Research map / Image generation / Image restoration
Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement
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 MDMS, a diffusion model for low-light image enhancement that adds a frequency-domain branch, multi-scale patch sampling to remove checkerboard artifacts, and a bright channel prior as conditioning.
Problem
Diffusion-based low-light image enhancement models learn noise distributions only in the spatial domain and use single-resolution patch sampling, which causes checkerboard artifacts from uneven patch overlap.
Approach
Jointly models spatial and FFT-based frequency features inside the diffusion U-Net (Multi-Domain Learning), samples overlapping patches at three resolutions and merges them (Multi-Scale Sampling), and conditions on a bright channel prior derived from the input.
Key results
- On LOL, MDMS reaches 27.12 dB PSNR, 0.882 SSIM and 0.078 LPIPS, beating the prior best PyDiff (27.07 dB, 0.100 LPIPS).
- On LOLv2-Real, MDMS reaches 33.30 dB PSNR, 2.19 dB above PyDiff (31.11 dB).
- Ablation: removing the frequency branch drops PSNR from 27.12 to 26.46 dB on LOL; the full model reaches the best result with sampling steps S=3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the spatial branch (ablation) | MDMS · LOL | PSNR: 26.36 -> 27.12 | Helped |
| add the frequency branch (ablation) | MDMS · LOL | PSNR: 26.46 -> 27.12 | Helped |
| use multi-scale sampling instead of only 64 × 64 patches (ablation) | MDMS · LOL | PSNR: 26.06 -> 27.12 | Helped |
| use multi-scale sampling instead of only 96 × 96 patches (ablation) | MDMS · LOL | PSNR: 26.40 -> 27.12 | Helped |
| use multi-scale sampling instead of only 128 × 128 patches (ablation) | MDMS · LOL | PSNR: 25.99 -> 27.12 | Helped |
| add 128 × 128 patches to 64 × 64 and 96 × 96 sampling (ablation) | MDMS · LOL | PSNR: 26.78 -> 27.12 | Helped |
| add Bright Channel Prior (ablation) | MDMS · LOL | PSNR: 26.80 -> 27.12 | Helped |
| use MDMS instead of Pydiff | low-light enhancement · LOL | PSNR: 27.07 -> 27.12 | Helped |
| use MDMS instead of Pydiff | low-light enhancement · LOLv2-Real | PSNR: 31.11 -> 33.30 | Helped |
| use MDMS instead of Pydiff | low-light enhancement · LOLv2-Syn | PSNR: 17.33 -> 17.40 | Helped |
Surprising findings
- Multi-scale sampling lets the model reach high quality with only 3 sampling steps, fewer than prior diffusion LLIE methods.
Limitations
- Evaluated only on LOL and LOLv2 paired datasets; generalization beyond them is not shown.
- Hardware and compute cost of training are not reported.
Experimental setup
Datasets: LOL LOLv2-Real LOLv2-Syn
Benchmarks: LOL LOLv2-Real LOLv2-Syn
Metrics: PSNR SSIM LPIPS
Baselines: PyDiff LLFlow SNR LLFormer IAT KinD++ MIRNetv2 Zero-DCE RUAS SCI
Builds on: DDIM PyDiff Bright Channel Prior Denoising diffusion probabilistic models
Closest papers by meaning
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.89)
- Efficient Diffusion as Low Light Enhancer (CVPR 2025 · similarity 0.88)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.88)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.88)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.87)
- Accelerated Diffusion via High-Low Frequency Decomposition for Pan-Sharpening (AAAI 2025 · similarity 0.87)
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.