search

Lotayou / Face-Renovation

Official repository of the paper "HiFaceGAN: Face Renovation via Collaborative Suppression and Replenishment".
https://arxiv.org/abs/2005.05005
GNU General Public License v3.0
284 stars 49 forks source link

blue spots on my own training results #21

Closed erjihaoshi closed 3 years ago

erjihaoshi commented 3 years ago

00001-1 when I train the HIFACEGAN，I use the FFHQ dataset and degrade them as train data I set netG = 'hifacegan'(not same as your 'lipspade') ， ngf = 32 （11GB only for me） others are same as yours 。 but I got asome results as picture illuminate(about 10%), left is low quality ,middle is my results ,right is high orginal qualty image, I wanna konw why the model generate those blue spots.

Lotayou commented 3 years ago

The blue spot you observed is a typical artifact in semantic-guided synthesis called "droplet artifact", which is discussed in detail in StyleGAN2

Please also refer to #12 for a similar issue.

erjihaoshi commented 3 years ago

Thanks for your reply.