Research map / Image generation / Super-resolution
See More Details: Efficient Image Super-Resolution by Experts Mining
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 SeemoRe, an efficient single-image super-resolution network that stacks residual groups of a rank-modulating expert (mixture of low-rank experts with top-1 routing) and a spatial-modulating expert (striped large-kernel convolution), achieving SOTA efficiency-accuracy trade-offs on standard SR benchmarks.
Problem
Efficient image super-resolution: existing SR models that stack disparate operations (attention, feature mixing, global-local context) incur high computational cost, limiting practical deployment on resource-constrained devices.
Approach
Combine a mixture of low-rank experts (routed top-1 over different low-rank dimensions with depth-dependent selection) inside a rank-modulating block with a striped depth-wise large-kernel spatial enhancement expert, arranged as alternating residual groups to capture channel-wise and spatial dependencies cheaply.
Key results
- SeemoRe-T (220K params, 45 GMACS) reaches 38.06 dB on Set5 x2 and 32.22 dB on Urban100 x2, outperforming SAFMN by 0.41 dB on Urban100 x2.
- SeemoRe-L (931K params) matches or exceeds lightweight Transformers such as SRFormer-Light and DAT-Light on Manga109 x4 (31.48 dB) at 50 GMACS, versus SwinIR-Light's 64 GMACS.
- SeemoRe-T runs in 5.66 ms versus 11.20 ms for DDistill-SR on a 320x180 input at x4 on an RTX 4090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add both RME and SME to the baseline (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 31.61 -> 32.22; +0.49 dB | Helped |
| add SME (ablation) | SeemoRe-T baseline · Urban100 ×2 | PSNR: 31.61 -> 31.97; 0.32 dB | Helped |
| add RME (ablation) | SeemoRe-T baseline · Urban100 ×2 | PSNR: 31.61 -> 31.96; 0.26 dB | Helped |
| order macro blocks Rank then Spatial rather than Spatial then Rank (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.17 -> 32.22 | Helped |
| increase SEE kernel size from 3 to 11 (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.04 -> 32.22; 0.18 dB | Helped |
| use exponential rank growth rather than linear or 2*i+2 growth (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.12 -> 32.22 | Helped |
| increase MoRE experts from 1 to 3 (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.16 -> 32.22 | Helped |
| use 3 recursive steps instead of 2 (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.22 -> 31.04 | Hurt |
| use 1 recursive step instead of 2 (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.22 -> 32.10 | Hurt |
| increase top-k expert routing from 1 to 3 (ablation) | SeemoRe-T · Set5 ×2 | PSNR: 38.06 -> 38.10 | Helped |
| replace SEE with FusedMB (ablation) | SeemoRe-T · Urban100 ×2 | PSNR: 32.22 -> 32.18 | Hurt |
| replace SEE with C2F (ablation) | SeemoRe-T · Manga109 ×2 | PSNR: 39.01 -> 39.04 | Helped |
Surprising findings
- Top-1 routing is used in both training and inference, and increasing top-k to 2 or 3 gives only marginal PSNR gains at higher GMACs.
- Scaling to 8 experts increases parameters without consistently improving PSNR.
Limitations
- Limited to a small number of experts because feature dimensionality grows rapidly when adding experts.
- Still produces blur artifacts in challenging cases, similar to Transformer-based SR methods.
- Real-SR evaluation is limited to a single model (SeemoRe-T) compared against SAFMN.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109 DIV2K-I DIV2K-II
Metrics: PSNR SSIM NIQE BRISQUE GMACs runtime GPU memory
Baselines: CARN-M IMDN PAN DRSAN DDistill-SR ShuffleMixer SAFMN SwinIR-Light ELAN-Light SRFormer-Light ESRT DAT-Light
Builds on: Mixture-of-Experts routing GatedFFN ShuffleMixer Conv2Former
Closest papers by meaning
- Image Processing GNN: Breaking Rigidity in Super-Resolution (CVPR 2024 · similarity 0.86)
- Low-Res Leads the Way: Improving Generalization for Super-Resolution by Self-Supervised Learning (CVPR 2024 · similarity 0.86)
- Towards Realistic Data Generation for Real-World Super-Resolution (ICLR 2025 · similarity 0.86)
- NSARM: Next-Scale Autoregressive Modeling for Robust Real-World Image Super-Resolution (NeurIPS 2026 · similarity 0.85)
- Learning Coupled Dictionaries from Unpaired Data for Image Super-Resolution (CVPR 2024 · similarity 0.85)
- Augmenting Perceptual Super-Resolution via Image Quality Predictors (CVPR 2025 · similarity 0.85)
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.