Research map / 3D vision / Gaussian splatting
EVPGS: Enhanced View Prior Guidance for Splatting-based Extrapolated View Synthesis
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 EVPGS, a coarse-to-fine view augmentation framework that fine-tunes Gaussian Splatting models on DDM-regularized and occlusion-aware, refined view priors to improve extrapolated view synthesis, and releases the Merchandise3D dataset.
Problem
Gaussian Splatting methods trained on limited view coverage produce artifacts when rendering extrapolated (out-of-coverage) views, a setting underexplored relative to interpolated view synthesis.
Approach
Coarse stage regularizes renderings at augmented views with a frozen diffusion model's noise estimate plus mesh-rasterized depth supervision; fine stage builds occlusion-aware reprojected view priors (via reverse reprojection check) and refines them in the frequency domain with coarse-stage renderings to restore view-dependent color.
Key results
- EVPGS(RaDe-GS) reaches 26.488 PSNR / 0.8991 SSIM / 0.0670 LPIPS on DTU vs. 25.778 / 0.8882 / 0.0753 for the RaDe-GS baseline.
- On Merchandise3D, EVPGS(RaDe-GS) reaches 25.136 PSNR vs. 23.559 for RaDe-GS.
- On Synthetic-NeRF, EVPGS(RaDe-GS) reaches 27.849 PSNR vs. 27.531 for RaDe-GS, and all five EVPGS variants improve over their backbones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EVPGS to RaDe-GS | RaDe-GS · DTU | PSNR: 25.778 -> 26.488 | Helped |
| Use only the coarse stage rather than the baseline alone (ablation) | RaDe-GS · DTU | PSNR: 25.778 -> 26.219 | Helped |
| Use only the fine stage rather than the baseline alone (ablation) | RaDe-GS · DTU | PSNR: 25.778 -> 26.165 | Helped |
| Add occlusion-aware reprojection (ablation) | EVPGS(RaDe-GS) · DTU | PSNR: 26.027 -> 26.488 | Helped |
| Add view prior refinement (ablation) | EVPGS(RaDe-GS) · DTU | PSNR: 26.155 -> 26.488 | Helped |
| Add the initial reconstructed mesh (ablation) | EVPGS(RaDe-GS) · DTU | PSNR: 26.337 -> 26.488 | Helped |
| Add the initial reconstructed mesh (ablation) | EVPGS(RaDe-GS) · Synthetic-NeRF | PSNR: 27.829 -> 27.849 | No clear effect |
Limitations
- Does not explicitly model lighting, leaving color discrepancies from complex view-dependent effects.
- Relies on the quality of the reconstructed mesh; performance drops without it.
Experimental setup
Datasets: DTU Synthetic-NeRF Merchandise3D
Benchmarks: DTU Merchandise3D Synthetic-NeRF
Metrics: PSNR SSIM LPIPS
Baselines: 3DGS Mip-Splatting 2DGS GOF RaDe-GS VEGS
Builds on: 3D Gaussian Splatting VEGS Stable Diffusion Mip-Splatting RaDe-GS
Closest papers by meaning
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.91)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.90)
- 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors (NeurIPS 2024 · similarity 0.89)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Taming Video Diffusion Prior with Scene-Grounding Guidance for 3D Gaussian Splatting from Sparse Inputs (CVPR 2025 · similarity 0.89)
- SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (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.