Research map / Image generation / Image restoration
MambaIRv2: Attentive State Space 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 MambaIRv2, a Mamba-based image restoration backbone whose attentive state-space equation injects learned prompts into the output matrix to attend beyond the scanned sequence with one scan, plus semantic-guided neighboring to link distant similar pixels.
Problem
Mamba's causal state-space modeling restricts each pixel to attend only to preceding tokens in a scanned sequence, forcing multi-directional scans and suffering long-range decay, which hurts non-causal image restoration.
Approach
Reinterpret the state-space output matrix C as an attention-like query and add instance-specific learned prompts (routed via Gumbel-softmax from a shared prompt pool) to C so each pixel can query similar unscanned pixels; then reorder the image by semantic label (SGN) so similar pixels are adjacent in the 1D sequence, enabling single-scan modeling.
Key results
- MambaIRv2-light outperforms SRFormer-light by 0.35 dB PSNR on 2x Urban100 with 79K fewer parameters.
- MambaIRv2-B exceeds HAT by 0.16 dB on 2x Manga109 and MambaIRv2-S uses 55% of CAT-A's MACs while beating it by 0.17 dB on Manga109.
- MambaIRv2 surpasses Restormer by 0.29 dB PSNR on Urban100 for gaussian color denoising at sigma=15.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add ASE to window MHSA (ablation) | MambaIRv2-light · Urban100, 2× SR | PSNR: 32.89 -> 32.94 | Helped |
| add ASE to window MHSA (ablation) | MambaIRv2-light · Manga109, 2× SR | PSNR: 39.11 -> 39.20 | Helped |
| add SGN to MHSA + ASE (ablation) | MambaIRv2-light · Urban100, 2× SR | PSNR: 32.94 -> 32.97 | Helped |
| add SGN to MHSA + ASE (ablation) | MambaIRv2-light · Manga109, 2× SR | PSNR: 39.20 -> 39.24 | Helped |
| inject prompts into B instead of C (ablation) | MambaIRv2-light ASE · Urban100, 2× SR | PSNR: 32.97 -> 32.96 | Hurt |
| inject prompts into Δ instead of C (ablation) | MambaIRv2-light ASE · Urban100, 2× SR | PSNR: 32.97 -> 32.93 | Hurt |
| inject prompts into y instead of C (ablation) | MambaIRv2-light ASE · Urban100, 2× SR | PSNR: 32.97 -> 32.94 | Hurt |
| replace four-direction Mamba scanning with MambaIRv2's single semantic-space scan (ablation) | MambaIR · Urban100, 2× lightweight SR | PSNR: 32.92 -> 33.26 | Helped |
| replace four-direction Mamba scanning with MambaIRv2's single semantic-space scan (ablation) | MambaIR · Manga109, 2× lightweight SR | PSNR: 39.31 -> 39.39 | Helped |
| use MambaIRv2-L instead of HAT | classic SR · Manga109, 2× | PSNR: 40.26 -> 40.55 | Helped |
| use MambaIRv2 instead of Restormer | color image denoising · Urban100, σ=15 | PSNR: 35.13 -> 35.42 | Helped |
Surprising findings
- Adding prompts to the output matrix C works better than injecting into B, delta, or output y, which the authors attribute to C's proximity to the output.
Limitations
- Gains over HAT/DAT on classic SR are small or mixed across scales and datasets (e.g., MambaIRv2-S trails on some 3x/4x entries).
- Computational cost of MambaIRv2-B/L is still large in MACs (445.8G-664.5G at 2x 256x256 output).
Experimental setup
Datasets: DIV2K Set5 Set14 BSDS100 Urban100 Manga109 Classic5 LIVE1 CBSD68 Kodak24 McMaster
Benchmarks: Set5 Set14 BSDS100 Urban100 Manga109 Classic5 LIVE1 CBSD68 Kodak24 McMaster
Metrics: PSNR SSIM #params MACs
Baselines: SwinIR MambaIR HAT DAT CAT-A SRFormer Restormer EDT ELAN LatticeNet
Builds on: Mamba MambaIR linear attention SwinIR
Closest papers by meaning
- MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration (CVPR 2025 · similarity 0.91)
- TSP-Mamba: The Travelling Salesman Problem Meets Mamba for Image Super-resolution and Beyond (CVPR 2025 · similarity 0.86)
- Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution (NeurIPS 2026 · similarity 0.86)
- Vision Mamba Mender (NeurIPS 2024 · similarity 0.86)
- ACL: Activating Capability of Linear Attention for Image Restoration (CVPR 2025 · similarity 0.86)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · 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.