Research map / Image generation / Image restoration
URWKV: Unified RWKV Model with Multi-state Perspective for Low-light 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 URWKV, a unified RWKV-based encoder-decoder for low-light image restoration with luminance-adaptive normalization, EMA multi-state token shift, and state-aware selective fusion, trained from scratch per dataset on LLIE and LLIE-deblur benchmarks.
Problem
Low-light image enhancement and joint low-light enhancement plus deblurring struggle with dynamically coupled degradations (noise, blur, uneven luminance) and unified restoration models lack luminance-adaptive mechanisms and are parameter-heavy.
Approach
Extends RWKV blocks with multi-state perspective: LAN modulates normalization scale using luminance vectors from inter-stage states, EMA aggregates intra-stage states in the token shift, and SSF selectively fuses multi-stage encoder features into the decoder via spatial guiding weights instead of naive skip connections.
Key results
- URWKV reaches 23.11 dB PSNR on LOL-v2-real, outperforming Retinexformer (22.79 dB) by 0.32 dB with 2.25M params.
- On LOL-blur, URWKV achieves 27.27 dB PSNR and 0.890 SSIM, beating the second-best PDHAT (26.71 dB) by 0.56 dB.
- On MIT-Adobe FiveK, URWKV reaches 26.08 dB PSNR, 1.50 dB above Retinexformer (24.58 dB).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · LOL-v2-real | PSNR: 22.79 -> 23.11 | Helped |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · LOL-v2-syn | PSNR: 25.88 -> 26.36 | Helped |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · SID | PSNR: 24.44 -> 23.11 | Hurt |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · SMID | PSNR: 29.15 -> 29.44 | Helped |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · SDSD-indoor | PSNR: 29.78 -> 31.24 | Helped |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · SDSD-outdoor | PSNR: 29.83 -> 29.99 | Helped |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · MIT-Adobe FiveK | PSNR: 25.60 -> 26.08 | Helped |
| Use URWKV instead of the strongest listed baseline | low-light image restoration models · LOL-blur | PSNR: 26.71 -> 27.27 | Helped |
| Add LAN (ablation) | URWKV · LOL-v2-real | PSNR: 21.33 -> 22.71 | Helped |
| Add SSF (ablation) | URWKV · LOL-v2-real | PSNR: 21.33 -> 21.40 | Helped |
| Add multi-state aggregation (ablation) | URWKV with Q-shift · LOL-v2-real | PSNR: 22.19 -> 23.11 | Helped |
| Add Q-shift to multi-state aggregation (ablation) | URWKV · LOL-v2-real | PSNR: 22.25 -> 23.11 | Helped |
Surprising findings
- Multi-state aggregation with naive concatenation skip connections (21.83 dB) performs worse than single-state concatenation (22.71 dB) on LOL-v2-real, showing that naive multi-state fusion hurts without SSF.
Limitations
- Evaluated only on paired low-light benchmarks; generalization to real dynamically coupled degradations is asserted rather than shown on unpaired or in-the-wild data.
- Ablations are conducted only on LOL-v2-real, a single dataset.
Experimental setup
Datasets: LOL-v2-real LOL-v2-syn SID SMID SDSD-indoor SDSD-outdoor MIT-Adobe FiveK LOL-blur
Benchmarks: LOL-v2-real LOL-v2-syn SID SMID SDSD-indoor SDSD-outdoor MIT-Adobe FiveK LOL-blur
Metrics: PSNR SSIM FLOPs parameters
Baselines: KinD EnlightenGAN SNR-Net FourLLIE UHDFour LLFormer Retinexformer BiFormer RetinexMamba LEDNet PDHAT Uformer
Builds on: RWKV Restormer Retinexformer BiFormer
Closest papers by meaning
- Distilling Semantic Priors from SAM to Efficient Image Restoration Models (CVPR 2024 · similarity 0.86)
- UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior (CVPR 2025 · similarity 0.85)
- Navigating Image Restoration with VAR's Distribution Alignment Prior (CVPR 2025 · similarity 0.83)
- SVFR: A Unified Framework for Generalized Video Face Restoration (CVPR 2025 · similarity 0.83)
- UARE: A Unified Vision-Language Model for Image Quality Assessment, Restoration, and Enhancement (CVPR 2026 · similarity 0.83)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.83)
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.