Research map / Image generation / Super-resolution
Progressive Focused Transformer for Single 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 Progressive Focused Transformer (PFT) with Progressive Focused Attention, which chains attention maps across layers via Hadamard products and sparsifies computation to enable larger windows for single image super-resolution.
Problem
Transformer-based single image super-resolution computes attention similarities with many irrelevant tokens, which hurts reconstruction and adds heavy computational cost.
Approach
Inherit attention maps from previous layers via row-normalized Hadamard products so consistently relevant tokens are amplified, and use the previous map as an index to skip similarity computations for irrelevant tokens, with a decaying top-k focus ratio across layers.
Key results
- PFT achieves 35.00 dB on Set14 x2 and 32.63 dB on Manga109... (Urban100 x4 32.63 dB vs ATD 32.62 dB) with 17.1% fewer FLOPs than ATD.
- PFT-light improves Urban100 x2 PSNR to 33.67 dB vs ATD-light 33.27 dB and reduces computation by 20.1%.
- Ablation on Urban100 x4: PFA reaches 26.62 dB at 50.9G FLOPs vs 26.26 dB for vanilla self-attention at 70.4G FLOPs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PFT instead of ATD | PFT · Set5 ×4 | PSNR: 33.10 -> 33.15 | Helped |
| Use PFT instead of ATD | PFT · Set14 ×4 | PSNR: 29.24 -> 29.29 | Helped |
| Use PFT instead of ATD | PFT · BSD100 ×4 | PSNR: 28.01 -> 28.02 | Helped |
| Use PFT instead of ATD | PFT · Urban100 ×4 | PSNR: 28.17 -> 28.20 | Helped |
| Use PFT instead of ATD | PFT · Manga109 ×4 | PSNR: 32.62 -> 32.63 | Helped |
| Add Progressive Attention (Hadamard inheritance) (ablation) | PFT-light · Set5 ×4 | PSNR: 32.28 -> 32.35 | Helped |
| Add Progressive Attention (Hadamard inheritance) (ablation) | PFT-light · Urban100 ×4 | PSNR: 26.26 -> 26.41 | Helped |
| Add Progressive Attention (Hadamard inheritance) (ablation) | PFT-light · Manga109 ×4 | PSNR: 30.62 -> 30.78 | Helped |
| Add focused sparse calculation to Progressive Attention (ablation) | PFT-light · Urban100 ×4 | PSNR: 26.41 -> 26.62 | Helped |
| Increase window size from 8 × 8 to 32 × 32 (ablation) | PFT-light · Urban100 ×4 | PSNR: 26.40 -> 26.81 | Helped |
| Set focus ratio to 0.1 rather than 0.5 (ablation) | PFT-light · Urban100 ×4 | PSNR: 26.62 -> 26.48 | Hurt |
| Set focus ratio to 0.9 rather than 0.5 (ablation) | PFT-light · Urban100 ×4 | PSNR: 26.62 -> 26.56 | Hurt |
Surprising findings
- Progressive Attention with a larger window (32x32) gives the best Urban100 result, showing that focused attention lets large windows be used at comparable cost.
Limitations
- Gains over ATD on classical SR are small (often within about 0.1 dB).
- Focus ratio and per-block top-k schedule are hand-chosen hyperparameters.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109
Metrics: PSNR SSIM FLOPs Params
Baselines: EDSR RCAN HAN IPT SwinIR CAT ART HAT IPG ATD CARN IMDN
Builds on: SwinIR HAT ATD top-k sparse attention LePE
Closest papers by meaning
- Recursive Generalization Transformer for Image Super-Resolution (ICLR 2024 · similarity 0.88)
- Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution (AAAI 2025 · similarity 0.86)
- CFAT: Unleashing Triangular Windows for Image Super-resolution (CVPR 2024 · similarity 0.86)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.85)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.85)
- Transformer-Based Selective Super-resolution for Efficient Image Refinement (AAAI 2024 · similarity 0.84)
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.