Research map / 3D vision / Gaussian splatting
Ref-GS: Directional Factorization for 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.
Ref-GS applies deferred shading and a spherical Mip-grid with vector outer-product geometry-lighting factorization to 2D Gaussian splatting, improving view-dependent reflection rendering and normal/geometry recovery.
Problem
Gaussian splatting methods query view-dependent color per primitive, causing ambiguity between primitive orientation and viewing angle, which yields blurry reflections and noisy geometry for reflective and near-field-lit scenes.
Approach
Blend Gaussian attributes (diffuse color, features, roughness) first via alpha blending into a G-buffer, then apply view-dependent shading per pixel using a spherical Mip-grid over reflected direction and roughness, factorized with spatial features via an outer product fed to a small MLP.
Key results
- On Glossy Synthetic, Ref-GS reaches 30.59 PSNR and 2.21 MAE normal error on Shiny Blender, versus 29.78 PSNR and 2.43 MAE for 3DGS-DR.
- On Shiny Real, Ref-GS averages 24.20 PSNR in Gardenspheres and 24.44 overall, the best among compared methods.
- Ref-GS trains in 2.63x the 3DGS time and renders at 0.37x 3DGS speed, while achieving over 45 FPS at 800x800 in the abstract's claim.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ref-GS instead of 3DGS-DR | 2DGS · Shiny Blender (Avg.) | PSNR: 33.94 -> 34.80 | Helped |
| Use Ref-GS instead of 3DGS-DR | 2DGS · Shiny Real (Avg.) | PSNR: 23.80 -> 24.44 | Helped |
| Use Ref-GS instead of 3iGS | 2DGS · NeRF Synthetic | PSNR: 33.60 -> 33.20 | Hurt |
| Use Ref-GS instead of 3DGS-DR | 2DGS · Glossy Synthetic | PSNR: 29.78 -> 30.59 | Helped |
| Add Sph-Mip encoding (ablation) | Ref-GS · Shiny Blender | PSNR: 29.95 -> 34.00 | Helped |
| Use the multi-level spherical feature grid instead of a single-level feature map (ablation) | Ref-GS · Shiny Blender | PSNR: 30.12 -> 34.00 | Helped |
| Add deferred shading (ablation) | Ref-GS · Shiny Blender | PSNR: 31.79 -> 34.00 | Helped |
| Add directional factorization K ⊗ S (ablation) | Ref-GS · Shiny Blender | PSNR: 33.37 -> 34.00 | Helped |
Limitations
- Training and rendering are slower than vanilla 3DGS (training 2.63x, rendering 0.37x).
- Lower PSNR than 3DGS-DR on some real scenes such as Sedan and Toycar, and worse on NeRF Synthetic than 3iGS and 3DGS.
Experimental setup
Datasets: NeRF Synthetic Shiny Blender Glossy Synthetic Shiny Real Mip-NeRF 360 Tanks & Temples Glass & Ball
Benchmarks: NeRF Synthetic Shiny Blender Glossy Synthetic Shiny Real Mip-NeRF 360 Tanks & Temples Glass & Ball
Metrics: PSNR SSIM LPIPS MAE
Baselines: Ref-NeRF NeRO ENVIDR 3DGS GaussianShader 3iGS 3DGS-DR GS-ROR
Builds on: 2D Gaussian Splatting Ref-NeRF TensoRF 3DGS-DR deferred shading
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- GI-GS: Global Illumination Decomposition on Gaussian Splatting for Inverse Rendering (ICLR 2025 · similarity 0.91)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.91)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.91)
- Hardware-Rasterized Ray-Based Gaussian Splatting (CVPR 2025 · similarity 0.91)
- 6DGS: Enhanced Direction-Aware Gaussian Splatting for Volumetric Rendering (ICLR 2025 · 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.