Research map / Architectures and efficiency / Attention and transformers
Bayesian Self-Attention with Local Pixel Correlations for Lightweight Denoising Transformers
Runyang He, Zuowei Shen, Hui JI
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 overlapped attention window (ablation) | NLformer · SIDD | PSNR: 39.73 -> 39.88 | Improved |
| add first-order SA branch (ablation) | NLformer · SIDD | PSNR: 39.63 -> 39.88 | Improved |
| add second-order SA branch (ablation) | NLformer · SIDD | PSNR: 39.55 -> 39.88 | Improved |
| add second-order SA branch (DND) (ablation) | NLformer · DND | PSNR: 39.56 -> 39.92 | Improved |
| use WIFN instead of plain linear FFN (ablation) | NLformer · SIDD | PSNR: 39.59 -> 39.88 | Improved |
| add Woodbury inversion to FFN (ablation) | NLformer · SIDD | PSNR: 39.74 -> 39.88 | Improved |
| add Woodbury inversion to FFN (ablation) | NLformer · DND | PSNR: 39.69 -> 39.92 | Improved |
| raise low-rank correction rank from 0 to 2 (ablation) | NLformer · SIDD | PSNR: 39.67 -> 39.88 | Improved |
| raise rank from 2 to 3 (ablation) | NLformer · SIDD | PSNR: 39.88 -> 39.89 | No clear effect |
| raise rank from 2 to 3 (ablation) | NLformer · DND | PSNR: 39.92 -> 39.85 | Hurt |
| use learned token embedding instead of raw-pixel patches (ablation) | NLformer · SIDD | PSNR: 39.68 -> 39.88 | Improved |
| scale model from lightweight to large (ablation) | NLformer · SIDD | PSNR: 39.88 -> 40.11 | Improved |
Closest papers by meaning
- Attention-Only Transformers via Unrolled Subspace Denoising (ICML 2025 · similarity 0.88)
- Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising (NeurIPS 2024 · similarity 0.88)
- ScaleDiff: Higher-Resolution Image Synthesis via Efficient and Model-Agnostic Diffusion (NeurIPS 2025 · similarity 0.87)
- Xformer: Hybrid X-Shaped Transformer for Image Denoising (ICLR 2024 · similarity 0.87)
- kNN Attention Demystified: A Theoretical Exploration for Scalable Transformers (ICLR 2025 · similarity 0.86)
- Efficient Lightweight Image Denoising with Triple Attention Transformer (AAAI 2024 · similarity 0.86)
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.
