Research map / 3D vision / Gaussian splatting
GaussianShader: 3D Gaussian Splatting with Shading Functions for Reflective Surfaces
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.
Adds a simplified shading function (diffuse, specular tint, GGX-style environment lighting, residual SH color) and a shortest-axis normal with depth-gradient consistency loss to 3D Gaussian Splatting to render reflective surfaces in real time.
Problem
3D Gaussian Splatting does not explicitly model appearance, so it fails on specular and reflective surfaces, while prior reflective-surface methods (Ref-NeRF, ENVIDR) are slow to train and render. Accurate normal estimation on discrete Gaussians is also hard.
Approach
Apply a simplified rendering-equation shading model on 3D Gaussians (diffuse color, specular tint times split-sum environment light, and an SH residual color term), with normals taken from each Gaussian's shortest axis plus a learned residual and a loss that ties predicted normals to depth-gradient normals.
Key results
- On Shiny Blender, average PSNR 31.94 vs 30.37 for 3D Gaussian Splatting (+1.57 dB), but below ENVIDR (32.88) and Ref-NeRF (32.32).
- On NeRF Synthetic, average PSNR 33.38 vs 33.30 for 3D Gaussian Splatting.
- Training takes about 0.58 h at 97 FPS, versus 23 h for Ref-NeRF and 6 h for ENVIDR, while 3D Gaussian Splatting runs at 274 FPS in 0.25 h.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GaussianShader shading and normal estimation to 3D Gaussian Splatting | 3D Gaussian Splatting · NeRF Synthetic, average | PSNR: 33.30 -> 33.38 | Helped |
| Add GaussianShader shading and normal estimation to 3D Gaussian Splatting | 3D Gaussian Splatting · Shiny Blender, average | PSNR: 30.37 -> 31.94 | Helped |
| Use GaussianShader instead of ENVIDR | Shiny Blender · Shiny Blender, average | PSNR: 32.88 -> 31.94 | Hurt |
| Add sparse loss (ablation) | GaussianShader · Shiny Blender, half resolution | PSNR: 31.79 -> 32.09 | Helped |
| Add normal-geometry consistency loss (ablation) | GaussianShader · Shiny Blender, half resolution | PSNR: 30.93 -> 32.09 | Helped |
| Add residual color for complex and indirect reflections (ablation) | GaussianShader · Shiny Blender, half resolution | PSNR: 31.49 -> 32.09 | Helped |
| Use the shortest-axis normal representation instead of directly optimizing a normal (ablation) | GaussianShader · Shiny Blender, half resolution | PSNR: 31.47 -> 32.09 | Helped |
| Use environment-map lighting instead of MLP lighting (ablation) | GaussianShader · Shiny Blender, half resolution | PSNR: 29.73 -> 32.09 | Helped |
| Add shading and normal attributes to Gaussian Splatting | 3D Gaussian Splatting · NeRF Synthetic and Shiny Blender, average | FPS: 274 -> 97 | Hurt |
Surprising findings
- On Shiny Blender, the method is still below ENVIDR and Ref-NeRF in average PSNR despite the reflective-surface focus.
- Rendering speed drops from 274 FPS (3D Gaussian Splatting) to 97 FPS, and training time roughly doubles relative to the baseline.
Limitations
- Slightly underperforms ENVIDR and Ref-NeRF on Shiny Blender in average PSNR and SSIM.
- Rendering is slower than vanilla 3D Gaussian Splatting due to the added shading attributes.
- Uses a single global environment light map per scene and a simplified shading approximation rather than full BRDF/indirect light modeling.
Experimental setup
Datasets: NeRF Synthetic Shiny Blender Glossy Synthetic Tanks and Temples
Benchmarks: NeRF Synthetic Shiny Blender Glossy Synthetic Tanks and Temples
Metrics: PSNR SSIM LPIPS training time FPS
Baselines: 3D Gaussian Splatting VolSDF Ref-NeRF ENVIDR NVDiffRec NVDiffRecMC NeRO NeRF
Builds on: 3D Gaussian Splatting Ref-NeRF NVDiffRec
Closest papers by meaning
- IRGS: Inter-Reflective Gaussian Splatting with 2D Gaussian Ray Tracing (CVPR 2025 · similarity 0.92)
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.89)
- Ref-GS: Directional Factorization for 2D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.88)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.87)
- DisC-GS: Discontinuity-aware Gaussian Splatting (NeurIPS 2024 · similarity 0.86)
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.