This is the implementation for HDR-GAN: HDR Image Reconstruction From Multi-Exposed LDR Images With Large Motions, Yuzhen Niu, Jianbin Wu, Wenxi Liu, Wenzhong Guo, Rynson W. H. Lau, in IEEE Transactions on Image Processing, 2021.
In this work, we proposed a novel GAN-based model, HDR-GAN, which produces high-quality HDR images from multi-exposed LDR images without the need to explicitly align the LDR images.

pip install -r requirements.txt

• Download training set of Kalantari's dataset, decompress and put it in ./dataset folder

dataset/
└── kalantari_dataset
    └── train
        ├── 16-09-28-01
        ├── 16-09-28-05
        ├── 16-09-28-06
        ├── 16-10-10-a-01
        ├── ....
        └── yyy

• Start training

python train.py \
    --epoch 256000 \
    --train_hw 512 512 \
    --batch_size 2 \
    --depth 3 \
    --unetpps \
    --gpu 0 \
    --loss_gan \
    --gan sphere \
    --lr 1e-4

You can change learning rate during training process by creating c{PID}.conf file with the content as follow:

LR: 1e-5

To monitor training, you can use Tensorboard in .\logs dir

python test.py \
     --unetpps \
     --gpu 0 \
     --ckpt pretrained/model \
     --cus_test_ds dataset/test

The output results are name as xxx_0.hdr and xxx_1.hdr, corresponding to the two results of deep HDR supervision in paper, respectively. You may use Photomatix for tonemapping .hdr files to obtain better visual effects.

@article{niu2021hdr,
  title={HDR-GAN: HDR image reconstruction from multi-exposed ldr images with large motions},
  author={Niu, Yuzhen and Wu, Jianbin and Liu, Wenxi and Guo, Wenzhong and Lau, Rynson WH},
  journal={IEEE Transactions on Image Processing},
  volume={30},
  pages={3885--3896},
  year={2021},
  publisher={IEEE}
}