Research map / Image generation / Super-resolution
Image Processing GNN: Breaking Rigidity in 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 Image Processing GNN (IPG), a graph-based super-resolution model that uses degree-flexible pixel-node graphs built from local and global sampling, achieving state-of-the-art SR results on standard benchmarks.
Problem
Convolution and window-attention super-resolution models are rigid: every pixel aggregates a fixed number of neighbors, ignoring that SR reconstruction effort is unbalanced toward detail-rich regions, and prior graph methods use patch nodes and uniform KNN degrees.
Approach
Assigns graph node degrees in proportion to a detail-rich indicator (absolute difference between features and their bilinear down-up sampled version), uses individual pixels as graph nodes, and builds local and global sampled graphs aggregated via edge-conditioned MetaFormer-style layers.
Key results
- IPG reaches 28.13 dB PSNR on Urban100 x4, exceeding prior SOTA by more than 0.1 dB under standard training.
- Pixel-node IPG beats 2x2 patch-node variant on Urban100 x4 (28.13 vs 27.81 dB).
- Degree-flexible graphs outperform plain KNN graphs on Urban100 x4 (28.13 vs 28.06 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IPG instead of HAT | ×4 SR models · Set5, Set14, B100, Urban100, and Manga109 | PSNR: 33.04 -> 33.15; 29.23 -> 29.24; 28.00 -> 27.99; 27.97 -> 28.13; 32.48 -> 32.53 | Mixed |
| use pixel nodes instead of 2 × 2 patch nodes (ablation) | IPG · Set5, Set14, and Urban100 ×4 SR | PSNR: 33.01 -> 33.15; 29.18 -> 29.24; 27.81 -> 28.13 | Helped |
| use degree-flexible graphs instead of fixed-degree KNN graphs (ablation) | IPG · Set5, Set14, and Urban100 ×4 SR | PSNR: 33.09 -> 33.15; 29.19 -> 29.24; 28.06 -> 28.13 | Helped |
| use both local and global graphs instead of either scale alone (ablation) | IPG · Set5, Set14, and Urban100 ×4 SR | PSNR: 33.08 -> 33.15; 33.04 -> 33.15; 29.19 -> 29.24; 29.11 -> 29.24; 28.10 -> 28.13; 27.59 -> 28.13 | Helped |
Surprising findings
- IPG runs slower than previous methods in practice despite competitive theoretical FLOPs, due to lack of hardware support for graph operations.
- Global-only graphs perform worse than local-only graphs on Urban100 x4 (27.59 vs 28.10 dB), while combining both gives the best result.
Limitations
- Wall-clock inference is slower than prior methods due to lack of hardware support for graph aggregation.
- Global sampling alone underperforms, and pixel-level graphs are expensive to search without sampling.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 BSDS100 Urban100 Manga109
Metrics: PSNR SSIM FLOPs parameters
Baselines: IGNN ELAN IPT SwinIR EDT CAT-A ART GRL-B HAT CARN IMDN LAPAR-A
Builds on: MetaFormer SwinIR HAT IGNN edge-conditioned aggregation
Closest papers by meaning
- Learning Efficient Image Super-Resolution Networks via Structure-Regularized Pruning (ICLR 2022 · similarity 0.89)
- FSR: A General Frequency-Oriented Framework to Accelerate Image Super-resolution Networks (AAAI 2023 · similarity 0.88)
- Lightweight Image Super-Resolution via Flexible Meta Pruning (ICML 2024 · similarity 0.88)
- Accurate Image Super-Resolution Using Very Deep Convolutional Networks (CVPR 2016 · similarity 0.88)
- Achieving Lightweight Super-Resolution for Real-Time Computer Graphics (AAAI 2025 · similarity 0.86)
- See More Details: Efficient Image Super-Resolution by Experts Mining (ICML 2024 · similarity 0.86)
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.