ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement
ControlLight is presented in the paper ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement.
ControlLight is a controllable low-light enhancement model built on top of FLUX.2 [klein] 9B. It is trained as a LoRA for continuous illumination enhancement, enabling users to adjust enhancement strength with a controllable parameter alpha. The model is designed to enhance low-light images while preserving the original scene structure, visual content, and fine-grained details.
🔥🔥🔥 News!!
- May 2026: 👋 We release ControlLight, its model weights, inference and training code.
- May 2026: 👋 We release Light100K, a continuous low-light enhancement dataset for controllable illumination learning.
⚡️ Model Usage
Installation
This project currently relies on the patched local diffusers/ checkout from the ControlLight repository.
git clone https://github.com/yfyang007/ControlLight.git
cd ControlLight
conda create -n controlight python=3.12 -y
conda activate controlight
python -m pip install --upgrade pip
python -m pip install -e diffusers
python -m pip install -r requirements.txt
python -m pip install -e .
You can verify the environment with:
bash scripts/predict.sh --help
bash scripts/demo.sh --help
bash -lc 'source scripts/project_env.sh; python run.py --help >/dev/null'
Inference with ControlLight
bash scripts/predict.sh predict-image \
--input /path/to/input.jpg \
--output /path/to/output.png \
--model-path /path/to/FLUX.2-klein-base-9B \
--lora-path /path/to/controllight.safetensors \
--alpha 0.50 \
--num-inference-steps 20 \
--guidance-scale 1.0 \
--seed 42 \
--device cuda \
--torch-dtype bfloat16
CLI Quick Start
bash scripts/predict.sh predict-four \
--input /path/to/images \
--output /path/to/out_four \
--model-path /path/to/FLUX.2-klein-base-9B \
--lora-path /path/to/controllight.safetensors \
--num-inference-steps 20 \
--seed 42 \
--device cuda \
--torch-dtype bfloat16
Recommended Inference Config
- Device:
cuda - Torch dtype:
bfloat16 - Inference steps:
20 - Guidance scale:
1.0 - Recommended seed:
42 - Enhancement strength:
alphain[0, 1], where larger values produce stronger low-light enhancement.
Example Settings
| Task | Setting |
|---|---|
| Mild Low-light Enhancement | alpha=0.25 |
| Medium Low-light Enhancement | alpha=0.50 |
| Strong Low-light Enhancement | alpha=0.75 |
| Full Low-light Enhancement | alpha=1.00 |
| Custom Enhancement Sweep | --alphas 0.20,0.40,0.60,0.80 |
Additional Resources
- Project Page: ControlLight Project Page
- GitHub Repository: yfyang007/ControlLight
- Model: ControlLight/ControlLight
- Dataset: ControlLight/Light100K
- Base Model: black-forest-labs/FLUX.2-klein-base-9B
License and Disclaimer
The code of ControlLight is intended to be released under the Apache License 2.0.
ControlLight is built on top of FLUX.2 [klein] 9B and uses third-party components, datasets, and model assets. All underlying base models and third-party components remain governed by their original licenses and terms. Users must comply with all applicable upstream licenses when using this project.
Citation
If you find ControlLight useful in your research, please star and cite:
@misc{yang2026controllightcontrollableconsistentgeneralizable,
title={ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement},
author={Yufeng Yang and Jianzhuang Liu and Jisheng Chu and Yuqi Peng and Xianfang Zeng and Jiancheng Huang and Shifeng Chen},
year={2026},
eprint={2605.25569},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2605.25569},
}