Research map / Image generation / Super-resolution
QMambaBSR: Burst Image Super-Resolution with Query State Space Model
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 QMambaBSR, a burst super-resolution network with a Query State Space Model that lets the base frame query all burst frames for sub-pixel details while suppressing noise, plus an adaptive upsampling module.
Problem
Burst super-resolution must extract consistent sub-pixel details from multiple noisy low-resolution RAW frames while suppressing random high-frequency noise; existing frame-by-frame fusion and static upsampling cause artifacts and over-smoothing.
Approach
A Query State Space Model where the base frame gates the state-space input matrix B and step size Δ so it queries all current frames jointly (inter-frame query and intra-frame scanning), combined with an AdaUp module that adapts transposed-convolution kernels to the scene's sub-pixel distribution.
Key results
- On Synthetic BurstSR (x4), reaches 43.12 dB PSNR, 0.29 dB above Burstormer (42.83 dB).
- On RealBSR-RAW, 27.558 dB PSNR vs. Burstormer 27.290 dB; on RealBSR-RGB, 31.401 dB PSNR vs. FBANet 31.012 dB.
- Ablation: MSFM adds 1.34 dB, QSSM adds 0.72 dB, AdaUp adds 0.26 dB over the baseline on synthetic data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QMambaBSR instead of Burstormer | BurstSR · Synthetic BurstSR ×4 | PSNR: 42.83 -> 43.12; 0.29 dB improvement | Helped |
| Use QMambaBSR instead of Burstormer | BurstSR · RealBSR-RAW ×4 | PSNR: 27.290 -> 27.558; 0.268 dB | Helped |
| Use QMambaBSR instead of Burstormer | BurstSR · RealBSR-RAW ×4 | L-PSNR: 32.533 -> 32.791; 0.258 dB | Helped |
| Use QMambaBSR instead of Burstormer | BurstSR · RealBSR-RGB ×4 | PSNR: 31.197 -> 31.401; 0.204 dB | Helped |
| Add MSFM to the baseline (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 39.81 -> 41.15; 1.34 dB | Helped |
| Add QSSM to the MSFM configuration (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 41.15 -> 41.87; 0.72 dB | Helped |
| Add AdaUp to the MSFM and QSSM configuration (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 41.87 -> 42.13; 0.26 dB | Helped |
| Use proposed fusion rather than concatenation (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 39.85 -> 42.13 | Helped |
| Use proposed fusion rather than PBFF (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 40.57 -> 42.13; 1.56 dB | Helped |
| Use proposed fusion rather than NRFE (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 41.72 -> 42.13; 0.41 dB | Helped |
| Use AdaUp rather than pixel shuffle (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 41.89 -> 42.13; 0.16 dB | Helped |
| Add Transformer branch to convolution-only MSFM (ablation) | QMambaBSR · Synthetic BurstSR ×4 (100 epochs) | PSNR: 41.57 -> 41.96; 0.39 dB | Helped |
Limitations
- Gains over Burstormer on synthetic and real benchmarks are modest (0.1 to 0.3 dB).
- Mamba long-range attenuation motivates adding a Transformer branch, so the global modeling of the SSM alone is limited.
Experimental setup
Datasets: Synthetic BurstSR Real BurstSR RealBSR-RAW RealBSR-RGB
Benchmarks: Synthetic BurstSR RealBSR-RAW RealBSR-RGB
Metrics: PSNR SSIM L-PSNR
Baselines: HighRes-net DBSR LKR MFIR BIPNet AFCNet FBAnet Burstormer RBSR GMTNet BSRT
Builds on: Mamba (Visual Mamba / MambaIR) Burstormer BIPNet DBSR
Closest papers by meaning
- Deep Burst Super-Resolution (CVPR 2021 · similarity 0.90)
- Burst Image Restoration and Enhancement (CVPR 2022 · similarity 0.88)
- Sequence Matters: Harnessing Video Models in 3D Super-Resolution (AAAI 2025 · similarity 0.86)
- Arbitrary-Scale Video Super-resolution Guided by Dynamic Context (AAAI 2024 · similarity 0.84)
- SinSR: Diffusion-Based Image Super-Resolution in a Single Step (CVPR 2024 · similarity 0.84)
- AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR 2024 · 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.