Research map / 3D vision / Gaussian splatting
MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference
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 MaterialRefGS, a 2D Gaussian Splatting method for reflective scenes that enforces multi-view consistent material maps, uses photometric-variation reflection priors for metallic supervision, and adds differentiable 2DGS ray tracing for indirect illumination.
Problem
Material/illumination decomposition for reflective scenes with Gaussian Splatting is under-constrained, causing illumination aliasing and poor generalization, and indirect reflections from inter-object occlusion are poorly modeled.
Approach
Enforce 2D Gaussians to yield multi-view consistent diffuse, roughness and metallic maps via patch homography warping, derive a multi-view reflection-strength prior from photometric variance across nearby views to supervise metallic, and model occlusion-aware indirect light with differentiable 2DGS ray tracing combined with a learnable environment map.
Key results
- On ShinyBlender, Ours reaches 35.57 PSNR, 0.976 SSIM, 0.049 LPIPS and 2.04 MAE, the best on all four metrics versus Ref-Gaussian (35.04 PSNR).
- On GlossySynthetic, Ours achieves 30.83 PSNR and 0.046 LPIPS, the best reported, versus Ref-Gaussian at 30.08 PSNR.
- On Ref-Real, Ours reaches 25.04 PSNR and 0.185 LPIPS, improving over EnvGS (24.85 PSNR, 0.215 LPIPS).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MaterialRefGS instead of Ref-Gaussian | MaterialRefGS · ShinyBlender | PSNR: 35.04 -> 35.57 | Helped |
| Use MaterialRefGS instead of Ref-Gaussian | MaterialRefGS · GlossySynthetic | PSNR: 30.08 -> 30.83 | Helped |
| Use MaterialRefGS instead of EnvGS | MaterialRefGS · Ref-Real | PSNR: 24.85 -> 25.04 | Helped |
| Use MaterialRefGS instead of EnvGS | MaterialRefGS · Mip-NeRF 360 | PSNR: 27.36 -> 27.06 | Hurt |
| Add multi-view consistency and reflection prior losses (ablation) | MaterialRefGS · ShinyBlender | PSNR: 34.87 -> 35.57 | Helped |
| Add multi-view consistency and reflection prior losses (ablation) | MaterialRefGS · Ref-Real | PSNR: 24.24 -> 25.04 | Helped |
| Add multi-view consistency loss without reflection prior loss (ablation) | MaterialRefGS · ShinyBlender | PSNR: 34.87 -> 35.21 | Helped |
| Add reflection prior loss after multi-view consistency loss (ablation) | MaterialRefGS · Ref-Real | PSNR: 24.47 -> 25.04 | Helped |
| Add normal prior (ablation) | MaterialRefGS · ShinyBlender | MAE: 2.59 -> 2.04 | Helped |
| Add normal prior, material regularization, and environment modeling (ablation) | MaterialRefGS · ShinyBlender | CD: 0.94 -> 0.60 | Helped |
| Add environment modeling with ray tracing (ablation) | MaterialRefGS · Ref-Real | PSNR: 24.76 -> 25.04 | Helped |
| Add environment modeling with ray tracing (ablation) | MaterialRefGS · ShinyBlender | LPIPS: 0.049 -> 0.049 | No clear effect |
Limitations
- Mip-NeRF 360 results are competitive but not best in PSNR/SSIM, with reduced gains on scenes with few reflective surfaces
- Metallic supervision relies on foreground masks and SAM2 region segmentation, which adds pipeline dependencies
Experimental setup
Datasets: ShinyBlender GlossySynthetic Ref-Real Mip-NeRF 360
Benchmarks: ShinyBlender GlossySynthetic Ref-Real Mip-NeRF 360
Metrics: PSNR SSIM LPIPS MAE Chamfer Distance
Baselines: Ref-NeRF ENVIDR 3DGS GaussianShader 2DGS 3DGS-DR Ref-Gaussian EnvGS
Builds on: 2D Gaussian Splatting 3D Gaussian Splatting Ref-Gaussian split-sum approximation Disney BRDF
Closest papers by meaning
- Reflective Gaussian Splatting (ICLR 2025 · similarity 0.94)
- Subsurface Scattering for Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- OMG: Opacity Matters in Material Modeling with Gaussian Splatting (ICLR 2025 · similarity 0.91)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Ref-GS: Directional Factorization for 2D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.89)
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.