Content

TideGS: Scalable Training of Over One Billion
3D Gaussian Splatting Primitives
via Out-of-Core Optimization

ICML 2026

Spotlight
Chonghao Zhong1 Linfeng Shi1 Hua Chen2 Tiecheng Sun2 Hao Zhao3,4 Binhang Yuan* 1 Chaojian Li* 1
1Hong Kong University of Science and Technology 2Great Wall Motor 3Tsinghua University 4Beijing Academy of Artificial Intelligence
Paper Code Video Gallery

Project Video

Abstract

TideGS city-scale out-of-core training teaser
Figure 1. TideGS targets city-scale 3D Gaussian Splatting with a hybrid SSD-CPU-GPU pipeline, avoiding single-GPU memory failure while sidestepping the cost and complexity of distributed multi-GPU training.

City-scale 3D Gaussian Splatting pushes training beyond the memory budget of a single GPU: a full urban scene can span many square kilometers and require more than one billion Gaussians. Naively loading the entire scene into GPU memory leads to out-of-memory failure, while distributed multi-GPU training introduces substantial engineering complexity and deployment cost.

TideGS addresses this scaling bottleneck with a hybrid SSD-CPU-GPU out-of-core training pipeline. The system keeps only the active working set and resident blocks in GPU memory, uses CPU memory for caching and prefetch scheduling, and stores cold scene blocks and update logs on SSD. By overlapping host-to-device prefetch, GPU computation, and device-to-host writeback, TideGS streams scene content through a single-GPU training loop instead of requiring the whole scene to reside in VRAM.

This project page summarizes the method through the project video, MatrixCity and Google Earth Rome visualizations, and interactive comparisons against Vanilla 3DGS and CLM.

Visual Quality Comparison

Switch scenes and drag the split bar to compare TideGS against Vanilla 3DGS.

TideGS Vanilla 3DGS
Scene 1 focused comparison between Vanilla 3DGS and TideGS
Scene 1
Scene 2 focused comparison between Vanilla 3DGS and TideGS
Scene 2
Scene 3 focused comparison between Vanilla 3DGS and TideGS
Scene 3
Scene 4 focused comparison between Vanilla 3DGS and TideGS
Scene 4

Method Comparison Gallery

MatrixCity Dataset

Sample

Vanilla 3DGS

CLM

TideGS (ours)

Test View

Method Overview

TideGS method overview

BibTeX

@inproceedings{zhong2026tidegs,
  title={{TideGS}: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization},
  author={Zhong, Chonghao and Shi, Linfeng and Chen, Hua and Sun, Tiecheng and Zhao, Hao and Yuan, Binhang and Li, Chaojian},
  booktitle={International Conference on Machine Learning},
  year={2026},
  organization={PMLR}
}