Slayerxxx

I have the same question...

I guess the conv in recursive unit is shared

Hi @xinntao Have you ever tried to use LPIPS as perceptual loss in ESRGAN? Are there any differences between your VGG based method and LPIPS?

> > @Feiyu-Zhang > > > > 1. 5k is not enough, at least, > 300k. > > 2. Your figure indicates there must be something wrong, which is not...

You can use different kernel and noise level at different position and make parameters change continuously

+1 Can't find the implemention of SE_ResNet50DCT

OK Thanks a lot