A Jax/Flax implementation of TIP 2021 paper, "Learning to Enhance Low-Light Image via Zero-Reference Deep Curve Estimation"
- Python >= 3.6
- JAX
- requirements.txt
Zero-DCE is trained with a selected subset of SICE dataset, which can be downloaded from Google Drive
Directly run on Colab
# Train Zero-DCE model
python train.py --model=Zero-DCE --train_dir=$DATA_DIR [--optional arguments]
# Train Zero-DCE++ model
python train.py --model=Zero-DCE++ --train_dir=$DATA_DIR [--optional arguments]
# Test Zero-DCE model
python test.py --model=Zero-DCE --ckpt_path=${CKPT_PATH} [--optional arguments]
# Test Zero-DCE++ model
python test.py --model=Zero-DCE++ --ckpt_path=${CKPT_PATH} [--optional arguments]
Method | Input shape | Flops | Params |
---|---|---|---|
Zero-DCE | (3, 1200, 900) | 84.990 GFlops | 0.079 M |
Zero-DCE++ | (3, 1200, 900) | 0.115 GFlops | 0.010 M |