Research map / 3D vision / Gaussian splatting
DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds
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.
Accelerates 3D Gaussian Splatting optimization by scheduling rendering resolution via a frequency-based schedule and coupling primitive growth to it, cutting training time about 45.7% across backbones without quality loss.
Problem
3DGS optimization time is dominated by rendering resolution and primitive count (optimization complexity); naive reductions in resolution or primitives hurt rendering quality due to aliasing or pruning.
Approach
Formulate 3DGS optimization as progressively fitting higher frequency components by increasing training resolution along a frequency-energy-based schedule, and co-schedule primitive growth with resolution via a concave-up curve and a momentum-based primitive budget that replaces dataset-prior caps.
Key results
- Plugging DashGaussian into various 3DGS backbones speeds optimization by 45.7% on average while keeping rendering quality (e.g., Taming-3DGS on Mip-NeRF 360: 5.51 to 3.23 min, PSNR 27.61 to 27.92).
- With Taming-3DGS as backbone, the full method reduces optimization time by 41.3% and uses fewer primitives (2.04M vs 2.28M) on Mip-NeRF 360.
- Combined with Taming-3DGS, optimization of a 3DGS model finishes within 200 seconds on a single GPU.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DashGaussian to Taming-3DGS | Taming-3DGS · Mip-NeRF 360 | Time (minutes): 5.51 -> 3.23 | Helped |
| Add DashGaussian to Taming-3DGS | Taming-3DGS · Deep Blending | PSNR: 29.69 -> 30.02 | Helped |
| Add DashGaussian to Taming-3DGS | Taming-3DGS · Tanks and Temples | Time (minutes): 4.02 -> 2.62 | Helped |
| Add DashGaussian | 3DGS · Mip-NeRF 360 | Time (minutes): 18.31 -> 10.16 | Helped |
| Add DashGaussian | Mip-Splatting · Deep Blending | Time (minutes): 23.57 -> 9.80 | Helped |
| Add resolution scheduler only (ablation) | Taming-3DGS · Mip-NeRF 360 | PSNR: 27.61 -> 27.99 | Helped |
| Add resolution scheduler only (ablation) | Taming-3DGS · Mip-NeRF 360 | Time (minutes): 5.51 -> 4.43 | Helped |
| Add primitive scheduler only (ablation) | Taming-3DGS · Mip-NeRF 360 | PSNR: 27.61 -> 27.85 | Helped |
| Add primitive scheduler only (ablation) | Taming-3DGS · Mip-NeRF 360 | Time (minutes): 5.51 -> 4.60 | Helped |
| Add both resolution and primitive schedulers (ablation) | Taming-3DGS · Mip-NeRF 360 | Time (minutes): 5.51 -> 3.23 | Helped |
Surprising findings
- Reducing the training resolution via frequency-based scheduling improved rendering quality over the backbone (e.g., LPIPS 0.210 to 0.196 for the resolution scheduler alone), not only speed.
Limitations
- Primitive-to-resolution relationship is heuristic and lacks deeper analysis, as the authors note.
- Hyperparameters (a, gamma, eta, power schedule) were set by default values and tuned in supplementary material.
Experimental setup
Datasets: Mip-NeRF 360 Deep Blending Tanks and Temples MatrixCity
Benchmarks: Mip-NeRF 360 Deep Blending Tanks and Temples MatrixCity
Metrics: PSNR SSIM LPIPS optimization time (minutes) number of Gaussian primitives
Baselines: 3DGS Reduced-3DGS Mini-Splatting Taming-3DGS Mip-Splatting Revising-3DGS
Builds on: 3D Gaussian Splatting Mip-Splatting Taming-3DGS Mini-Splatting Revising-3DGS
Closest papers by meaning
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.93)
- 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 0.92)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Hardware-Rasterized Ray-Based Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (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.