CLODE

Continuous Exposure Learning for Low-light Image Enhancement using Neural ODEs <br> (ICLR 2025 Spotlight)

This repository is the official implementation of "Continuous Exposure Learning for Low-light Image Enhancement using Neural ODEs" @ ICLR25.

Donggoo Jung*, Daehyun Kim*, Tae Hyun Kim $^\dagger$ (*Equal Contribution)

[ICLR2025] Paper

Method

We propose the unsupervised low-light image enhancement problem by reframing discrete iterative curve-adjustment methods into a continuous space using Neural Ordinary Differential Equations (NODE).

| Under-exposure | Over-exposure | Normal-exposure | | :------------: | :-----------: | :-------------: | | TBD | TBD | TBD |

Evaluation

Download the pre-trained model and place it in ./pth/.

# In inference.py, only modify the following paths:
# file_path: Path to the input images
# gt_path: Path to the ground truth images
# file_path = '/path/to/your/input'
# gt_path = '/path/to/your/corresponding_gt'

$ python inference.py

User Controllablity

CLODE learns the low-light exposure adjustment mechanism in the continuous-space, and is trained to output $I_T$ by integrating the states from $0$ to $T$ using a fixed $T=3$. However, users can manually adjust the integration interval by changing the final state value $T$ at the test stage, allowing them to output images with the preferred exposure level and even produce images darker than the input. In practice, by controlling the final state from $-(T+\Delta t)$ to $(T+\Delta t)$, the exposure level of the output image can be easily controlled to provide a more user-friendly exposure level.

$ python inference.py --T 4.8    # set to 3.5, more brighter
$ python inference.py --T -1.4   # set to -1.4, more darker
$ python inference.py --T 2.5    # set to 2.5, Adjust to the brightness desired by the user

Results

We provide our results for the LOL and SICE Part2 dataset. (CLODE/CLODE$\dagger$) | Dataset | PSNR | SSIM | Images| | :------:| :---:| :---:| :---: | | LOL | 19.61/23.58 | 0.718/0.754 | Link/Link | | SICE | 15.01/16.18 | 0.687/0.707 | Link/Link |

Train

python main_experiment.py

Citation

If you find our work useful in your research, please consider citing our paper.

@article{jung2025continuous,
  title={Continuous Exposure Learning for Low-light Image Enhancement using Neural {ODE}s},
  author={Donggoo Jung and Daehyun Kim and Tae Hyun Kim},
  booktitle={ICLR},
  year={2025},
}

Acknowledgement

We are using torchdiffeq as the Neural ODEs library. We thank the author for sharing their codes.