Grendel-GS provides a distributed training system for 3D Gaussian Splatting (3DGS), enabling faster training, support for more Gaussians, and reconstruction of larger, higher-resolution scenes. It's designed as a drop-in replacement for existing 3DGS workflows, targeting researchers and practitioners needing to scale 3DGS capabilities.

How It Works

Grendel-GS extends the original 3DGS implementation with distributed training capabilities using PyTorch's torchrun. It supports larger batch sizes and leverages multiple GPUs to accelerate training and reduce memory per GPU. The system maintains the core 3DGS algorithm, ensuring compatibility while offering performance gains.

Quick Start & Requirements

• Install: Clone with --recursive and use conda env create --file environment.yml.
• Prerequisites: Conda, compatible GPU drivers, CUDA environment, PyTorch >= 1.10. Requires compiling custom CUDA kernels for diff-gaussian-rasterization and simple-knn.
• Dataset: COLMAP format datasets (e.g., Mip360, Tanks&Temples).
• Links: Paper, Project Page

Highlighted Details

• Achieves >3.5x faster training on Mip360 with 4 GPUs and batch size 4.
• Enables training large-scale 4K scenes with >40 million Gaussians without OOM.
• Trains scenes like "Truck" from Tanks&Temples to PSNR 23.79 in 45 seconds (7k iterations) on 4 GPUs.
• Supports gsplat as an alternative CUDA backend.

Maintenance & Community

• Actively developed by NYU Systems.
• Contact: hz3496@nyu.edu for collaboration/contributions.

Licensing & Compatibility

• Apache License Version 2.0.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The repository currently focuses on training and rendering; it does not include interactive or network viewers found in the original 3DGS codebase. Support for additional features is under development.