[ECCV2024] DualDn:
Dual-domain Denoising via Differentiable ISP

👉 Project Page (For more visual results)

Ruikang Li, Yujin Wang, Shiqi Chen, Fan Zhang, Jinwei Gu and Tianfan Xue

News

• Sept 29, 2024: Paper accepted at ECCV 2024. 😊
• Nov 4, 2024: Training and inferencing code released. 🌹
• Nov 10, 2024: Pre-trained models and visual results released. Feel free to download! 📝
• Nov 15, 2024: Project page has evolved to the next generation—check it out! 💢

TODO Lists

• We have captured a new test set with real_captured RAW images, and we will release a new benchmark for dual-denoising.

Trained with Synthetic Noise.
ONE Model. NO Cherry-Picking. NO Fine-Tuning.

---

DualDn achieves greater generalizability compared to most learning-based denoising methods, as it can adapt to different unseen noises, ISP params, and even black-box ISPs. Experiments show that DualDn achieves SOTA performance and can adapt to various denoising backbones. Moreover, DualDn can be used as a plug-and-play denoising module with real cameras without retraining, and still demonstrate better performance than original commercial on-camera denoising, which can be seen from the visual results below.

---

Previous Methods

(1) Raw-domain denoising:

(2) sRGB-domain denoising:

Our DualDn

Network Architecture

Installation

See INSTALL.md for the installation of environment and dependencies required to run DualDn.

Demo

Train and Evaluation

We trained DualDn on a single GPU ONLY using: (1) clean raw images, (2) a reasonable noise model.

We chose the MIT-Adobe FiveK Dataset for training, as it's a robust dataset containing multiple raw images in DNG format with EXIF metadata. Although some images contain noise, MIT-Adobe FiveK is sufficient for training DualDn to generalize effectively to in-the-wild scenes.

And we believe that if more clean raws or a more accurate noise model are given, DualDn’s performance could improve even further.

Training on	Evaluating on	Test Sets	Instructions	Pre-trained Model	Visual Results
MIT-Adobe FiveK Download
	Synthetic Images	Download	Prerequisites \ Train \ Test \ Inference	Download	Download
	Real_captured Images	Download	Prerequisites \ Train \ Test \ Inference	Download	Download
	DND Benchmark	Download	Prerequisites \ Train \ Test \ Inference		Download

HINTS: We only provide models and results trained using Restormer backbone, as it demonstrated the best performance in DualDn. However, you can also train DualDn with other backbones, such as SwinIR or MIRNet-v2, by following our instructions. Reference evaluation metrics are also available in our paper.

Results

Citation

@inproceedings{li2025dualdn,
  title={Dualdn: Dual-domain denoising via differentiable isp},
  author={Li, Ruikang and Wang, Yujin and Chen, Shiqi and Zhang, Fan and Gu, Jinwei and Xue, Tianfan},
  booktitle={European Conference on Computer Vision},
  pages={160--177},
  year={2025},
  organization={Springer}
}

License

This repository is licensed under Attribution-NonCommercial 4.0 International

1. Attribution — Proper credit must be given, including a link to the license and an indication of any modifications made. This should be done in a reasonable manner, without implying endorsement by the licensor

2. NonCommercial — The Algorithm may NOT be used for commercial purposes. This includes, but is not limited to, the sale, licensing, or integration of the Software into commercial products or services.

For collaboration or inquiries, please contact us.

Acknowledgement

This code is based on the BasicSR toolbox.