optimization strategies like MCMC-style Gaussian optimization and 3DGUT extensions.

Who it’s for

Researchers and developers working on 3D scene reconstruction, real-time rendering, and neural radiance fields (NeRFs) who need a high-performance, PyTorch-compatible library for training and rendering Gaussian splats.

Highlights

• High Performance: Training takes up to 4x less GPU memory and finishes up to 15% faster than the official implementation.
• Wide Sensor Support: Includes native support for pinhole, FTheta, fisheye, and LiDAR camera models.
• Multi-GPU Support: Distributed support for the dense rendering path.
• Inference Optimization: An experimental HiGS-based path for low-latency, inference-only rendering of pre-trained scenes.
• Flexible Integration: Supports arbitrary batching over multiple scenes and viewpoints.