Reproduce “GS-Scale: Unlocking Large-Scale 3D Gaussian Splatting Training via Host Offloading” 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
Donghyun Lee, Dawoon Jeong, Jae W. Lee, Hongil Yoon
GS-Scale introduces a system-level approach to enable large-scale 3D Gaussian Splatting training on a single consumer-grade GPU by offloading data to host memory. The system reduces peak GPU memory usage by 3.98x geomean and allows training with up to four times more Gaussians, resulting in improved rendering quality while maintaining competitive training speeds.
Key insights
- Only a small fraction (averaging 8.28%) of total Gaussians are actively involved in the forward and backward passes during any given training iteration, making offloading feasible.
- Keeping only geometric parameters on the GPU is sufficient for rapid frustum culling, allowing most other data to reside on the CPU.
- Optimizer updates for Gaussians with zero gradients can be precisely deferred for multiple iterations, significantly reducing CPU memory access for updates without quality loss.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.