Research map / 3D vision / Gaussian splatting
Large Images Are Gaussians: High-Quality Large Image Representation with Levels of 2D Gaussian Splatting
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 LIG, a 2D Gaussian Splatting image representation that directly optimizes covariance matrices and uses a two-level Level-of-Gaussian scheme to fit large images (histopathology, satellite, 2K natural images) with many Gaussians.
Problem
Fitting large images with 2D Gaussian Splatting is hard because optimization becomes difficult with a very large number of Gaussian points, and INR-based methods suffer from high training memory and slow decoding.
Approach
Optimize the 2D covariance matrix directly (3 parameters, custom CUDA kernels) with negative-sigma points filtered at render time, and fit large images in two levels: a coarse low-frequency Gaussian set on a downsampled image, then a high-frequency Gaussian set on the normalized residual.
Key results
- On STimage 9K, LIG reaches 37.47 dB PSNR at 3.5e7 Gaussians versus 29.86 dB for GaussianImage at the same count, with lower training memory (16.67 GB vs 20.47 GB).
- On FGF2 4K, LIG reaches 51.81 to 56.05 dB PSNR across 3.5e7 to 5.5e7 Gaussians, compared with about 27.5 dB for GaussianImage.
- On DIV-HR 2K, LIG reaches 44.89 to 52.22 dB PSNR, compared with 40.09 to 29.45 dB for GaussianImage as the Gaussian count grows.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LIG instead of GaussianImage at the listed large-image Gaussian settings | LIG · STimage (9K) | PSNR: 29.86 -> 42.19 | Helped |
| Use LIG instead of GaussianImage at the listed large-image Gaussian settings | LIG · FGF2 (4K) | PSNR: 27.53 -> 56.05 | Helped |
| Use LIG instead of GaussianImage at the listed large-image Gaussian settings | LIG · DIV-HR (2K) | PSNR: 40.09 -> 52.22 | Helped |
| Add the Level-of-Gaussian mechanism (ablation) | LIG · STimage (9K) | PSNR: 32.68 -> 44.49 | Helped |
| Add the Level-of-Gaussian mechanism (ablation) | LIG · FGF2 (4K) | PSNR: 51.35 -> 56.05 | Helped |
| Add the Level-of-Gaussian mechanism (ablation) | LIG · DIV-HR (2K) | PSNR: 42.30 -> 52.22 | Helped |
| Allocate some Gaussian points to L0 for low-frequency initialization (ablation) | LIG · STimage (9K) | PSNR: 34.12 -> 37.47; 33.81 -> 39.86; 33.22 -> 42.19 | Helped |
| Allocate some Gaussian points to L0 for low-frequency initialization (ablation) | LIG · STimage (9K) | FPS: 25.78 -> 20.19; 21.04 -> 17.89; 17.34 -> 15.72 | Hurt |
| Increase training iterations (ablation) | LIG · STimage (9K) and FGF2 (4K) | PSNR: 37.68 -> 40.76; 53.22 -> 57.20 | Helped |
| Use NeuRBF instead of LIG | LIG · FGF2 (4K) | PSNR: 41.31 -> 44.69 | Hurt |
Surprising findings
- GaussianImage's PSNR decreases as the Gaussian count increases on DIV-HR (40.09 dB at 5e5 down to 29.45 dB at 9e5), indicating optimization difficulty at scale.
- On the NeuRBF comparison on FGF2, NeuRBF achieves higher PSNR (44.69 vs 41.31 dB) at similar parameter count, though LIG has slightly lower FPS gap.
Limitations
- Gaussian-based representation lags behind NeuRBF in PSNR on FGF2 at similar parameter count
- Two-stage training adds sequential training and inference cost; level count fixed at 2
- Compression to match INR-level efficiency is not addressed; representation only, as acknowledged in conclusion
Experimental setup
Datasets: STimage FGF2 DIV-HR
Benchmarks: STimage (9K) FGF2 (4K) DIV-HR (2K)
Metrics: PSNR training memory (GB) FPS
Baselines: GaussianImage SIREN Gauss WIRE FINER NeuRBF
Builds on: 3D Gaussian Splatting GaussianImage gsplat Image-GS NeuRBF
Closest papers by meaning
- GaussianSR: High Fidelity 2D Gaussian Splatting for Arbitrary-Scale Image Super-Resolution (AAAI 2025 · similarity 0.91)
- Gaussian Graph Network: Learning Efficient and Generalizable Gaussian Representations from Multi-view Images (NeurIPS 2024 · similarity 0.90)
- Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 · similarity 0.90)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.90)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
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.