Research map / Architectures and efficiency / CNN design
MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space
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 MambaLLIE, a U-shaped state space model for low light image enhancement that combines a local-enhanced state space module and an implicit Retinex-aware selective kernel module in a global-then-local block.
Problem
Low light image enhancement needs both global illumination correction and local denoising/deblurring, but CNNs and Transformers have limited receptive fields or quadratic cost, and vanilla vision SSMs neglect local 2D structure.
Approach
A Global-then-Local State Space Block pairs a 2D selective scan with an augmented local bias (LESSM) and a spatially varying selective-kernel module (IRSK) guided by max/mean illumination priors, avoiding explicit Retinex decomposition.
Key results
- MambaLLIE reaches 22.95 dB PSNR on LOL-V2-real and 25.87 dB on LOL-V2-syn, the best PSNR in the table, versus 22.80 and 25.67 for RetinexFormer.
- On SDSD-outdoor MambaLLIE reaches 30.00 dB PSNR, while SSIM on LOL-V2-real and SDSD-outdoor is second-best behind RetinexFormer and SNR-Net.
- In a 70-participant user study MambaLLIE scores highest on overall rating (4.243 vs 4.100 for RetinexFormer), and it has the best mean detection score on ExDark (0.768).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add implicit Retinex-aware guidance instead of estimating the enhanced result directly without a prior (ablation) | MambaLLIE · Ablation study | PSNR: 1.25 dB | Helped |
| use implicit Retinex-aware guidance instead of illumination-map estimation and element-wise multiplication (ablation) | MambaLLIE · Ablation study | PSNR: 1.00 dB | Helped |
| add LESSM to vanilla state space blocks (ablation) | MambaLLIE · Ablation study | PSNR: 0.33 dB and 0.08 dB | Helped |
| add IRSK to the model (ablation) | MambaLLIE · Ablation study | PSNR: 0.96, 0.74, and 0.63 dB | Helped |
| use a 3*3 selective kernel instead of 3*5 (ablation) | MambaLLIE · Ablation study | PSNR: 30.12 -> 29.55 | Hurt |
| use a 5*5 selective kernel instead of 3*5 (ablation) | MambaLLIE · Ablation study | PSNR: 30.12 -> 29.48 | Hurt |
| use a 5*7 selective kernel instead of 3*5 (ablation) | MambaLLIE · Ablation study | PSNR: 30.12 -> 28.88 | Hurt |
| use a 5*3 selective kernel instead of 3*5 (ablation) | MambaLLIE · Ablation study | PSNR: 30.12 -> 29.31 | Hurt |
Surprising findings
- Its PSNR gains over RetinexFormer are small (about 0.2 dB on average), and SSIM is not best on several datasets.
- Mapping small selective kernels to early layers and larger ones to later layers reverses the LSKNet design, which the authors attribute to padding effects in enhancement.
Limitations
- Requires hand-designed illumination priors that depend on prior experience.
- Relies on PSNR and SSIM, which the authors acknowledge are biased; they add only a user study and real-world evaluation.
Experimental setup
Datasets: LOL-V2-real LOL-V2-syn SMID SDSD-indoor SDSD-outdoor ExDark
Benchmarks: LOL-V2-real LOL-V2-syn SMID SDSD-indoor SDSD-outdoor ExDark
Metrics: PSNR SSIM FLOPS parameters mAP-style per-class detection score user study rating
Baselines: RetinexNet DeepUPE SID KinD MIRNet EnGAN Restormer SNR-Net QuadPrior MambaIR RetinexFormer ZeroDCE
Builds on: Mamba VMamba MambaIR LSKNet RetinexFormer
Closest papers by meaning
- IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective (AAAI 2025 · similarity 0.89)
- ECMamba: Consolidating Selective State Space Model with Retinex Guidance for Efficient Multiple Exposure Correction (NeurIPS 2024 · similarity 0.86)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.85)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.84)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (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.