IdolGAN project is inspired by our pure passion which is restoring our favorite K-pop idols to high quality images. So we googled about it. But many SR models like Topaz can't generate an image we are satisfied with. We think models trained by FFHQ or Celeb-A never restore our favorite idols to high quality properly. Therefore, we gather many high-quality K-pop Idol images from the internet. The most important part is the face. The satisfactory thershold of face image quality is higher than the other parts of human images (etc. arm, legs, or feet). So we collect about 6,000 K-pop girl group images at least 512*512 size, even the face alone! We build these images to a dataset called KID-F).

In this repo, we will archive our journey to restore perfect high quality K-pop idol face images. We finished our first step which is proving our hypothesis: fat pops and wrinkled old women in FFHQ, Celeb-A, etc can't generate our beautiful idol's face. ❤️

We used HiFaceGAN and KID-F Dataset.

• Ubuntu
• PyTorch 1.0+
• CUDA 10.1
• python packages: opencv-python, tqdm,
• Data augmentation tool: imgaug
• Face Recognition Toolkit for evaluation
• tqdm to make you less anxious when testing :)
• cv2

• cv2
• mediapipe
• tqdm

• cv2
• tqdm

We provide Model Checkpoint which trained from KID-F Dataset. We used train_dataset of KID-F for training, which takes 2 days on a single GTX 1080 Ti. We upload the checkpoint on our Google Drive.

The configurations is stored in options/config_hifacegan.py in HiFaceGAN You need to edit train&test dataset root, each(train/test) option name, and netG to 'lipspade' for using KID-F dataset.

We used PSNR, SSIM, FID, LPIPS, IDD for evaluation.

We compared HiFaceGAN's performance of KID-F test images with 5 metrics (PSNR, SSIM, FID, LPIPS, IDD). Column FFHQ is metrics calculated from pretrained model (HiFaceGAN) with FFHQ Downloaded from HiFaceGAN. Column Idol Dataset (Ours) calculated from HiFaceGAN trained with KID-F.

The metrics is calculated from KID-F test images (total 300 images, all 512*512 resolution) inference of each model and original images.

• Restore videos, not only images.
• Fine-tuning SR model to one particular idol.
• Restore other parts of idol. (like body, clothes, etc.)
• Restore background at stages or tv shows.

If you have any questions, opinions or advices, please feel free to contact to develop@edai.club

• Instagram
• Twitter

@notyet{
  title = {},
  author = {Dongkyu Kim, Donggeon Han, Hyunwook Kwon, Dain Jeong, Cheol H. Jeong},
  month = {August},
  year = {2022}
}