Domain Transfer in Latent Space (DTLS) Wins on Image Super-Resolution - a Non-Denoising Model

EDIT: We are working on a better version base on this concept.

This is the official code page of Domain Transfer in Latent Space (DTLS)

Please refer the paper on arXiv from arXiv

Citation:

@misc{hui2023domain,
      title={Domain Transfer in Latent Space (DTLS) Wins on Image Super-Resolution - a Non-Denoising Model}, 
      author={Chun-Chuen Hui and Wan-Chi Siu and Ngai-Fong Law},
      year={2023},
      eprint={2311.02358},
      archivePrefix={arXiv},
      primaryClass={eess.IV}
}

Update

We are building this GitHub page and will update more information later.

Dataset

To prepare FFHQ dataset, you can follow: FFHQ

Training

Follow the command lines below

32 --> 512

python main.py --mode train --hr_size 512 --lr_size 32 --stride 16 --train_steps 100000 --save_folder '32_512_s16' --data_path 'your_dataset_directory' --batch_size 2

16 --> 128

python main.py --mode train --hr_size 128 --lr_size 16 --stride 4 --train_steps 50000 --save_folder '16_128_s4' --data_path 'your_dataset_directory' --batch_size 32

Evaluation

You can download the pretrained model from Google Drive