Research map / 3D vision / Gaussian splatting
UVGS: Reimagining Unstructured 3D Gaussian Splatting using UV Mapping
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.
Maps 3D Gaussian Splatting objects to structured 2D UV images via spherical mapping (14-channel UVGS), compresses them to a 3-channel Super UVGS with a multi-branch CNN, and uses pretrained image VAEs and latent diffusion models for 3DGS compression, generation and inpainting.
Problem
3D Gaussian Splatting is discrete, unstructured and permutation-invariant, so it is hard to feed into image-based autoencoders and generative diffusion models.
Approach
Spherical mapping turns each 3DGS object into an image-like UV map whose channels are Gaussian attributes; a multi-branch CNN compresses these 14 channels into a 3-channel Super UVGS image so that pretrained 2D image VAEs and diffusion models can be used with no fine-tuning.
Key results
- Super UVGS reconstructions from pretrained image AE reach 30.9 PSNR (cars) with 99.5% compression relative to fitted 3DGS, while Super UVGS alone compresses 89.7%.
- UVGS with K=4 layers reaches 33.2 PSNR on the full Objaverse set vs 34.2 for fitted 3DGS.
- Unconditional generation achieves FID 26.20 and KID 3.24, better than Get3D (53.17), EG3D (74.51) and GaussianCube (34.67) in Table 2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase UVGS layers from K=1 to K=4 (ablation) | UVGS · Objaverse 3DGS | PSNR: 31.1 -> 33.2 | Helped |
| increase UVGS layers from K=1 to K=4 (ablation) | UVGS · Objaverse 3DGS | LPIPS: 0.06 -> 0.03 | Helped |
| reduce UVGS resolution from 512 × 512 to 256 × 256 (ablation) | UVGS · Objaverse 3DGS | PSNR: 31.1 -> 28.2 | Hurt |
| reduce UVGS resolution from 512 × 512 to 256 × 256 (ablation) | UVGS · Objaverse 3DGS | LPIPS: 0.08 -> 0.23 | Hurt |
| add branching to the mapping network (ablation) | UVGS mapping network · Objaverse 3DGS | PSNR: 27.8 -> 31.1 | Helped |
| add branching to the mapping network (ablation) | UVGS mapping network · Objaverse 3DGS | LPIPS: 0.31 -> 0.08 | Helped |
| use UVGS for unconditional generation instead of GaussianCube | 3D object generation · Unconditional generation | FID: 34.67 -> 26.20 | Helped |
| use UVGS for unconditional generation instead of GaussianCube | 3D object generation · Unconditional generation | KID: 3.72 -> 3.24 | Helped |
| use UVGS for text-conditioned generation instead of LGM | 3D object generation · Text-conditioned generation | CLIP Score: 30.74 -> 32.62 | Helped |
Surprising findings
- Pretrained image VAEs generalize to Super UVGS images without any fine-tuning.
- Without branched mapping networks, Super UVGS reconstruction drops sharply (PSNR 27.8 vs 31.1).
Limitations
- Single-layer UVGS loses appearance detail and yields washed-out objects.
- Single-layer UV maps may not represent highly detailed or complex objects or scenes.
- Representation does not yet efficiently use empty UV pixels.
Experimental setup
Datasets: Objaverse ShapeNet cars
Benchmarks: Objaverse 3DGS dataset ShapeNet cars
Metrics: PSNR LPIPS SSIM FID KID CLIP Score compression %
Baselines: Get3D DiffTF EG3D GaussianCube DreamGaussian LGM Shap-E
Builds on: 3D Gaussian Splatting latent diffusion models LPIPS
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- DiffGS: Functional Gaussian Splatting Diffusion (NeurIPS 2024 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- Textured Gaussians for Enhanced 3D Scene Appearance Modeling (CVPR 2025 · similarity 0.92)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
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.