Research map / 3D vision / Gaussian splatting
On Scaling Up 3D Gaussian Splatting Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Grendel is a distributed multi-GPU 3D Gaussian Splatting training system that partitions Gaussians and pixels, uses sparse all-to-all communication and dynamic load balancing, and adds batched training with a sqrt(batch size) learning-rate scaling rule.
Problem
3DGS training runs on a single GPU, which caps the number of Gaussians and the image resolution for large-scale, high-resolution scenes due to memory and compute limits.
Approach
Mixed parallelism: distribute Gaussians across GPUs for the transformation step and contiguous pixel blocks for rendering/loss, linked by sparse all-to-all transfer of only the Gaussians that intersect each pixel partition, with iterative rebalancing from prior-iteration render times, plus batched training with Adam hyperparameter scaling derived under an independent-gradients hypothesis (lr scale
Key results
- On 4K Rubble, 40.4M Gaussians across 16 GPUs reach test PSNR 27.28 versus 26.28 with 11.2M Gaussians on one GPU.
- Throughput on Rubble rises from 5.55 images/s (4 GPUs, batch 1) to 38.03 images/s (32 GPUs, batch 64).
- Versus CityGaussian Official, Grendel (200K images) achieves 3x-6.7x training speedups with comparable or higher PSNR on Rubble, Building, and MatrixCity.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CityGS Official with Grendel using the same 200K-image training count | Grendel · Rubble (4x downsampled) | PSNR: 25.88 -> 27.39 | Helped |
| replace CityGS Official with Grendel using the same 200K-image training count | Grendel · Building (4x downsampled) | PSNR: 22.14 -> 22.69 | Helped |
| replace CityGS Official with Grendel using the same 200K-image training count | Grendel · MatrixCity Block_All (1600px width) | PSNR: 27.41 -> 27.33 | Hurt |
| use Grendel instead of CityGS Official | Grendel · Rubble, Building, and MatrixCity Block_All | total training time: 2.88 hrs -> 0.85 hrs; 4.57 hrs -> 0.90 hrs; 8.25 hrs -> 1.22 hrs | Helped |
| increase from 4 GPUs and batch size 1 to 32 GPUs and batch size 64 | Grendel · Rubble | throughput: 5.55 -> 38.03 | Helped |
| scale training from 1 GPU and batch size 1 to 16 GPUs and batch size 32 | Grendel · Train | throughput: 34.72 -> 185.19 | Helped |
| scale training from 1 GPU and batch size 1 to 16 GPUs and batch size 32 | Grendel · Train | PSNR: 21.84 -> 21.76 | No clear effect |
| increase the number of Gaussians through densification (ablation) | Grendel · Rubble | PSNR: 24.84 -> 25.85 -> 26.14 -> 26.28 -> 26.61 -> 26.91 -> 27.06 -> 27.28 | Helped |
| add iterative load balancing (ablation) | Grendel · Mip-NeRF360 and Rubble | throughput | Helped |
| increase batch size beyond one (ablation) | Grendel · Train, Mip-NeRF360, Tanks & Temples, and DeepBlending | throughput | Helped |
| use square-root learning-rate scaling instead of constant or linear scaling (ablation) | Grendel Adam optimizer · Rubble | update cosine similarity to batch-size-1 updates | Helped |
| use exponential momentum scaling instead of leaving momentum coefficients unchanged (ablation) | Grendel Adam optimizer · Rubble | update cosine similarity to batch-size-1 updates | Helped |
Surprising findings
- Balancing pixel workloads does not balance the number of Gaussians touched per GPU, so random Gaussian redistribution trained fastest in experiments despite higher communication volume.
Limitations
- Gradient independence assumption is heuristic and holds only approximately in practice.
- Scaling experiments rely on densification tuning to reach target Gaussian counts, and rely on NVLink/Slingshot-connected A100 clusters.
Experimental setup
Datasets: Rubble MatrixCity Block_All Mip-NeRF 360 Tanks & Temples Deep Blending Bicycle Train
Benchmarks: Rubble MatrixCity Block_All Mip-NeRF 360 Tanks & Temples Deep Blending Bicycle Train
Metrics: PSNR SSIM LPIPS throughput (images/s) training time
Baselines: CityGaussian 3DGS single-GPU RetinaGS DOGS
Builds on: 3D Gaussian Splatting Adam FSDP CityGaussian
Closest papers by meaning
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.92)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (NeurIPS 2024 · similarity 0.91)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 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.