Research map / Image generation / Super-resolution
GaussianSR: High Fidelity 2D Gaussian Splatting for Arbitrary-Scale Image Super-Resolution
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 GaussianSR, which replaces discrete latent codes in INR-based arbitrary-scale super-resolution with 2D Gaussian splatting (selective per-pixel Gaussian kernels plus dual-stream channel decoupling), trained end-to-end on DIV2K.
Problem
Arbitrary-scale super-resolution (ASSR) with INR/LIIF-style methods is limited by discrete latent codes, causing checkerboard-like artifacts from abrupt latent selection and requiring costly weighted interpolation of neighboring codes.
Approach
Model each LR feature as a continuous 2D Gaussian field whose amplitudes are the LR features; stack and render these fields to refine and upsample features, and use a Gumbel-Softmax classifier to assign one of 100 learnable Gaussian kernels to each pixel (selective Gaussian splatting), with a dual-stream split so only 8 channels go through Gaussian splatting while the rest are bicubically upsample
Key results
- On integer-scale x4 SR, EDSR-baseline-GaussianSR reaches 26.19 dB on Urban100 vs 26.15 for LIIF and 26.14 for A-LIIF, with the largest gains on high-resolution datasets.
- On non-integer scales (x1.5, x2.4, x3.6), GaussianSR achieves the best PSNR on all five datasets, e.g. 36.27 dB on Urban100 at x1.5 vs 36.13 for LIIF.
- Latency at 128x128 target is 12.56 ms vs 14.52 ms for LIIF, but at 384x384 GaussianSR (106.72 ms) is slower than LIIF (86.43 ms).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace LIIF with GaussianSR using the same EDSR-baseline encoder | EDSR-baseline · General100 ×4 and ×3.6 | PSNR (dB): 31.53 -> 31.55; 32.37 -> 32.40 | Helped |
| Replace LIIF with GaussianSR using the same EDSR-baseline encoder | EDSR-baseline · BSD100 ×4 and ×3.6 | PSNR (dB): 27.60 -> 27.61; 28.11 -> 28.13 | Helped |
| Replace LIIF with GaussianSR using the same EDSR-baseline encoder | EDSR-baseline · Urban100 ×4 and ×3.6 | PSNR (dB): 26.15 -> 26.19; 26.89 -> 26.94 | Helped |
| Replace LIIF with GaussianSR using the same EDSR-baseline encoder | EDSR-baseline · Manga109 ×4 and ×3.6 | PSNR (dB): 30.54 -> 30.54; 31.59 -> 31.60 | Mixed |
| Replace LIIF with GaussianSR using the same EDSR-baseline encoder | EDSR-baseline · DIV2K100 ×4 and ×3.6 | PSNR (dB): 29.00 -> 29.03; 29.67 -> 29.70 | Helped |
| Add Gaussian splatting for 8 channels, with bicubic upsampling on the remaining 56, instead of bicubic upsampling on all 64 channels (ablation) | GaussianSR · DIV2K100 ×2, ×3, and ×4 | PSNR (dB): 34.4128 -> 34.7134; 30.8311 -> 31.0010; 28.8788 -> 29.0281 | Helped |
| Increase Gaussian-splatted channels from 8 to 16 (ablation) | GaussianSR · DIV2K100 ×2, ×3, and ×4 | PSNR (dB): 34.7134 -> 34.6990; 31.0010 -> 30.9853; 29.0281 -> 29.0157 | Hurt |
| Increase Gaussian bank from 100 to 400 classes (ablation) | GaussianSR · DIV2K100 ×2, ×3, and ×4 | PSNR (dB): 34.7134 -> 34.7041; 31.0010 -> 30.9914; 29.0281 -> 29.0174 | Hurt |
| Increase Gaussian bank from 100 to 900 classes (ablation) | GaussianSR · DIV2K100 ×2, ×3, and ×4 | PSNR (dB): 34.7134 -> 34.7072; 31.0010 -> 30.9906; 29.0281 -> 20.0229 | Hurt |
Surprising findings
- At 384x384 target resolution, GaussianSR is slower than LIIF (106.72 ms vs 86.43 ms), contrary to the efficiency claims.
- Ablation shows increasing the Gaussian bank from 100 to 400 or 900 classes does not help, and fewer than 20 Gaussians are frequently selected.
Limitations
- The five-dimensional rendering tensor per Gaussian field has substantial memory overhead.
- Gains over LIIF and A-LIIF are marginal on General100 and BSD100, and GaussianSR is slower than LIIF at 384x384 target resolution.
- Evaluation uses only the EDSR-baseline encoder and synthetic bicubic-downsampled LR inputs.
Experimental setup
Datasets: DIV2K General100 BSD100 Urban100 Manga109 Set5
Benchmarks: General100 BSD100 Urban100 Manga109 DIV2K100
Metrics: PSNR runtime (ms)
Baselines: MetaSR LIIF ITSRN A-LIIF DIINN Bicubic
Builds on: LIIF 3D Gaussian Splatting EDSR ITSRN
Closest papers by meaning
- Large Images Are Gaussians: High-Quality Large Image Representation with Levels of 2D Gaussian Splatting (AAAI 2025 · similarity 0.91)
- Splatter Image: Ultra-Fast Single-View 3D Reconstruction (CVPR 2024 · similarity 0.89)
- S2Gaussian: Sparse-View Super-Resolution 3D Gaussian Splatting (CVPR 2025 · similarity 0.88)
- Superpoint Gaussian Splatting for Real-Time High-Fidelity Dynamic Scene Reconstruction (ICML 2024 · similarity 0.88)
- Arbitrary-Scale Image Generation and Upsampling using Latent Diffusion Model and Implicit Neural Decoder (CVPR 2024 · similarity 0.87)
- GaussHDR: High Dynamic Range Gaussian Splatting via Learning Unified 3D and 2D Local Tone Mapping (CVPR 2025 · similarity 0.87)
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.