Research map / Image generation / Super-resolution
SRFormerV2: Taking a Closer Look at Permuted Self-Attention for Image Super-Resolution
Yupeng Zhou, Zhen Li, Chunle Guo, Li Liu, Ming‐Ming Cheng, Qibin Hou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enlarge attention window from 8x8 to 16x16 (ablation) | SwinIR · Urban100 x2 SR | PSNR: 33.09 -> 33.40 | Improved |
| enlarge PSA window from 12x12 to 24x24 (ablation) | SRFormer · Urban100 x2 SR | PSNR: 33.08 -> 33.51 | Improved |
| add ConvFFN (5x5 depthwise conv) to FFN (ablation) | SRFormer 24x24 · Urban100 x2 SR | PSNR: 33.38 -> 33.51 | Improved |
| use 5x5 rather than 3x3 depthwise conv in ConvFFN (ablation) | SRFormer · Manga109 x2 SR | PSNR: 39.34 -> 39.49 | Improved |
| 3x3 depthwise conv in ConvFFN versus none (ablation) | SRFormer · Manga109 x2 SR | PSNR: 39.44 -> 39.34 | Hurt |
| use PSA instead of token reduction at S=24 (ablation) | large-window self-attention · Urban100 x2 SR | PSNR: 33.24 -> 33.38 | Improved |
| use PSA instead of random token sampling at S=24 (ablation) | large-window self-attention · Urban100 x2 SR | PSNR: 32.34 -> 33.38 | Improved |
| raise token reduction factor r from 2 to 3 in PSA (S=24) (ablation) | SRFormer PSA · Urban100 x2 SR | PSNR: 33.38 -> 33.09 | Hurt |
| scale attention window from 24x24 to 40x40 (ablation) | SRFormer (DF2K) · Urban100 x2 SR | PSNR: 34.09 -> 34.52 | Improved |
| scale attention window from 24x24 to 40x40 (ablation) | SRFormer (DF2K) · Set5 x2 SR | PSNR: 38.58 -> 38.58 | No clear effect |
| replace window attention in HAT with PSA (36x36 window) | HAT · Urban100 x2 SR | PSNR: 34.45 -> 34.57 | Improved |
| scale SRFormer into SRFormerV2 (larger window, channels, mixed small/large windows) | SRFormer · Urban100 x2 SR (DF2K-trained) | PSNR: 34.09 -> 34.59 | Improved |
Closest papers by meaning
- SRFormer: Permuted Self-Attention for Single Image Super-Resolution (ICCV 2023 · similarity 0.98)
- Recursive Generalization Transformer for Image Super-Resolution (ICLR 2024 · similarity 0.89)
- Transcending the Limit of Local Window: Advanced Super-Resolution Transformer with Adaptive Token Dictionary (CVPR 2024 · similarity 0.88)
- CFAT: Unleashing Triangular Windows for Image Super-resolution (CVPR 2024 · similarity 0.88)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 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.
