Reproduce “On Scaling Up 3D Gaussian Splatting Training” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hexu Zhao, Haoyang Weng, Daohan Lu, Ang Li, Jinyang Li, Aurojit Panda, Saining Xie
Grendel, a distributed system for 3D Gaussian Splatting (3DGS) training developed by researchers from NYU and Pacific Northwest National Laboratory, overcomes single-GPU memory and computational limits to enable high-resolution and large-scale 3D scene reconstruction. It achieves up to a 5.33x speedup on 16 GPUs for medium scenes and attains a 27.28 PSNR on the 4K Rubble dataset by supporting over 230 million Gaussians, surpassing previous single-GPU results.
Key insights
- Reconstruction quality in 3DGS is directly tied to the number of Gaussians, often requiring capacities far beyond what a single high-end GPU can manage for large, complex scenes.
- Effective distributed 3DGS training necessitates a specialized system design that carefully manages mixed parallelism and dynamically balances computational loads, as generic distributed deep learning frameworks are insufficient.
- Careful hyperparameter scaling for Adam (square-root for learning rate, exponential for momentum) is crucial for maintaining training stability and convergence characteristics when moving from batch size 1 to larger batches in 3DGS.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.