Research map / Image generation / Image restoration
Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions
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 a RAW-domain camera ISP pipeline (unprocessing, low-light/noise/white-balance/CCM/tone-curve simulation, reprocessing) to generate unlimited synthetic low-light pairs, used to pre-train and fine-tune existing LLIE networks for better practicality and generalization.
Problem
Deep LLIE models need large paired datasets, but real captured pairs are scarce and existing sRGB-domain synthetic low-light data (gamma/Lightroom) has limited size and degradation diversity, causing poor white balance, artifacts, and poor generalization.
Approach
Synthesize low-light pairs by unprocessing sRGB to RAW, applying exposure reduction, heteroscedastic shot/read noise, white balance variation, and CCM/tone-curve diversity in RAW, then reprocessing through a forward ISP, and use this data to pre-train then fine-tune existing LLIE networks.
Key results
- Retinexformer trained with synthetic data then fine-tuned on LOL-v1 reaches 23.999 PSNR / 0.854 SSIM vs 23.622 / 0.828 for LOL-v1-only training.
- On MIT-Adobe FiveK, RetinexMamba+ improves PSNR from 18.287 to 20.262 over the baseline RetinexMamba.
- In the synthetic-data ablation, the RAW-domain noise synthesis reaches 22.168 PSNR vs 21.830 for sRGB-domain noise and 21.449 without noise.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with the proposed synthetic data and fine-tune on LOL-v1 instead of training only on LOL-v1 | RetinexMamba · LOL-v1 | PSNR: 23.527 -> 24.133 | Helped |
| Train with the proposed synthetic data and fine-tune on LOL-v2 instead of training only on LOL-v2 | Retinexformer · LOL-v2 | PSNR: 22.236 -> 22.514 | Helped |
| Train with the proposed synthetic data and fine-tune on LOL-v2 | RetinexMamba · MIT-Adobe FiveK | PSNR: 18.287 -> 20.262 | Helped |
| Train with the proposed synthetic data and fine-tune on LOL-v2 | Retinexformer · NPE | CLIPIQA: 0.475 -> 0.680 | Helped |
| Train with the proposed synthetic data and fine-tune on LOL-v2 | Retinexformer · DICM | CLIPIQA: 0.463 -> 0.610 | Helped |
| Train with the proposed synthetic data and fine-tune on LOL-v2 | Retinexformer · LIME | CLIPIQA: 0.490 -> 0.617 | Helped |
| Train with the proposed synthetic data and fine-tune on LOL-v2 | Retinexformer · VV | CLIPIQA: 0.490 -> 0.514 | Helped |
| Add noise synthesis (ablation) | Retinexformer trained with the proposed synthetic pipeline · LOL-v2 | PSNR: 21.449 -> 22.168 | Helped |
| Synthesize noise in the RAW domain instead of the sRGB domain (ablation) | Retinexformer trained with the proposed synthetic pipeline · LOL-v2 | PSNR: 21.830 -> 22.168 | Helped |
Limitations
- Synthetic-only training is combined with fine-tuning on target paired data, so standalone synthetic performance is not fully isolated.
- Evaluation of face detection uses only a single detector (YOLO-v5s) on DARK FACE.
Experimental setup
Datasets: ImageNet COCO LOL-v1 LOL-v2 MIT-Adobe FiveK NPE DICM LIME VV DARK FACE
Benchmarks: LOL-v1 LOL-v2 MIT-Adobe FiveK NPE DICM LIME VV DARK FACE
Metrics: PSNR SSIM LPIPS MUSIQ CLIPIQA Q-Align average precision
Baselines: SNR-LLIE Retinexformer RetinexMamba RetinexNet synthetic data MBLLEN synthetic data GladNet synthetic data AGLLNet synthetic data
Builds on: Brooks et al. 2019 unprocessing pipeline Retinexformer SNR-LLIE RetinexMamba
Closest papers by meaning
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.91)
- IniRetinex: Rethinking Retinex-type Low-Light Image Enhancer via Initialization Perspective (AAAI 2025 · similarity 0.88)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.87)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.87)
- DarkIR: Robust Low-Light Image Restoration (CVPR 2025 · similarity 0.87)
- HVI: A New Color Space for Low-light Image Enhancement (CVPR 2025 · similarity 0.86)
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.