Research map / 3D vision / Gaussian splatting
Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering
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.
Proposes multi-scale 3D Gaussian splatting that keeps Gaussians at several levels of detail and selects them by screen-space pixel coverage, reducing aliasing and speeding up low-resolution and distant rendering.
Problem
3D Gaussian splatting suffers aliasing and slow rendering when the scene is rendered at low resolution or from far away, because small Gaussians collapse into single pixels and per-pixel alpha blending over many Gaussians is slow.
Approach
Aggregate small Gaussians in voxels into coarser-level Gaussians during training, then at render time select only the Gaussians whose screen-space pixel coverage matches the current resolution, filtering out too-small and too-large ones instead of applying a per-Gaussian low-pass filter.
Key results
- Achieves 13%-66% PSNR and 160%-2400% rendering speed improvement at 4x-128x scale rendering on Mip-NeRF 360 versus 3D Gaussian splatting, with fewer than 5% extra Gaussians.
- At 64x downsampling on Mip-NeRF 360, PSNR reaches 25.35 versus 15.23 for 3D Gaussian splatting.
- Quality and speed at 1x resolution remain comparable to the original 3D Gaussian splatting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use multi-scale Gaussians with selective rendering | 3D Gaussian · 360 dataset, 4x | PSNR: 22.50 -> 24.82 | Helped |
| Use multi-scale Gaussians with selective rendering | 3D Gaussian · 360 dataset, 64x | Time: 103.3 -> 4.9 | Helped |
| Use multi-scale Gaussians with selective rendering | 3D Gaussian · Tank and Temples, 64x | PSNR: 13.88 -> 19.67 | Helped |
| Use multi-scale Gaussians with selective rendering | 3D Gaussian · Deep Blending, 64x | PSNR: 17.75 -> 25.70 | Helped |
| Add multi-scale training (ablation) | 3D Gaussian · 360 dataset, 4x | PSNR: 22.50 -> 23.50 | Helped |
| Add multi-scale training (ablation) | 3D Gaussian · 360 dataset, 4x | Time: 9.3 -> 7.7 | Hurt |
| Filter small Gaussians using pixel coverage (ablation) | 3D Gaussian with multi-scale training · 360 dataset, 64x | Time: 103.3 -> 4.6 | Helped |
| Filter small Gaussians using pixel coverage (ablation) | 3D Gaussian with multi-scale training · 360 dataset, 64x | PSNR: 15.23 -> 17.38 | Helped |
| Insert large Gaussians by aggregating small Gaussians (ablation) | 3D Gaussian with multi-scale training · 360 dataset, 64x | PSNR: 15.23 -> 21.53 | Helped |
| Insert large Gaussians by aggregating small Gaussians without filtering small Gaussians (ablation) | 3D Gaussian with multi-scale training · 360 dataset, 1x | PSNR: 27.52 -> 18.02 | Hurt |
| Use multi-scale training, small-Gaussian filtering, and large-Gaussian insertion together (ablation) | 3D Gaussian · Tank and Temples, 64x | PSNR: 13.88 -> 19.67 | Helped |
| Use multi-scale training, small-Gaussian filtering, and large-Gaussian insertion together (ablation) | 3D Gaussian · Deep Blending, 64x | Time: 59.7 -> 3.4 | Helped |
Surprising findings
- Naively inserting large aggregated Gaussians without small-Gaussian filtering drastically degrades quality (e.g., 18.02 PSNR at 1x on Mip-NeRF 360).
Limitations
- All Gaussians must still be splatted before pixel-coverage filtering, so rendering time does not decrease linearly with resolution.
- Filtering relies on post-splatting pixel coverage, adding overhead at very low resolution.
- Gains are reported on static benchmark scenes with downsampling simulation only.
Experimental setup
Datasets: Mip-NeRF 360 Tanks and Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks and Temples Deep Blending
Metrics: PSNR LPIPS rendering time (ms)
Baselines: 3D Gaussian Splatting Mip-NeRF 360 Zip-NeRF Tri-MipNeRF
Builds on: 3D Gaussian Splatting EWA Splatting Mip-NeRF mipmap
Closest papers by meaning
- 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)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting (NeurIPS 2024 · 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.