Denoising Diffusion Post-Processing for Low-Light Image Enhancement

---

News

• Accepted for presentation in the Journal to Conference track at 2025 International Joint Conference on Neural Networks (IJCNN 2025) 🥳
• The article is officially published in Pattern Recognition and is openly accessible [here] 🎉

---

Abstract: Low-light image enhancement (LLIE) techniques attempt to increase the visibility of images captured in low-light scenarios. However, as a result of enhancement, a variety of image degradations such as noise and color bias are revealed. Furthermore, each particular LLIE approach may introduce a different form of flaw within its enhanced results. To combat these image degradations, post-processing denoisers have widely been used, which often yield oversmoothed results lacking detail. We propose using a diffusion model as a post-processing approach, and we introduce Low-light Post-processing Diffusion Model (LPDM) in order to model the conditional distribution between under-exposed and normally-exposed images. We apply LPDM in a manner which avoids the computationally expensive generative reverse process of typical diffusion models, and post-process images in one pass through LPDM. Extensive experiments demonstrate that our approach outperforms competing post-processing denoisers by increasing the perceptual quality of enhanced low-light images on a variety of challenging low-light datasets. Source code is available at https://github.com/savvaki/LPDM.

---

Download Results

The image results for all techniques and datasets are available for download here.

• undarken directories contain $\hat{\boldsymbol{x}}_0^\eta$ for each low-light enhancement technique $\eta$. In other words, these are the results before post-processing.
• denoised directories contain the results of LPDM in the lpdm_lol subdirectory, as well as the results for the ablation studies. For the LOL dataset, the results of NAFNet and BM3D are also provided. BM3D results are provided for different values of $\sigma$.
• The LPDM results have subdirectories in the format phi_s which indicate the $\phi$ and $s$ parameters used.

---

Requirements

Set up and activate the virtual conda environment:

• conda env create -f environment.yml
• conda activate lpdm

---

Test

• Download the pre-trained model .ckpt file and .yaml file here and place them in the checkpoints directory:

  ───checkpoints
       ├───lpdm_lol.ckpt
       └───lpdm_lol.yaml
  

• Modify the s and phi (φ) parameters in configs/test/denoise.yaml as desired.
• Place your low-light images (c) in the test/dark directory
• Place your enhanced images by any enhancer η in the test/eta directory. Make sure the images have the same names as the files in test/dark
• Run the following commands:

  cd scripts
  python denoise_config.py
  

• Denoised results can be found in test/denoised
• Run the following command for info about optional extra settings: python denoise_config.py --help.
• By default, the device is set to CPU. To use GPU pass the argument --device "cuda"

Calculate Metrics

• Modify configs/test/metrics.yaml and specify the denoised image glob path pred_path and the ground truth glob path target_path. To use non-reference metrics, remove target_path from the config.
• Run the following commands:

  cd scripts
  python calculate_metrics.py
  

• Metrics will be saved in test/results.csv.
• For additional options run python calculate_metrics.py --help

---

Train

• Download the LOL Dataset here. Unzip the dataset and place the contents in the datasets/lol directory:

  ───datasets
     └───lol
          ├───eval15
          │   ├───high
          │   └───low
          └───our485
              ├───high
              └───low
  

• For the commands below, PyTorch lightning will create a logs directory where checkpoints will be saved.

• For GPU training options pass the --gpu argument according to PyTorch Lighting documentation. The examples below use one GPU.

• Train and log locally:

  • python main.py --base configs/train/lpdm_lol.yaml --gpu 0,

• Train and log metrics with wandb:

  • Set your wandb environment variables: WANDB_API_KEY, WANDB_ENTITY, WANDB_PROJECT
  • python main.py --base configs/train/lpdm_lol.yaml --logger_to_use wandb --gpu 0,

• An example command to resume a saved checkpoint in the directory RUN_DIRECTORY_NAME_HERE using wandb logging:

  • python main.py --resume logs/RUN_DIRECTORY_NAME_HERE --logger_to_use wandb --gpu 0,

• To reduce the size of the model, reduce the model_channels parameter in configs/train/lpdm_lol.yaml

---

BibTeX Citation

@article{panagiotou2024denoising,
title = {Denoising diffusion post-processing for low-light image enhancement},
journal = {Pattern Recognition},
volume = {156},
pages = {110799},
year = {2024},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2024.110799},
url = {https://www.sciencedirect.com/science/article/pii/S0031320324005508},
author = {Savvas Panagiotou and Anna S. Bosman}
}

---

References

This repository is a derivative of the original Stable Diffusion repository.