This repository includes the official implementation of the paper AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution (CVPR2024).

A suitable conda environment named adabm can be created and activated with:

conda env create -f environment.yaml
conda activate adabm

• For training, we use LR images sampled from DIV2K.
• For testing, we use benchmark datasets and large input datasets Test2K,4K,8K. Test8K contains the images (index 1401-1500) from DIV8K. Test2K/4K contain the images (index 1201-1300/1301-1400) from DIV8K which are downsampled to 2K and 4K resolution. After downloading the datasets, the dataset directory should be organized as follows:

datasets
  -DIV2K
    - DIV2K_train_LR_bicubic # for training
    - DIV2K_train_HR
    - test2k # for testing
    - test4k
    - test8k
  -benchmark # for testing

Please download the pretrained models from here and place them in pretrained_model.

sh run.sh edsr 0 6 8 # gpu_id a_bit w_bit 
sh run.sh edsr 0 4 4 # gpu_id a_bit w_bit 

sh run.sh edsr_eval 0 6 8 # gpu_id a_bit w_bit 
sh run.sh edsr_eval 0 4 4 # gpu_id a_bit w_bit

• set --dir_data to the directory path for datasets.
• set --pre_train to the saved model path for testing model.
• the trained model is saved in experiment directory.
• set --test_own to the own image path for testing.

More running scripts can be found in run.sh.

Our implementation is based on EDSR(PyTorch).

• parallel patch inference

If you found our implementation useful, please consider citing our paper:

@misc{hong2024adabm,
      title={AdaBM: On-the-Fly Adaptive Bit Mapping for Image Super-Resolution}, 
      author={Cheeun Hong and Kyoung Mu Lee},
      year={2024},
      eprint={2404.03296},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Email: cheeun914@snu.ac.kr