Research map / Architectures and efficiency / Attention and transformers
ACL: Activating Capability of Linear Attention 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 ACL, a Mamba-style image restoration network that replaces the SSM with linear attention (LAMA) plus a multi-scale dilated convolution module (MDC) for de-raining and de-blurring.
Problem
Mamba's unidirectional, recursive 1D scanning is ill-suited to 2D image restoration, losing local pixel relations and being less efficient, while Transformers have quadratic attention cost.
Approach
Shows that linear attention and the SSM in Mamba share a similar recursive formulation, and replaces the SSM with linear attention inside a Mamba-like block to get global, multi-directional pixel connections at linear cost, with an added multi-scale dilated convolution module for local detail.
Key results
- ACL reaches 39.18 PSNR / 0.983 SSIM on Rain100L and 32.22 / 0.920 on Rain100H, the best among compared methods in Table 1.
- On GoPro deblurring, ACL achieves 33.25 PSNR / 0.964 SSIM with 55 GFLOPs and 4.6M parameters, while LoFormer reports 33.73 PSNR at 47 GFLOPs.
- Ablation on Rain100L: replacing the baseline's blocks with LAMA raises PSNR from 38.24 to 38.89, and adding MDC raises it to 39.18.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace baseline linear-attention blocks with LAMA (ablation) | ACL baseline · Rain100L | PSNR: 38.24 -> 38.89 | Helped |
| Add MDC to the baseline with LAMA (ablation) | ACL baseline · Rain100L | PSNR: 38.89 -> 39.18 | Helped |
| Replace LAMA with one-dimensional scanning SSM (ablation) | ACL · Rain100L | PSNR: 39.18 -> 36.47 | Hurt |
| Replace LAMA with bidirectional scanning SSM (ablation) | ACL · Rain100L | PSNR: 39.18 -> 38.07 | Hurt |
| Replace LAMA with standard linear attention (ablation) | ACL · Rain100L | PSNR: 39.18 -> 38.24 | Hurt |
| Set encoder/decoder block counts to (3,3,3) instead of (6,3,3) (ablation) | ACL · Rain100L | PSNR: 39.18 -> 38.95 | Hurt |
| Set encoder/decoder block counts to (4,4,4) instead of (6,3,3) (ablation) | ACL · Rain100L | PSNR: 39.18 -> 39.03 | Hurt |
| Set encoder/decoder block counts to (6,6,6) instead of (6,3,3) (ablation) | ACL · Rain100L | PSNR: 39.18 -> 39.19 | No clear effect |
| Set MDC dilation rate to 1 instead of 2 (ablation) | ACL · Rain100L | PSNR: 39.18 -> 38.12 | Hurt |
| Set MDC dilation rate to 3 instead of 2 (ablation) | ACL · Rain100L | PSNR: 39.18 -> 39.10 | Hurt |
Surprising findings
- On GoPro deblurring, ACL is not SOTA (LoFormer and SSAMAN report higher PSNR), which the authors acknowledge as a gap to Transformer models.
Limitations
- Does not outperform CNN-based models on large-size images.
- Gap to SOTA Transformer models on image deblurring.
- Evaluated only on de-raining and de-blurring.
Experimental setup
Datasets: Rain100L Rain100H Rain200L Rain200H DID-Data GoPro
Benchmarks: Rain100L Rain100H Rain200L Rain200H DID-Data GoPro
Metrics: PSNR SSIM FLOPs Param
Baselines: Restormer MAXIM DRT MPRNet DAWN IRNeXt MambaIR NAFNet Uformer Stripformer LoFormer MIMO
Builds on: Mamba MambaIR linear attention Restormer NAFNet
Closest papers by meaning
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.87)
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.87)
- Sharing Key Semantics in Transformer Makes Efficient Image Restoration (NeurIPS 2024 · similarity 0.86)
- MambaIRv2: Attentive State Space Restoration (CVPR 2025 · similarity 0.86)
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.86)
- UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · 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.