Research map / Image generation / Super-resolution
CFAT: Unleashing Triangular Windows 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 CFAT, a super-resolution transformer that alternates non-overlapping triangular and rectangular window self-attention with channel attention and overlapping cross-fusion attention, reporting gains over HAT on standard SR benchmarks.
Problem
Rectangular shifted-window self-attention in transformer SR models causes boundary-level distortion and offers a limited set of unique shifting modes.
Approach
Introduces a non-overlapping triangular window attention (each rectangular window split into four triangular windows) that works in series with rectangular window attention, giving more unique shift modes and reducing boundary distortion, combined with channel attention and overlapping cross-fusion attention.
Key results
- CFAT reaches 33.19 dB PSNR on Set5 x4 (DF2K training) and 28.17 dB on BSD100 x4, exceeding HAT's 33.04 dB and 28.00 dB.
- Ablation: combining rectangular and triangular windows yields 27.77 dB vs 27.61 dB (no RWAB) and 27.56 dB (no TWAB) on BSD100 x4.
- CFAT has 22.07M params and 90.59 GMac multi-adds, vs HAT's 20.8M and 103.7 GMac.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CFAT instead of HAT | CFAT · Set5 ×4 (DIV2K+Flickr2K) | PSNR: 33.04 -> 33.19 | Helped |
| use CFAT instead of HAT | CFAT · Set14 ×4 (DIV2K+Flickr2K) | PSNR: 29.23 -> 29.30 | Helped |
| use CFAT instead of HAT | CFAT · BSD100 ×4 (DIV2K+Flickr2K) | PSNR: 28.00 -> 28.17 | Helped |
| use CFAT instead of HAT | CFAT · Urban100 ×4 (DIV2K+Flickr2K) | PSNR: 27.97 -> 28.11 | Helped |
| use CFAT instead of HAT | CFAT · Manga109 ×4 (DIV2K+Flickr2K) | PSNR: 32.48 -> 32.63 | Helped |
| add CWAB (ablation) | CFAT · BSD100 ×4 | PSNR: 27.65dB -> 27.75dB | Helped |
| use depthwise-pointwise convolution in CWAB instead of standard convolution (ablation) | CFAT · BSD100 ×4 | PSNR: 27.75dB -> 27.77dB | Helped |
| add RWAB (ablation) | CFAT · BSD100 ×4 | PSNR: 27.61dB -> 27.77dB | Helped |
| add TWAB (ablation) | CFAT · BSD100 ×4 | PSNR: 27.56dB -> 27.77dB | Helped |
| use triangular window size 16*16 (L) instead of 8*8 (L) (ablation) | CFAT · BSD100 ×4 | PSNR: 27.58 -> 27.77 | Helped |
| use rectangular window size 16 (S) instead of 8 (S) (ablation) | CFAT · BSD100 ×4 | PSNR: 27.52 -> 27.77 | Helped |
| use variable shifts 0 →8 →16 →24 instead of a fixed shift size (ablation) | CFAT · BSD100 ×4 | PSNR: 27.58 -> 27.77 | Helped |
Limitations
- Gains over HAT are small on most benchmarks (e.g., about 0.1 dB on BSD100 x4 and Set14 x4), while the abstract claims a 0.7 dB improvement.
- Hyperparameters (window size, shift, alpha, beta, interval) were tuned on BSD100 x4 for 70 epochs, which is also a reported test set.
- Triangular window benefit is argued qualitatively; the complexity analysis is approximate.
Experimental setup
Datasets: DIV2K Flickr2K
Benchmarks: Set5 Set14 BSD100 Urban100 Manga109
Metrics: PSNR SSIM
Baselines: EDSR HAN SAN IPT SwinIR Swin2SR ACT ART EDT HAT
Builds on: SwinIR HAT ART Swin Transformer
Closest papers by meaning
- Recursive Generalization Transformer for Image Super-Resolution (ICLR 2024 · similarity 0.88)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.86)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.86)
- Progressive Focused Transformer for Single Image Super-Resolution (CVPR 2025 · similarity 0.86)
- Efficient Attention-Sharing Information Distillation Transformer for Lightweight Single Image Super-Resolution (AAAI 2025 · similarity 0.86)
- Transformer-Based Selective Super-resolution for Efficient Image Refinement (AAAI 2024 · 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.