Research map / 3D vision / Gaussian splatting
BudSplat: Feed-forward 3D Gaussian Splatting under a Rendering Budget
MUYU XU, Fangneng Zhan, Yu Wei, Ling Shao, Shijian Lu
Read the paperCheck your idea against 137,217 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 BudSplat, a feed-forward 3D Gaussian Splatting approach that operates under a fixed rendering budget so it scales to dense multi-view inputs without exploding the number of Gaussians.
Problem
Feed-forward 3D Gaussian Splatting rendering efficiency degrades as inputs scale from sparse to dense multi-view images, because the number of candidate Gaussians grows rapidly.
Approach
Frames feed-forward 3DGS as operating under an explicit rendering budget, so the number of predicted Gaussians is controlled rather than growing with the number of input views. Specific mechanism not stated in the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add detail-token importance estimation (ablation) | BudSplat · RE10K low-overlap (24 views, same budget) | PSNR: 20.04 -> 23.02 | Improved |
| add importance-preserving voxel fusion (ablation) | BudSplat · RE10K low-overlap (24 views, same budget) | PSNR: 22.47 -> 23.02 | Improved |
| add quality-aware pruning score (ablation) | BudSplat · RE10K low-overlap (24 views, same budget) | PSNR: 22.93 -> 23.02 | Improved |
| add patch-level quota redistribution (ablation) | BudSplat · RE10K low-overlap (24 views, same budget) | PSNR: 22.55 -> 23.02 | Improved |
| add all proposed modules vs budgeted AnySplat-style baseline (ablation) | AnySplat-style pipeline · RE10K low-overlap (24 views, same budget) | LPIPS: 0.686 -> 0.201 | Improved |
| prune to 5% of fused Gaussians instead of 40% (ablation) | BudSplat · RE10K low-overlap (32 views) | PSNR: 23.69 -> 23.08 | Hurt |
| prune to 5% of fused Gaussians instead of 40% (ablation) | BudSplat · DL3DV low-overlap (32 views) | PSNR: 18.91 -> 18.08 | Hurt |
| BudSplat (5% budget) vs AnySplat Gaussian count | BudSplat · RE10K low-overlap (72 views) | Number of Gaussians: 3461K -> 173K | Improved |
| BudSplat vs WorldMirror on DL3DV at low overlap | BudSplat · DL3DV low-overlap (16 views) | PSNR: 18.28 -> 18.14 | Hurt |
| BudSplat vs posed VolSplat on ACID at 32 views | BudSplat · ACID low-overlap (32 views) | PSNR: 23.76 -> 23.42 | Hurt |
Closest papers by meaning
- 3D Gaussian Rendering Can Be Sparser: Efficient Rendering via Learned Fragment Pruning (NeurIPS 2024 · similarity 0.92)
- CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction (NeurIPS 2026 · similarity 0.91)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.91)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- DeSplat: Decomposed Gaussian Splatting for Distractor-Free Rendering (CVPR 2025 · similarity 0.91)
- CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting (NeurIPS 2026 · 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.
