Research map / Image generation / Image restoration
MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration
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 MaIR, a Mamba image restoration model with a Nested S-shaped Scanning strategy (locality and continuity preserving) and Sequence Shuffle Attention for aggregating directional sequences, reporting SOTA on SR, denoising, deblurring and dehazing.
Problem
Mamba-based image restoration flattens 2D images into 1D sequences along rows and columns, which breaks local relationships and spatial continuity, and aggregates the resulting sequences by naive summation that ignores their differing contexts.
Approach
A nested S-shaped stripe-based scan that keeps locality (within stripes) and continuity (S-shaped path within and across stripes, with shift-stripe at boundaries) at no extra cost, plus an attention module (SSA) that weights the four directional sequences via channel-wise group convolution over shuffled channels.
Key results
- Classic SR x2 on Urban100: MaIR+ reaches 34.38 dB vs MambaIR 34.15 dB (MaIR 34.19).
- Real-image denoising on SIDD: MaIR 39.92 dB PSNR vs MambaIR 39.89 dB and Uformer 39.89 dB.
- Deblurring on GoPro: MaIR 33.69 dB, equal to NAFNet, and 31.57 dB on HIDE vs NAFNet 31.32 dB; dehazing outdoor SOTS 39.45 dB indoor vs UVM-Net 40.17 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MaIR instead of MambaIR | classic image super-resolution · Urban100 ×4 | PSNR: 27.68 -> 27.71 | Helped |
| Use MaIR instead of ART | image denoising · Urban100, σ=15 | PSNR: 35.29 -> 35.35 | Helped |
| Use MaIR instead of ART | image denoising · Urban100, σ=25 | PSNR: 33.14 -> 33.22 | Helped |
| Use MaIR instead of ART | image denoising · Urban100, σ=50 | PSNR: 30.19 -> 30.30 | Helped |
| Use MaIR instead of MambaIR | real image denoising · SIDD-Medium | PSNR: 39.89 -> 39.92 | Helped |
| Use MaIR instead of NAFNet | image deblurring · GoPro | PSNR: 33.69 -> 33.69 | No clear effect |
| Use MaIR instead of CU-Mamba | image deblurring · HIDE | PSNR: 31.47 -> 31.57 | Helped |
| Replace the proposed scanning with Z-shaped scanning (ablation) | MaIR · Urban100 ×2 lightweight SR | PSNR: 32.97 -> 32.94 | Hurt |
| Remove the shift-stripe mechanism (ablation) | MaIR · Urban100 ×2 lightweight SR | PSNR: 32.97 -> 32.93 | Hurt |
| Replace NSS with window-based scanning (ablation) | MaIR · Urban100 ×2 lightweight SR | PSNR: 32.97 -> 32.93 | Hurt |
| Replace NSS with Z-shaped scanning [19] (ablation) | MaIR · Urban100 ×2 lightweight SR | PSNR: 32.97 -> 32.88 | Hurt |
| Replace NSS with Peano-Hilbert curve scanning (ablation) | MaIR · Urban100 ×2 lightweight SR | PSNR: 32.97 -> 32.95 | Hurt |
Surprising findings
- Ablation on stripe width shows near-constant performance across widths 2 to 16, with only the largest (32) dropping.
- On indoor dehazing MaIR is slightly below UVM-Net in PSNR while using about 0.3% of its parameters.
Limitations
- Training hardware and compute details are not reported in the main text.
- Stripe-width study shows near-flat performance, so the stripe design's contribution is not strongly tuned.
- Indoor dehazing PSNR is below UVM-Net, and some ablations are limited to Urban100 ×2.
Experimental setup
Datasets: DIV2K Flickr2K DF2K WED BSD400 GoPro SIDD-Medium RESIDE ITS OTS RESIDE-6K
Benchmarks: Set5 Set14 B100 Urban100 Manga109 BSD68 Kodak24 McMaster SIDD GoPro HIDE SOTS-Indoor
Metrics: PSNR SSIM
Baselines: SAN HAN IGNN NLSA ELAN IPT SwinIR SRFormer MambaIR CARN IMDN LAPAR-A
Builds on: Vmamba MambaIR LocalMamba ZigMa SwinIR
Closest papers by meaning
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.91)
- DCTMamba: Advancing JPEG Image Restoration Through Long-Sequence Modeling and Adaptive Frequency Strategy (AAAI 2025 · similarity 0.88)
- ACL: Activating Capability of Linear Attention for Image Restoration (CVPR 2025 · similarity 0.87)
- TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond (CVPR 2025 · similarity 0.86)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.86)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · 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.