Towards Robust Event-guided Low-Light Image Enhancement: <br> A Large-Scale Real-World Event-Image Dataset and Novel Approach

[CVPR 2024 Oral & TPAMI 2025]

:loudspeaker: News

[2025.09.23] :tada: Our extension paper "EvLight++" is now published in IEEE TPAMI! This work extends the original EvLight to low-light video enhancement with improved methodology and extensive applications (Source code is released).
[2024.12.12] Normal-light event streams are released.
[2024.08.24] Source code is released.
[2024.06.15] SDE dataset and synthetic event dataset of SDSD are released.
[2024.04.06] Dataset and code release plan announced.

:pushpin: Roadmap & Status

[x] Release of synthetic event dataset of SDSD
[x] Release of our collected SDE dataset
[x] Release of source code
[x] Release of split normal-light event streams and the whole normal-light event streams

:file_folder: Dataset Preparation

1. SDE Dataset (Real-World)

The SDE dataset contains 91 image+event paired sequences (43 indoor, 48 outdoor) captured with a DAVIS346.

Resolution: 346 × 260
Split: 76 training sequences, 15 testing sequences.

| Dataset Content | Baidu Netdisk | OneDrive | Password | | :--- | :---: | :---: | :---: | | Aligned Dataset | Link | Link | w7qe | | Normal-Light Events | - | Link | - |

Note: We focus on the consistency between normal/low-light images. Consistency between event streams has not yet been fully verified.

<details> <summary>Click to view SDE Directory Structure</summary>

--indoor/outdoor 
├── test 
│   ├── pair1 
│   │   ├── low 
│   │   │   ├── xxx.png (low-light RGB frame) 
│   │   │   ├── xxx.npz (split low-light event streams) 
│   │   │   └── lowlight_event.npz (the whole low-light event stream) 
│   │   └── normal 
│   │       └── xxx.png (normal-light RGB frame) 
└── train 
    └── pair1 
        ├── low 
        │   ├── xxx.png 
        │   ├── xxx.npz 
        │   └── lowlight_event.npz 
        └── normal 
            └── xxx.png

</details>

2. SDSD Dataset (Synthetic Events)

We incorporated events into the SDSD dataset using the v2e simulator (resized to 346x260).

:warning: Notice:

Please download the latest version (we fixed previous issues).

We recommend skipping the first/last 3 split event files due to sparse events caused by slow motion.

<details> <summary>Click to view SDSD Directory Structure</summary>

--indoor/outdoor 
├── test 
│   └── pair1 
│       ├── low (split low-light event streams for each RGB frame) 
│       └── low_event (whole synthetic low-light event stream) 
└── train 
    └── pair1 
        ├── low 
        └── low_event

</details>

:computer: Usage

1. Dependencies

pip install -r requirements.txt

2. Pretrained Models

Download models from Baidu Pan (pwd: 8agv) or OneDrive.

Video-based checkpoints from Baidu Pan (pwd: n1b7) or OneDrive.

3. Training

Modify the dataset path in options/train/xxx.yaml.
Run the training script:

sh options/train/xxx.sh

For video enhancement, use the corresponding *_vid.sh scripts.

4. Testing

Modify the model and dataset paths in options/test/xxx.yaml.
Run the test script:

sh options/test/xxx.sh

For video enhancement, use the corresponding *_vid.sh scripts.

:mortar_board: Citation

If this work is helpful for your research, please consider citing:

@ARTICLE{11192751,
  author={Chen, Kanghao and Liang, Guoqiang and Lu, Yunfan and Li, Hangyu and Wang, Lin},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  title={EvLight++: Low-Light Video Enhancement With an Event Camera: A Large-Scale Real-World Dataset, Novel Method, and More},
  year={2026},
  volume={48},
  number={2},
  pages={1608-1625},
  keywords={Cameras;Videos;Semantic segmentation;Depth measurement;Feature extraction;Signal to noise ratio;Lighting;Semantics;Image color analysis;Training;Low light enhancement;high dynamic range;event camera;real-world dataset;downstream applications},
  doi={10.1109/TPAMI.2025.3617801}
}

@inproceedings{liang2024towards,
  title={Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach},
  author={Liang, Guoqiang and Chen, Kanghao and Li, Hangyu and Lu, Yunfan and Wang, Lin},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={23--33},
  year={2024}
}

:heart: Acknowledgment

We thank the authors of INR-Event-VSR and Retinexformer for their open-source contributions.

EvLight

Install / Use

README