Research map / Image generation / Image restoration

DarkIR: Robust Low-Light Image Restoration

CVPR 2025 · Image generation › Image restoration · approach: Priors for restoration networks · 96 citations (OpenAlex)

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 DarkIR, a lightweight CNN with Fourier-domain encoder blocks and dilated-spatial-attention decoder blocks for joint low-light enhancement, denoising and deblurring, achieving SOTA on LOLBlur, LOLv2 and Real-LOLBlur with fewer parameters and MACs.

Problem

Night and low-light photos suffer jointly from noise, low light and blur, but most restoration methods solve deblurring and low-light enhancement separately, and existing joint models are heavy.

Approach

An encoder-decoder CNN with asymmetric blocks: Fourier-domain (amplitude-only) attention and Fre-MLP in the low-resolution encoder for illumination correction, and a decoder with Dilated-Spatial Attention Module (dilations 1, 4, 9) for large-receptive-field deblurring, trained with an intermediate low-light guiding loss.

Key results

• On LOLBlur, DarkIR-l reaches 27.30 dB PSNR vs 25.74 dB for LEDNet (about +1 dB) with 12.96M params vs 7.4M for LEDNet's ... and DarkIR-m (3.31M params, 7.25 GMACs) reaches 27.00 dB.
• DarkIR-mt achieves 23.87 dB PSNR on LOLv2-Real, the best reported in Table 4 (vs 22.80 for RetinexFormer), and 18.93 dB on LSRW vs 17.67 for the next best listed method.
• DarkIR-m uses about 55% fewer parameters than LEDNet and 88% fewer than Restormer, with 4x fewer MACs than LEDNet.

Measured effects reported in the paper

Surprising findings

• Inference time is not reduced proportionally to the MAC reduction because depth-wise convolutions have low arithmetic intensity on GPUs.
• The multi-task model trained on all datasets loses about 0.4 dB on LOLBlur (26.62 dB) relative to the single-task model.

Limitations

• Depth-wise convolutions do not map efficiently onto some GPU architectures, so inference time falls less than MAC count.
• Multi-task model shows a slight PSNR drop on LOLBlur compared with the single-task model.
• Training is on synthetic LOLBlur; real-world evaluation relies on no-reference metrics (MUSIQ, NRQM, NIQE) since Real-LOLBlur lacks ground truth.

Experimental setup

Datasets: LOLBlur LOLv2-Real LOLv2-Synthetic LSRW

Benchmarks: LOLBlur Real-LOLBlur LOLv2-Real LOLv2-Synthetic LSRW

Metrics: PSNR SSIM LPIPS MUSIQ NRQM NIQE MACs Params

Baselines: LEDNet Restormer RetinexFormer NAFNet MIMO-UNet DeblurGAN-v2 KinD++ DRBN

Builds on: NAFNet LEDNet Metaformer Large Kernel Attention FourLLIE

Closest papers by meaning

• Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.87)
• DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.87)
• Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.87)
• VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook (AAAI 2024 · similarity 0.86)
• Contrastive Learning for Low-Light Raw Denoising (Student Abstract) (AAAI 2024 · similarity 0.86)
• Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)