NoiseDiff

Dark Noise Diffusion: Noise Synthesis for Low-Light Image Denoising
📄 View on arXiv

---

Authors
Liying Lu  |  Raphaël Achddou  |  Sabine Süsstrunk

Get Started

1. Clone the repository

   git clone https://github.com/IVRL/NoiseDiff.git
   cd NoiseDiff

2. Install dependencies
   Run the installation script to set up the required environment.
   Please install PyTorch and torchvision versions matching your device.

   bash install.sh

3. Download the SID and ELD datasets
   Download the SID dataset from the Learning-to-See-in-the-Dark repository and the ELD dataset from the ELD repository.
   Please follow the instructions in each repository for downloading and preparing the datasets. Also change the data path in the dataloader/dataset.py, dataloader/dataset_denoising.py and test_denoising.py accordingly.

4. Download the dark shading resources for denoising training
   Download the dark shading resources from the PMN repository, place the resources folder under the root directory.

5. Download the pre-trained models
   Download the pre-trained models from Google Drive, and place the pretrained_ckpts folder under the root directory.

6. (Optional) Download our synthetic data
   The noise images generated by our model can be downloaded from HuggingFace.

---

Usage

All commands for training and testing are provided in script.sh. You can modify the parameters in the script as needed.

1. Diffusion Network Training & Testing (Data Generation)

• Training:
  Run the training command in script.sh to train the diffusion network. You can train it using our provided synthetic data or your own generated data.

• Testing / Data Generation:
  Use the provided command to generate synthetic noise data. Set the --resume argument to the path of your trained model checkpoint, or use the provided pretrained checkpoints in the --pretrained_ckpts folder.

• Logs: Training logs will be saved in the ./logs folder.

• Output: Generated data will be saved in the ./output folder.
  (Save paths can be changed by changing the --save_folder in the commands.)

2. Denoising Network Training & Testing

• Training:
  Use the corresponding command in script.sh to train the denoising network.

• Testing:
  Run the test command to evaluate the denoising performance on the SID or ELD dataset, with or without dark shading correction. Set the --resume argument to the path of your trained model checkpoint, or use the provided pretrained checkpoints in the --pretrained_ckpts folder.

• Logs: Training logs will be saved in the ./logs folder.

• Output: Denoising results will be saved in the ./output folder.
  (Save paths can be changed by changing the --save_folder in the commands.)

---

Notes

• The synthetic dataset is a bit large, we are still trying to upload the complete dataset.
• For any issues or questions, please open an issue on this repository, or contact us via email at [email protected].

---

Acknowledgement

This project refers to some code from the following repositories, we thank the authors for their contributions and publishing their code:

• Denoising-Diffusion-Pytorch
• Learning-to-See-in-the-Dark
• PMN

---

Citation

If you find this work useful, please cite our paper:

@inproceedings{NoiseDiff2025,
  title     = {Dark Noise Diffusion: Noise Synthesis for Low-Light Image Denoising},
  author    = {Lu, Liying and Achddou, Rapha{\"e}l and S{\"u}sstrunk, Sabine},
  booktitle = {Proceedings of the International Conference on Computational Photography (ICCP)},
  year      = {2025}
}