Research map / Image generation / Super-resolution
Recursive Generalization Transformer for 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 the Recursive Generalization Transformer (RGT) for image super-resolution, with recursive-generalization self-attention for linear-complexity global attention and hybrid adaptive integration to fuse local and global block features.
Problem
Transformer-based image super-resolution relies on local self-attention, which limits global context modeling, while global self-attention is quadratic in image size and impractical for high-resolution inputs.
Approach
RG-SA recursively compresses input features into small fixed-size representative maps and applies cross-attention from the full-resolution input to them, with scaled query/key/value channels; HAI adds a learnable per-block adaptor on the block input to fuse features across alternating local and global attention blocks.
Key results
- RGT outperforms CAT-A by 0.21 dB on Urban100 at x2 (34.47 vs 34.26 for RGT vs CAT-A reported in table, with RGT+ at 34.63).
- RGT at x4 reaches 27.98 dB on Urban100 and 32.50 dB on Manga109 with 13.37M params and 251.07G FLOPs, lower than CAT-A's 16.60M and 360.67G.
- Removing HAI and using a vanilla skip connection drops Urban100 x2 PSNR from 33.68 to 32.71 dB in the ablation, while HAI adds 0.16 dB over no HAI.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RGT instead of CAT-A | Image super-resolution · Urban100 ×2 | PSNR: +0.21 dB | Helped |
| Add RG-SA to the L-SA baseline (ablation) | RGT-S · Urban100 ×2 | PSNR: 33.43 -> 33.52 | Helped |
| Add HAI to L-SA and RG-SA (ablation) | RGT-S · Urban100 ×2 | PSNR: 33.52 -> 33.68 | Helped |
| Remove recursive aggregation from RGM (ablation) | RGT-S · Urban100 ×2 | PSNR: 33.68 -> 33.57 | Hurt |
| Use channel scaling with c_r=0.5 rather than c_r=1 (ablation) | RGT-S · Urban100 ×2 | PSNR: 33.54 -> 33.68; 0.14 dB | Helped |
| Replace HAI with a vanilla skip connection (ablation) | RGT-S · Urban100 ×2 | PSNR: 33.52 -> 32.71; 0.81 dB | Hurt |
| Add HAI to an L-SA-only model (ablation) | L-SA-only model · Urban100 ×2 | PSNR: 33.43 -> 33.44 | No clear effect |
Surprising findings
- A vanilla skip connection around each Transformer block degrades PSNR by 0.81 dB, while the learnable adaptor in HAI recovers and improves performance.
Limitations
- Local self-attention and RG-SA are used in alternation, so the global context still depends on stacked blocks and the RGM may lose local detail, as the authors note.
- Reported gains over the strongest baseline are modest (e.g., about 0.1 dB on several benchmarks).
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 B100 Urban100 Manga109
Metrics: PSNR SSIM Params FLOPs
Baselines: EDSR RCAN SRFBN SAN HAN CSNLN NLSA CRAN DFSA ELAN SwinIR CAT-A
Builds on: SwinIR CAT Swin Transformer EDSR
Closest papers by meaning
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.90)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.88)
- Progressive Focused Transformer for Single Image Super-Resolution (CVPR 2025 · similarity 0.88)
- Transformer-Based Selective Super-resolution for Efficient Image Refinement (AAAI 2024 · similarity 0.88)
- CFAT: Unleashing Triangular Windows for Image Super-resolution (CVPR 2024 · similarity 0.88)
- Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution (AAAI 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.