Research map / 3D vision / Gaussian splatting
TexGaussian: Generating High-quality PBR Material via Octree-based 3D 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.
Proposes TexGaussian, an octree-based 3D U-Net that places one 3D Gaussian per finest octree leaf of an input mesh and regresses PBR albedo, roughness and metallic maps in a single feed-forward pass.
Problem
Generating PBR materials (albedo, roughness, metallic) for untextured 3D meshes is slow and inconsistent with geometry when methods rely on multi-view 2D diffusion models or point-cloud diffusion, which yield blurry, illumination-baked textures.
Approach
Octant-aligned 3D Gaussian Splatting: each Gaussian sits at the center of the finest octree leaf built from surface points, extended with roughness and metallic channels, and an octree-based 3D U-Net regresses these Gaussian parameters directly from 2D rendering supervision (MSE+LPIPS) rather than diffusion denoising.
Key results
- On ShapeNet, average FID 49.76 vs. 61.49 for Point-UV (2-stage) and 59.45 for TexOct, in about 11 s inference.
- On the Objaverse PBR test set, FID 123.72 (albedo) and 129.52 (PBR rendering) vs. 145.49 to 213.21 for DreamMat, FlashTex and Fantasia3D, running in about 21 s versus 20+ minutes for baselines.
- Octant-aligned Gaussians fit albedo with PSNR 32.60 using 79,480 Gaussians, versus PSNR 23.75 with 16.8M voxel-aligned Gaussians.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TexGaussian instead of TexOct | ShapeNet per-category texture generation · ShapeNet average | FID: 59.45 -> 49.76 | Helped |
| Use TexGaussian instead of TexOct | ShapeNet per-category texture generation · ShapeNet average | KID (×102): 2.60 -> 2.07 | Helped |
| Use TexGaussian instead of TexOct | ShapeNet per-category texture generation · ShapeNet average | inference time: 17.44s -> 11.02s | Helped |
| Use TexGaussian instead of Paint3D | ShapeNet 2D diffusion-based comparison · ShapeNet average | FID: 135.25 -> 100.97 | Helped |
| Use TexGaussian instead of Paint3D | ShapeNet 2D diffusion-based comparison · ShapeNet average | KID (×102): 4.67 -> 1.88 | Helped |
| Use TexGaussian instead of DreamMat | Text-conditioned PBR material generation · Objaverse test set (albedo) | FID: 152.63 -> 123.72 | Helped |
| Use TexGaussian instead of Fantasia3D | Text-conditioned PBR material generation · Objaverse test set (albedo) | KID (×102): 0.96 -> 0.20 | Helped |
| Use TexGaussian instead of DreamMat | Text-conditioned PBR material generation · Objaverse test set (PBR rendering) | FID: 145.49 -> 129.52 | Helped |
| Use TexGaussian instead of DreamMat | Text-conditioned PBR material generation · Objaverse test set (PBR rendering) | KID (×102): 0.19 -> 0.02 | Helped |
| Use octant-aligned rather than full voxel-aligned Gaussians (ablation) | 3D Gaussian fitting · Objaverse subset (albedo) | PSNR: 23.75 -> 32.60 | Helped |
| Use octant-aligned rather than full voxel-aligned Gaussians (ablation) | 3D Gaussian fitting · Objaverse subset (albedo) | LPIPS: 0.17 -> 0.039 | Helped |
| Use octant-aligned rather than full voxel-aligned Gaussians (ablation) | 3D Gaussian fitting · Objaverse subset (albedo) | SSIM: 0.92 -> 0.97 | Helped |
Limitations
- Generalization limited by the scale of the training set; struggles with extremely complex objects beyond training data.
- Trained only on a 29k-object Objaverse subset and per-category ShapeNet models; larger models and data are left to future work.
Experimental setup
Datasets: ShapeNet Objaverse Cap3D captions
Benchmarks: ShapeNet (bench, car, chair, table) Objaverse PBR test subset (200 objects)
Metrics: FID KID PSNR LPIPS SSIM inference time
Baselines: Point-UV TexOct TEXTure Text2Tex Paint3D TexPainter Fantasia3D FlashTex DreamMat Make-it-Real
Builds on: 3D Gaussian Splatting LGM octree-based convolutional neural networks (O-CNN) Point-UV TexOct
Closest papers by meaning
- Material Palette: Extraction of Materials from a Single Image (CVPR 2024 · similarity 0.89)
- MaterialRefGS: Reflective Gaussian Splatting with Multi-view Consistent Material Inference (NeurIPS 2025 · similarity 0.89)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.89)
- DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (ICLR 2024 · similarity 0.88)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.88)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.88)
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.