search

VinAIResearch / HyperInverter

HyperInverter: Improving StyleGAN Inversion via Hypernetwork (CVPR 2022)
https://di-mi-ta.github.io/HyperInverter/
Apache License 2.0
117 stars 13 forks source link

It seems exits much noise on face in some images #6

Open tengshaofeng opened 2 years ago

tengshaofeng commented 2 years ago

Hi, thank you for your great work. When I test on many images, I found comes several defect as follows：

  1. It seems exits much noise on face in some images(zoom in) orgin image: tbqorg (2) invert image: tbqinv (2)

  2. It seems sensitive to hair on the face: origin image:
    tbqorg (3) invert image: tbqinv (3)

Can you give me some advise to solve that?

Mayorc1978 commented 1 year ago

Have you solved that? I strongly think that even if the output is set to 1024, actual output without distortion/noise artifacts will probably happen only for the size of 256x256 set as the output size as to get 1024 size there will be an interpolation, and that will give you all those errors. After all, you see the pre-trained models specify a D size for the lightweight Hyperinverter of 128, and 256 for the large version. image