Research map / 3D vision / Gaussian splatting

Textured Gaussians for Enhanced 3D Scene Appearance Modeling

CVPR 2025 · 3D vision › Gaussian splatting · approach: Regularization and primitive design for 3DGS · 12 citations (OpenAlex)

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.

Augments each 3D Gaussian with an alpha, RGB, or RGBA texture map (custom CUDA ray-Gaussian intersection) so Gaussians can model spatially varying color and opacity, improving novel-view synthesis especially with few Gaussians.

Problem

3D Gaussian Splatting limits per-Gaussian expressivity: pixels covered by one Gaussian share a color (up to falloff) and each Gaussian is only a simple ellipsoid, so detailed appearance is hard to model.

Approach

Attach a per-Gaussian 2D texture map (alpha, RGB, or RGBA) on the Gaussian's local tangent plane, sampled via ray-plane intersection and UV mapping, and composite spatially varying color and alpha; alpha textures let Gaussians represent non-ellipsoidal shapes.

Key results

• With 1% of the default Gaussians, Textured Gaussians improve over 3DGS* by nearly 2 dB PSNR on scene-level datasets (e.g., Deep Blending 24.83 vs 23.97).
• Full RGBA textures give better PSNR than 3DGS* on most datasets at the default Gaussian count (e.g., Blender 33.24 vs 33.08).
• At equal model size, alpha-only textures outperform 3DGS* on Blender, Tanks and Temples, and Deep Blending (e.g., Deep Blending 28.36 vs 28.04 PSNR).

Measured effects reported in the paper

Surprising findings

• Alpha-only textures outperform RGB textures and perform comparably to RGBA textures while using one-third and one-fourth of the model size.
• Alpha-only textures let Gaussians represent complex shapes and high-frequency textures via spatially varying opacity, despite being simpler than RGBA.

Limitations

• 2D diffuse texture maps assume textures lie on a local surface and do not model spatially varying specular color.
• Texture representation limited to planar 2D maps; does not extend to volumetric textures or 5D radiance fields without future work.
• Baseline 3DGS* is the authors' own reimplementation, and Xu et al. results are partly from others' runs, complicating comparisons.

Experimental setup

Datasets: Blender Mip-NeRF 360 DTU Tanks and Temples Deep Blending

Benchmarks: Blender Mip-NeRF 360 DTU Tanks and Temples Deep Blending

Metrics: PSNR SSIM LPIPS

Baselines: 3DGS Xu et al. (texture-mapped Gaussians) Instant-NGP Mip-NeRF 360 GStex

Builds on: 3D Gaussian Splatting Gaussian Opacity Fields mesh texture mapping

Closest papers by meaning

• IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.95)
• GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
• UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping (CVPR 2025 · similarity 0.92)
• Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
• Spec-Gaussian: Anisotropic View-Dependent Appearance for 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
• Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.91)