Research map / 3D vision / Gaussian splatting
TideGS: Scalable Training of Over One Billion 3D Gaussian Splatting Primitives via Out-of-Core Optimization
Chonghao Zhong, Linfeng Shi, Chen Hua, Tiecheng Sun, Hao Zhao, Binhang Yuan, Chaojian Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add trajectory-adaptive differential streaming (ablation) | TideGS · MatrixCity BigCity/Aerial (~102M Gaussians) | PCIe traffic (GB/iter): 0.85 -> 0.10 | Improved |
| add trajectory-adaptive differential streaming (ablation) | TideGS · MatrixCity BigCity/Aerial (~102M Gaussians) | iteration time (ms): 145.3 -> 90.7 | Improved |
| add asynchronous overlap of I/O and compute (ablation) | TideGS · MatrixCity BigCity/Aerial (~102M Gaussians) | iteration time (ms): 210.5 -> 90.7 | Improved |
| add Morton-ordered block layout (ablation) | TideGS · MatrixCity BigCity/Aerial (~102M Gaussians) | CPU cache hit rate (%): 42.1 -> 95.2 | Improved |
| add Morton-ordered block layout (ablation) | TideGS · MatrixCity BigCity/Aerial (~102M Gaussians) | PCIe traffic (GB/iter): 0.45 -> 0.10 | Improved |
| add differential streaming or overlap (ablation) | TideGS · MatrixCity BigCity/Aerial (~102M Gaussians) | PSNR: 24.79 / 24.86 vs 24.83 | No clear effect |
| use trajectory (TSP-ordered) camera sequence instead of shuffle (ablation) | TideGS · Mip-NeRF 360 in-memory | throughput (img/s): 3.24 -> 3.50 | Improved |
| scale Gaussians from ~102M to ~1.1B with out-of-core training | TideGS · MatrixCity | PSNR: 24.8 -> 26.1 | Improved |
| use block virtualization instead of native in-memory 3DGS | TideGS · Mip-NeRF 360 | PSNR: 29.0252 -> 28.9157 | No clear effect |
| replace CLM offloading with TideGS SSD-CPU-GPU hierarchy | CLM vs TideGS · MatrixCity ~1.1B | feasibility | Improved |
Closest papers by meaning
- On Scaling Up 3D Gaussian Splatting Training (ICLR 2025 · similarity 0.94)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
