Research map / 3D vision / Gaussian splatting
Sort-free Gaussian Splatting via Weighted Sum 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 Weighted Sum Rendering (GS-WSR), a sort-free Gaussian Splatting method that replaces alpha blending with learned commutative weighted sums plus view-dependent opacity, and runs faster on a mobile GPU.
Problem
3DGS rendering depends on costly, view-dependent depth sorting before non-commutative alpha blending, which hurts speed on mobile GPUs and causes popping artifacts.
Approach
Replace sorted alpha-compositing with a commutative weighted sum where each Gaussian's weight is a learned function of depth (DIR, EXP, or linear-correction variants), plus view-dependent opacity from spherical harmonics, so rendering order does not matter.
Key results
- LC-WSR averages 27.19 PSNR on Mip-NeRF360 (vs 3DGS 27.21) and 23.61 on Tanks & Temples (vs 23.14), 29.63 on Deep Blending (vs 29.41).
- On a Snapdragon 8 Gen 3 GPU at 1920x1080, the method is on average 1.23x faster in total time than the Vulkan 3DGS-Graphics implementation when mobile resources are not exhausted.
- Runtime memory is about 63% of 3DGS-Graphics memory.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace DIR-WSR's constant weights with LC-WSR's linear depth correction (ablation) | Weighted Sum Rendering · Mip-NeRF 360 | PSNR: 25.99 -> 27.19 | Helped |
| Replace DIR-WSR's constant weights with LC-WSR's linear depth correction (ablation) | Weighted Sum Rendering · Tanks & Temples | PSNR: 22.80 -> 23.61 | Helped |
| Replace DIR-WSR's constant weights with LC-WSR's linear depth correction (ablation) | Weighted Sum Rendering · Deep Blending | PSNR: 28.85 -> 29.63 | Helped |
| Replace EXP-WSR's exponential depth weights with LC-WSR's linear correction (ablation) | Weighted Sum Rendering · Mip-NeRF 360 | PSNR: 26.97 -> 27.19 | Helped |
| Replace EXP-WSR's exponential depth weights with LC-WSR's linear correction (ablation) | Weighted Sum Rendering · Tanks & Temples | PSNR: 23.32 -> 23.61 | Helped |
| Replace EXP-WSR's exponential depth weights with LC-WSR's linear correction (ablation) | Weighted Sum Rendering · Deep Blending | PSNR: 29.77 -> 29.63 | Hurt |
| Add learnable rendering parameters (ablation) | LC-WSR · Mip-NeRF360 | PSNR: 23.19 -> 27.19 | Helped |
| Add learnable rendering parameters (ablation) | LC-WSR · Tanks & Temples | PSNR: 21.55 -> 23.61 | Helped |
| Add learnable rendering parameters (ablation) | LC-WSR · Deep Blending | PSNR: 27.92 -> 29.63 | Helped |
| Add view-dependent opacity (ablation) | LC-WSR · Mip-NeRF360 | PSNR: 25.88 -> 27.19 | Helped |
| Add view-dependent opacity (ablation) | LC-WSR · Tanks & Temples | PSNR: 21.83 -> 23.61 | Helped |
| Add view-dependent opacity (ablation) | LC-WSR · Deep Blending | PSNR: 29.27 -> 29.63 | Helped |
Surprising findings
- Linear-correction weights produce better PSNR than the exponential OIT-style weights for Gaussian splatting.
- View-dependent opacity contributes substantially: removing it drops Mip-NeRF360 PSNR from 27.19 to 25.88.
Limitations
- No early termination possible in the fragment-shader-based implementation, so it cannot skip occluded Gaussians.
- Dark objects in front of bright backgrounds appear apparently transparent (failure case).
- Speedup is not realized on scenes where the 3DGS-Graphics baseline does not exhaust mobile resources; kitchen and train show little or negative gain.
Experimental setup
Datasets: Mip-NeRF360 Tanks & Temples Deep Blending
Benchmarks: Mip-NeRF360 Tanks & Temples Deep Blending
Metrics: PSNR SSIM LPIPS runtime (ms) runtime memory (MB)
Baselines: 3DGS Plenoxels INGP-Base INGP-Big Mip-NeRF360 Compact 3DGS Compressed 3DGS LightGaussian 3DGS-Compute 3DGS-Graphics
Builds on: 3D Gaussian Splatting Order Independent Transparency (weighted blended OIT) KPConv
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 0.91)
- DashGaussian: Optimizing 3D Gaussian Splatting in 200 Seconds (CVPR 2025 · similarity 0.91)
- Steepest Descent Density Control for Compact 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Multi-Scale 3D Gaussian Splatting for Anti-Aliased Rendering (CVPR 2024 · similarity 0.90)
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.