Research map / 3D vision / Gaussian splatting
Lightweight Predictive 3D Gaussian Splats
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 Lightweight Predictive 3D Gaussian Splats, storing only a sparse set of parent Gaussians and predicting child Gaussians and attributes with a hash grid, tiny MLPs and attention, cutting storage while keeping quality and running on a mobile device.
Problem
3D Gaussian Splatting scenes need millions of Gaussians and gigabytes of storage, which hinders transmission and deployment on resource-constrained devices such as phones.
Approach
Store only a small set of parent points (position and scale) organized as depth-1 trees; children positions and all other attributes are predicted on the fly from a shared Instant-NGP hash grid, tiny 2-layer MLPs, and a modified self-attention over each tree, with adaptive tree manipulation that promotes children to parents.
Key results
- On Mip-NeRF 360, Ours-C3 reaches 29.11 PSNR at 35.60MB versus 3DGS 28.69 PSNR at 693MB (about 19.5x smaller).
- On Tanks&Temples, Ours-C1 uses 22.00MB versus LightGS 22.43MB, with PSNR 23.19 versus 22.83 for LightGS.
- Ours-C3 achieves 29.89 PSNR on Deep Blending at 35.40MB, versus 3DGS 29.41 PSNR at 676MB, and runs in real time on an iPhone 14.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the predictive Gaussian splat representation instead of 3DGS | Mip-NeRF 360° · Mip-NeRF 360° | PSNR: 28.69 -> 29.11 | Helped |
| Use the predictive Gaussian splat representation instead of 3DGS | Mip-NeRF 360° · Mip-NeRF 360° | Storage: 693MB -> 35.60MB | Helped |
| Use the predictive Gaussian splat representation instead of 3DGS | Tanks&Temples · Tanks&Temples | PSNR: 23.14 -> 23.82 | Helped |
| Use the predictive Gaussian splat representation instead of 3DGS | Tanks&Temples · Tanks&Temples | Storage: 411MB -> 35.32 MB | Helped |
| Use the predictive Gaussian splat representation instead of 3DGS | Deep Blending · Deep Blending | PSNR: 29.41 -> 29.89 | Helped |
| Use the predictive Gaussian splat representation instead of 3DGS | Deep Blending · Deep Blending | Storage: 676MB -> 35.40MB | Helped |
| Use frequency encoding instead of the hash grid (ablation) | Ours-C3 · Bicycle and Playroom | PSNR: 23.68 -> 19.72; 29.27 -> 23.36 | Hurt |
| Remove self-attention (equivalently, add self-attention) (ablation) | Ours-C3 · Bicycle and Playroom | PSNR: 23.68 -> 22.80; 29.27 -> 28.74 | Hurt |
| Remove Adaptive Tree Manipulation (equivalently, add ATM) (ablation) | Ours-C3 · Bicycle and Playroom | PSNR: 23.68 -> 22.66; 29.27 -> 28.65 | Hurt |
| Remove scene contraction (equivalently, add scene contraction) (ablation) | Ours-C3 · Bicycle | PSNR: 23.68 -> 22.88 | Hurt |
| Remove distance-to-AABB-center input for attribute prediction (ablation) | Ours-C3 · Bicycle and Playroom | PSNR: 23.68 -> 23.32; 29.27 -> 29.18 | Hurt |
| Remove 3D position input for attribute prediction (ablation) | Ours-C3 · Bicycle and Playroom | PSNR: 23.68 -> 9.72; 29.27 -> 6.19 | Hurt |
Surprising findings
- Ours-C1 has lower SSIM and higher LPIPS than LightGS and CompactGS on Mip-NeRF 360 despite being smaller, so the storage-quality trade-off is not uniformly better.
- Removing position input from attribute MLPs collapses PSNR (Bicycle 9.72, Playroom 6.19), showing position is critical.
Limitations
- Some configurations show lower SSIM and higher LPIPS than prior compact methods.
- ScaffoldGS achieves higher PSNR than the method on Tanks&Temples.
- Evaluation focuses on static scenes from three standard benchmarks.
Experimental setup
Datasets: Mip-NeRF 360 Tanks&Temples Deep Blending
Benchmarks: Mip-NeRF 360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS storage (MB) FPS on-device
Baselines: 3DGS ScaffoldGS LightGS CompactGS Mip-NeRF 360 iNGP Plenoxels
Builds on: 3D Gaussian Splatting Instant-NGP ScaffoldGS
Closest papers by meaning
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.93)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- DirectTriGS: Triplane-based Gaussian Splatting Field Representation for 3D Generation (CVPR 2025 · similarity 0.91)
- Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis (CVPR 2024 · similarity 0.91)
- LODGE: Level-of-Detail Large-Scale Gaussian Splatting with Efficient Rendering (NeurIPS 2025 · similarity 0.91)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · 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.