Research map / 3D vision / Gaussian splatting
Gaussian Graph Network: Learning Efficient and Generalizable Gaussian Representations from Multi-view Images
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 Gaussian Graph Network, which models multi-view pixel-aligned Gaussian groups as graph nodes and uses Gaussian-level graph linear and pooling layers to produce fewer, better feed-forward Gaussians.
Problem
Feed-forward generalizable 3D Gaussian Splatting methods combine pixel-aligned Gaussians from multiple views independently, producing redundant Gaussians, artifacts, and extra memory cost that grows with the number of input views.
Approach
Build a Gaussian Graph whose nodes are per-view pixel-aligned Gaussian groups, extend graph message passing to Gaussian level via a per-pixel relation matrix between views, and aggregate connected groups with a Gaussian pooling layer that merges duplicate Gaussians.
Key results
- On RealEstate10K with 4 input views, GGN reaches 24.76 PSNR with 102K Gaussians versus MVSplat's 20.86 PSNR with 262K Gaussians.
- On ACID with 16 input views, GGN reaches 27.69 PSNR and 150K Gaussians, while MVSplat drops to 17.64 PSNR with 1049K Gaussians.
- Cross-dataset generalization (train ACID, test RealEstate10K, 8 views) yields 24.75 PSNR for GGN versus 19.17 for MVSplat.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GGN instead of MVSplat† | Novel view synthesis · RealEstate10K, 4 views | PSNR: 21.48 -> 24.76 | Helped |
| Use GGN instead of MVSplat† | Novel view synthesis · RealEstate10K, 8 views | PSNR: 21.39 -> 25.15 | Helped |
| Use GGN instead of MVSplat† | Novel view synthesis · RealEstate10K, 16 views | PSNR: 21.34 -> 26.18 | Helped |
| Use GGN instead of MVSplat† | Novel view synthesis · ACID, 4 views | PSNR: 23.78 -> 26.46 | Helped |
| Use GGN instead of MVSplat† | Novel view synthesis · ACID, 8 views | PSNR: 23.72 -> 26.94 | Helped |
| Use GGN instead of MVSplat† | Novel view synthesis · ACID, 16 views | PSNR: 23.70 -> 27.69 | Helped |
| Use GGN instead of MVSplat | Cross-dataset novel view synthesis · ACID → RealEstate10K, 8 views | PSNR: 19.17 -> 24.75 | Helped |
| Use GGN instead of MVSplat | Cross-dataset novel view synthesis · RealEstate10K → ACID, 8 views | PSNR: 20.58 -> 25.21 | Helped |
| Add Gaussian Graph linear layer (ablation) | GGN · RealEstate10K, 8 views | PSNR: 24.72 -> 25.15 | Helped |
| Add Gaussian Graph pooling layer (ablation) | GGN · RealEstate10K, 8 views | PSNR: 20.25 -> 25.15 | Helped |
| Add Gaussian Graph pooling layer (ablation) | GGN · RealEstate10K, 8 views | Gaussians (K): 524 -> 126 | Helped |
| Replace vanilla model with full Gaussian Graph Network (ablation) | Novel view synthesis · RealEstate10K, 8 views | PSNR: 19.74 -> 25.15 | Helped |
Surprising findings
- Baseline pixelSplat and MVSplat get worse as input views increase, whereas GGN improves with more views.
- Removing the pooling layer raises Gaussian count from 126K to 524K and drops PSNR by about 4.9 dB on RealEstate10K.
Limitations
- Pixel-aligned predictions make representations sensitive to input resolution; 1024x1024 inputs yield over 1M Gaussians per view.
- Does not model generative completion of unseen scene parts.
- Focuses on color field and does not fully capture scene geometry.
Experimental setup
Datasets: RealEstate10K ACID
Benchmarks: RealEstate10K ACID
Metrics: PSNR SSIM LPIPS number of Gaussians FPS inference time
Baselines: pixelSplat MVSplat
Builds on: 3D Gaussian Splatting pixelSplat MVSplat
Closest papers by meaning
- Large Images Are Gaussians: High-Quality Large Image Representation with Levels of 2D Gaussian Splatting (AAAI 2025 · similarity 0.90)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.90)
- Graph-Guided Scene Reconstruction from Images with 3D Gaussian Splatting (ICLR 2025 · similarity 0.90)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.89)
- Structure Consistent Gaussian Splatting with Matching Prior for Few-shot Novel View Synthesis (NeurIPS 2024 · 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.