LLFlow (Fine-Tuned and Optimized Fork)

Low-light image enhancement with normalizing flow, adapted from the original LLFlow work and updated for this project’s fine-tuning and runtime optimization workflow.

Original Paper | Original Project

The training and inference pipeline has been adapted for:

• Fine-tuning on custom paired low/high datasets
• Stability-focused inference behavior
• Preprocessing options such as log-space low-light input and histogram-equalization concatenation
• Practical paired and unpaired test scripts with result export

Framework

Visual Example

Environment Setup

• Python 3.8
• CUDA-enabled PyTorch
• GPU for training and inference

git clone git@github.com:StrikerEurika/LLFlow.git
cd LLFlow/code
pip install -r requirements.txt

# Install PyTorch for your CUDA version
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

# For Conda
conda create -n llflow python=3.8 -y
conda activate llflow
cd LLFlow/code
pip install -r requirements.txt

Dataset Layout

LOL

Set dataset root to the folder that contains:

<LOL_ROOT>/
  our485/
    low/
    high/
  eval15/
    low/
    high/

This is required by Lol_dataset.py

LOL-v2

Set dataset root to the folder that contains:

<LOLV2_ROOT>/
  Synthetic/
    Train/
      Low/
      Normal/
    Test/
      low/
      high/
  Real_captured/
    Train/
      Low/
      Normal/
    Test/
      low/
      high/

This is required by Lol_dataset.py

Custom Paired Dataset

Set dataset root to:

<CUSTOM_ROOT>/
  low/
  high/

Used by dataset type Custom in Custom_smallNet_custom.yml

Training

cd code
python train.py --opt confs/LOL_smallNet.yml --tfboard

# Other examples
python train.py --opt confs/LOL-pc.yml --tfboard
python train.py --opt confs/LOLv2-pc.yml --tfboard
python train.py --opt confs/Custom_smallNet_custom.yml --tfboard

Outputs are created under:

LLFlow/experiments/<experiment_name>/

Including checkpoints, logs, validation images, and TensorBoard logs.

Paired Evaluation (PSNR / SSIM / LPIPS)

Run:

python test.py --opt confs/LOL_smallNet.yml

This script:

• Loads dataroot_LR and dataroot_GT from config
• Runs enhancement
• Saves enhanced images under:
• Saves metrics CSV with timestamp under:
• Implementation: test.py

Unpaired Evaluation / Inference Batch

Run:

python test_unpaired.py --opt confs/LOL_smallNet.yml -n unpaired_results

This script:

• Loads images from dataroot_unpaired
• Saves outputs to:
• Implementation: test_unpaired.py

Using In Your Own Project

Use using.py for folder-level enhancement.

Command:

python using.py \
  --input /path/to/input_folder \
  --output /path/to/output_folder \
  --model_path /path/to/latest_G.pth \
  --conf confs/Custom_smallNet_custom.yml

Arguments:

• --input: folder containing low-light images
• --output: folder to save enhanced images
• --model_path: path to trained weights
• --conf: model architecture/config yaml
• Implementation: using.py

License and Attribution

This project is based on LLFlow and includes third-party licensed components. See:

• LICENSE
• README.md If you use this work academically, please cite the original LLFlow paper:

@article{wang2021low,
  title={Low-Light Image Enhancement with Normalizing Flow},
  author={Wang, Yufei and Wan, Renjie and Yang, Wenhan and Li, Haoliang and Chau, Lap-Pui and Kot, Alex C},
  journal={arXiv preprint arXiv:2109.05923},
  year={2021}
}