[ICCV 25] CWNet: Causal Wavelet Network for Low-Light Image Enhancement [[Paper]](https://openaccess.thecvf.com/content/ICCV2025/html/ZhangCWNetCausalWaveletNetworkforLow-LightImageEnhancementICCV2025_paper.html)

• Tongshun Zhang, Pingping Liu, Yubing Lu, Mengen Cai, Zijian Zhang, Zhe Zhang, Qiuzhan Zhou
• College of Computer Science and Technology, Jilin University
• Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education
• College of Communication Engineering, Jilin University

🔥 News

• [02/09/2025] We have updated our results based on the LOL-v2-Real [Google Drive] training set and evaluated on the Real test set:

Dataset	SSIM	PSNR	LPIPS
LOL-v2-Real (Updated)	0.8604	21.65	0.0651
Based on LOL-v1 (Previous/Paper)	0.9005	27.39	0.0383

The previous results reported in our paper for LOL-v2-Real were based on the LOL-v1 dataset, which we refer to as "Based_v1".

1. Abstract

Traditional Low-Light Image Enhancement (LLIE) methods primarily focus on uniform brightness adjustment, often neglecting instance-level semantic information and the inherent characteristics of different features. To address these limitations, we propose CWNet (Causal Wavelet Network), a novel architecture that leverages wavelet transforms for causal reasoning. Specifically, our approach comprises two key components: 1) Inspired by the concept of intervention in causality, we adopt a causal reasoning perspective to reveal the underlying causal relationships in low-light enhancement. From a global perspective, we employ a metric learning strategy to ensure causal embeddings adhere to causal principles, separating them from non-causal confounding factors while focusing on the invariance of causal factors. At the local level, we introduce an instance-level CLIP semantic loss to precisely maintain causal factor consistency. 2) Based on our causal analysis, we present a wavelet transform-based backbone network that effectively optimizes the recovery of frequency information, ensuring precise enhancement tailored to the specific attributes of wavelet transforms. Extensive experiments demonstrate that CWNet significantly outperforms current state-of-the-art methods across multiple datasets, showcasing its robust performance across diverse scenes.

CWNet Overall Architecture

Performance

2. Prepare Dataset

LOL-v1 Baidu Disk (code: cyh2), Google Drive. From Retinexformer.

LOL-v2 Baidu Disk (code: cyh2), Google Drive. From Retinexformer.

LSRW-Huawei Baidu Disk (code:wmrr). From R2RNet.

3. Create Environment

Installation

conda env create -f environment.yml
conda activate cwnet

(1) First download the pre-trained model (DA-CLIP, Segmentation model) from Google Drive.

(2) Download the pre-trained model DA-CLIP.

(3) Download the pre-trained segmentation model hrnet_w48_pascal_context_cls59_480x480.pth, and configure line 21 of models/archs/segment/hrseg_model.py: pretrained_dict.

(4) Configure the path in /options/xx.yml:

lightmap_GT: Generate the light map path via dataroot_GT, generated by LIME (LIME: Low-Light Image Enhancement via Illumination Map Estimation.) It can also be obtained from Google Drive.

dataroot_GT: your image path

dataroot_LQ: your image path

4. Train

python train.py -opt ./options/train/LSRW-Huawei.yml

5. Test

python test.py -opt ./options/test/LSRW-Huawei.yml

6. Pre-trained Models and Outputs

✨ Google Drive.

Citation Information

If you find the project useful, please cite:

@inproceedings{zhang2025cwnet,
  title={Cwnet: Causal wavelet network for low-light image enhancement},
  author={Zhang, Tongshun and Liu, Pingping and Lu, Yubing and Cai, Mengen and Zhang, Zijian and Zhang, Zhe and Zhou, Qiuzhan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={8789--8799},
  year={2025}
}