Research map / Architectures and efficiency / Attention and transformers
Enhancing Image Restoration Transformer via Adaptive Translation Equivariance
JiaKui Hu, Zhengjian Yao, Lujia Jin, Hangzhou He, Yanye Lu
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 |
|---|---|---|---|
| add slide-indexed key-value attention (SkvSA) to SwinIR-Large (ablation) | SwinIR-Large · Urban100 x4 SR | PSNR: 27.94 -> 28.31 | Improved |
| add adaptive indices (ASkvSA) on top of SkvSA (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.31 -> 28.47 | Improved |
| add downsampled self-attention (DSA) to SkvSA (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.31 -> 28.49 | Improved |
| add both ASkvSA and DSA (full TEA) (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.31 -> 28.67 | Improved |
| replace window attention with TEA in SwinIR (translation equivariance) | SwinIR · Urban100 x4 SR | NTK condition number: 1746.49 -> 236.78 | Improved |
| replace window attention with TEA in SwinIR | SwinIR · Urban100 x4 SR | SRGA generalization value: 3.655 -> 3.275 | Improved |
| increase sliding window size w from 7 to 15 (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.35 -> 28.67 | Improved |
| increase sliding window size w from 15 to 31 (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.67 -> 28.73 | No clear effect |
| increase sliding stride s from 1 to 4 (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.43 -> 28.67 | Improved |
| increase adaptive index kernel size k from 1 to 3 (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.53 -> 28.67 | Improved |
| increase DSA token number Nd from 16 to N (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.67 -> 28.80 | No clear effect |
| use average pooling instead of max pooling for DSA downsampling (ablation) | TEAFormer · Urban100 x4 SR | PSNR: 28.61 -> 28.67 | Improved |
Closest papers by meaning
- Rethinking Expressivity and Degradation-Awareness in Attention for All-in-One Blind Image Restoration (ICLR 2026 · similarity 0.87)
- DSwinIR: Rethinking Window-Based Attention for Image Restoration (TPAMI 2025 · similarity 0.87)
- Sharing Key Semantics in Transformer Makes Efficient Image Restoration (NeurIPS 2024 · similarity 0.87)
- Intra and Inter Parser-Prompted Transformers for Effective Image Restoration (AAAI 2025 · similarity 0.85)
- Adapt or Perish: Adaptive Sparse Transformer with Attentive Feature Refinement for Image Restoration (CVPR 2024 · similarity 0.85)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · 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.
