csxmli2016 / DFDNet

Blind Face Restoration via Deep Multi-scale Component Dictionaries (ECCV 2020)
915 stars 215 forks source link

VGGFace loss details. #75

Open funnylittleman opened 3 years ago

funnylittleman commented 3 years ago

For perceptual loss do you use features obtained by VGGFeat class(same features as in main model) or some other model? In paper it is written that you use first 4 conv layers of VGGFace for perceptual loss. Did you mean last conv layers of first 4 blocks of VGG (as it is usually done in perceptual losses) or you mean literally first 4 conv blocks?