Research map / Architectures and efficiency / Attention and transformers
Xformer: Hybrid X-Shaped Transformer for Image Denoising
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 Xformer, a dual-branch X-shaped Transformer for image denoising that pairs spatial-wise window attention with channel-wise cross-covariance attention, coupled by a Bidirectional Connection Unit.
Problem
Image denoising with Transformers typically captures only one type of global interaction (spatial-window or channel-wise), leaving patch-level and channel-level information poorly combined.
Approach
Two parallel U-shaped encoder-decoder branches, one using spatial-wise window Transformer blocks (STB) and one using channel-wise cross-covariance Transformer blocks (CTB), fused at multiple levels by a Bidirectional Connection Unit (BCU) that adds convolutionally refined features across branches.
Key results
- On Gaussian color denoising at sigma=50, Xformer reaches 30.38 dB on McMaster and 30.36 dB on Urban100, the best among compared methods (Restormer 30.02 dB on Urban100).
- On Urban100 grayscale sigma=50, Xformer obtains 28.71 dB, a 0.42 dB gain over Restormer (28.29 dB).
- On real denoising, Xformer reaches 40.19 dB PSNR on DND (best) and 39.98 dB on SIDD (comparable to Restormer's 40.02 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine STB and CTB in dual branches rather than use all STB (ablation) | Xformer · Urban100, Gaussian color σ=50 | PSNR (dB): 29.87 -> 29.94 | Helped |
| combine STB and CTB in dual branches rather than use all CTB (ablation) | Xformer · Urban100, Gaussian color σ=50 | PSNR (dB): 29.67 -> 29.94 | Helped |
| add complete BCU (ablation) | Xformer dual branches · Urban100, Gaussian color σ=50 | PSNR (dB): 29.82 -> 29.94 | Helped |
| use complete bidirectional BCU instead of single-direction BCU-1 (ablation) | Xformer dual branches · Urban100, Gaussian color σ=50 | PSNR (dB): 29.84 -> 29.94 | Helped |
| use complete bidirectional BCU instead of single-direction BCU-2 (ablation) | Xformer dual branches · Urban100, Gaussian color σ=50 | PSNR (dB): 29.92 -> 29.94 | Helped |
| use BCU-2 Conv direction instead of BCU-1 DWConv direction (ablation) | Xformer · Urban100, Gaussian color σ=50 | PSNR (dB): 29.84 -> 29.92 | Helped |
| add shifted-window strategy (ablation) | Xformer spatial-wise branch · Urban100, Gaussian color σ=50 | PSNR (dB): 29.88 -> 29.94 | Helped |
| use STB single branch instead of CTB single branch (ablation) | Xformer · Urban100, Gaussian color σ=50 | PSNR (dB): 29.68 -> 29.84 | Helped |
| add a CTB branch to the STB branch and use dual branches with BCU (ablation) | Xformer · Urban100, Gaussian color σ=50 | PSNR (dB): 29.84 -> 29.94 | Helped |
| use dual branches with BCU rather than dual branches without BCU (ablation) | Xformer · Urban100, Gaussian color σ=50 | PSNR (dB): 29.82 -> 29.94 | Helped |
Surprising findings
- A dual-branch model with plain concatenation fusion (no BCU) performed worse than a single STB branch, showing naive branch combination can hurt.
Limitations
- Gains over Restormer on SIDD are not present; Xformer is slightly below Restormer on SIDD PSNR.
- Model complexity is comparable to Restormer rather than substantially lower, so efficiency gains are modest.
Experimental setup
Datasets: DIV2K Flickr2K BSD500 WED SIDD
Benchmarks: Set12 BSD68 Urban100 CBSD68 Kodak24 McMaster SIDD DND
Metrics: PSNR SSIM FLOPs parameters
Baselines: BM3D DnCNN IRCNN FFDNet NLRN MWCNN RNAN RDN DRUNet SwinIR IPT Restormer
Builds on: SwinIR Restormer Uformer Conformer
Closest papers by meaning
- Attention-Only Transformers via Unrolled Subspace Denoising (ICML 2025 · similarity 0.87)
- AdaFormer: Efficient Transformer with Adaptive Token Sparsification for Image Super-resolution (AAAI 2024 · similarity 0.85)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.85)
- FrameQuant: Flexible Low-Bit Quantization for Transformers (ICML 2024 · similarity 0.84)
- U-DiTs: Downsample Tokens in U-Shaped Diffusion Transformers (NeurIPS 2024 · similarity 0.84)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (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.