[IJCAI 2023 oral] Pyramid Diffusion Models For Low-light Image Enhancement

Paper | Project Page | Supplement Materials

Pyramid Diffusion Models For Low-light Image Enhancement
Dewei Zhou, Zongxin Yang, Yi Yang
In IJCAI'2023

Overall

Quantitative results

Evaluation on LOL

The evauluation results on LOL are as follows

Method	PSNR	SSIM	LPIPS
KIND	20.87	0.80	0.17
KIND++	21.30	0.82	0.16
Bread	22.96	0.84	0.16
IAT	23.38	0.81	0.26
HWMNet	24.24	0.85	0.12
LLFLOW	24.99	0.92	0.11
PyDiff (Ours)	27.09	0.93	0.10

Dependencies and Installation

git clone https://github.com/limuloo/PyDIff.git
cd PyDiff
conda create -n PyDiff python=3.7
conda activate PyDiff
conda install pytorch==1.7.0 torchvision torchaudio cudatoolkit=11.0 -c pytorch
cd BasicSR-light
pip install -r requirements.txt
BASICSR_EXT=True sudo $(which python) setup.py develop
cd ../PyDiff
pip install -r requirements.txt
BASICSR_EXT=True sudo $(which python) setup.py develop

Dataset

You can refer to the following links to download the LOL dataset and put it in the following way:

PyDiff/
    BasicSR-light/
    PyDiff/
    dataset/
        LOLdataset/
            our485/
            eval15/

Pretrained Model

You can refer to the following links to download the pretrained model and put it in the following way:

PyDiff/
    BasicSR-light/
    PyDiff/
    pretrained_models/
        LOLweights.pth

Test

cd PyDiff/
CUDA_VISIBLE_DEVICES=0 python pydiff/train.py -opt options/infer.yaml

NOTE: When testing on your own dataset, set 'use_kind_align' in 'infer.yaml' to false. For details, please refer to #6.

Train

Training with 2 GPUs

For training purposes, the utilization of the following commands is advised if you possess 2 GPUs with a memory capacity of 24GB or higher, as outlined in the paper.

cd PyDiff/
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=22666 pydiff/train.py -opt options/train_v1.yaml --launcher pytorch

Training with a single GPU

Otherwise, you can use the following commands for training, which requires 1 GPU with memory >=24GB.

cd PyDiff/
CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch --nproc_per_node=1 --master_port=22666 pydiff/train.py -opt options/train_v2.yaml --launcher pytorch

Training on a Custom Low-Level Task Dataset

Please update the following fields in the PyDiff/options/train_v3.yaml file: YOUR_TRAIN_DATASET_GT_ROOT, YOUR_TRAIN_DATASET_INPUT_ROOT, YOUR_EVAL_DATASET_GT_ROOT, and YOUR_EVAL_DATASET_INPUT_ROOT. If required, please also update the PyDiff/pydiff/data/lol_dataset.py. Finally, please employ the subsequent commands to initiate the training process:

cd PyDiff/
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=22666 pydiff/train.py -opt options/train_v3.yaml --launcher pytorch

Please feel free to customize additional parameters to meet your specific requirements. To enable PyDiff to train on a single GPU, the PyDiff/options/train_v2.yaml file can be consulted.

Citation

If you find our work useful for your research, please cite our paper

@article{zhou2023pyramid,
  title={Pyramid Diffusion Models For Low-light Image Enhancement},
  author={Zhou, Dewei and Yang, Zongxin and Yang, Yi},
  journal={arXiv preprint arXiv:2305.10028},
  year={2023}
}

Acknowledgement

Our code is partly built upon BasicSR. Thanks to the contributors of their great work.