Research map / 3D vision / Gaussian splatting
Anti-Aliased 2D Gaussian Splatting
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 AA-2DGS, an anti-aliased 2D Gaussian Splatting that adds a world-space flat smoothing kernel and an object-space Mip filter, improving multi-scale novel view synthesis while keeping 2DGS mesh accuracy.
Problem
2D Gaussian Splatting produces severe aliasing when rendered at sampling rates different from training (zoom in/out), and its screen-space clamping heuristic worsens these artifacts.
Approach
Band-limit planar 2D Gaussians with a flat smoothing kernel set from the maximal training-view sampling rate, and derive an affine approximation of the ray-splat intersection so a Gaussian Mip filter can be applied directly in each splat's local space, replacing the clamping heuristic.
Key results
- Single-scale training (Mip-NeRF 360, 1/8 res) multi-scale testing: AA-2DGS average PSNR 27.08 vs 25.28 for 2DGS, and 26.66 for 2DGS w/o clamping.
- Single-scale training on Blender: AA-2DGS average PSNR 32.11 vs 24.47 for 2DGS across four scales.
- DTU Chamfer distance: AA-2DGS mean 0.74 vs 2DGS 0.80, comparable to GOF (0.74) and Neuralangelo (0.61 best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add world-space flat smoothing and object-space Mip filtering | 2DGS · Blender single-scale training, multi-scale testing | PSNR (Avg.): 24.47 -> 32.11 | Helped |
| add world-space flat smoothing and object-space Mip filtering | 2DGS · Mip-NeRF 360 single-scale training, multi-scale testing | PSNR (Avg.): 25.28 -> 27.08 | Helped |
| add world-space flat smoothing and object-space Mip filtering | 2DGS · Mip-NeRF 360 same-scale training and testing | PSNR: 27.56 -> 27.38 | Hurt |
| add world-space flat smoothing and object-space Mip filtering | 2DGS · DTU | Chamfer distance (Mean): 0.80 -> 0.74 | Helped |
| add world-space flat smoothing filter (ablation) | AA-2DGS · Mip-NeRF 360 single-scale training, multi-scale testing | PSNR (Avg.): 26.69 -> 27.08 | Helped |
| add object-space Mip filter (ablation) | AA-2DGS · Blender single-scale training, multi-scale testing | PSNR (Avg.): 30.76 -> 32.11 | Helped |
| add world-space flat smoothing filter (ablation) | AA-2DGS · Blender single-scale training, multi-scale testing | PSNR (Avg.): 32.06 -> 32.11 | No clear effect |
| remove screen-space clamping (ablation) | 2DGS · Blender single-scale training, multi-scale testing | PSNR (Avg.): 24.47 -> 30.54 | Helped |
| remove screen-space clamping (ablation) | 2DGS · Mip-NeRF 360 single-scale training, multi-scale testing | PSNR (Avg.): 25.28 -> 26.66 | Helped |
| remove screen-space clamping (ablation) | 2DGS · Mip-NeRF 360 same-scale training and testing | PSNR: 27.56 -> 27.29 | Hurt |
Surprising findings
- Removing the screen-space clamping from 2DGS alone substantially improves multi-scale performance, indicating the clamping heuristic itself worsens aliasing.
- AA-2DGS slightly lowers single-scale PSNR on Mip-NeRF 360 (27.38 vs 27.56 for 2DGS) because band-limiting attenuates high frequencies.
Limitations
- Needle-like artifacts persist for planar primitives at grazing angles and extreme magnification because of zero thickness along the normal.
- Fixed filter parameters (sigma=0.1, s=0.2) may over-smooth details in some scenes.
- Slight reduction in single-scale PSNR due to band-limiting trade-off.
Experimental setup
Datasets: Blender Mip-NeRF 360 DTU
Benchmarks: Blender Mip-NeRF 360 DTU
Metrics: PSNR SSIM LPIPS Chamfer distance
Baselines: 2DGS 2DGS w/o Clamping Mip-Splatting 3DGS 3DGS + EWA Mip-NeRF Tri-MipRF Zip-NeRF NeuralAngelo GOF GaussianSurfels SuGaR
Builds on: 2D Gaussian Splatting Mip-Splatting EWA splatting 3D Gaussian Splatting
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.92)
- Mip-Splatting: Alias-free 3D Gaussian Splatting (CVPR 2024 · similarity 0.92)
- Hardware-Rasterized Ray-Based Gaussian Splatting (CVPR 2025 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.91)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (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.