Train with w-plus pSp encoder.
RewindL opened this issue · comments
I'm wondering if this repository can be trained with w-plus pSp encoder since original pSp is designed to predict wplus latent code from input image.
I have read the DualStyleGAN code, and I found when wplus encoder is employed during training DualStyleGAN, it will work for destylezation if I change z_plus_latent=True
ot z_plus_latent=False
, which means latent need not be projected by decoder.style (the 8-layer-MLP of stylegan2) anymore.
But when pretraining/finetuning DualStyleGAN, random styles are sampled by torch.randn limiting these styles are from Z+ space, and should be projected by decoder.style when injected into nework. So I have no plan to train with w-plus encoder, and I want to ask you if there is any solution or I must train z-plus pSp first?
Your reply will be appreciated.
DualStyleGAN can be trained with wplus encoder.
All you need to do is to revise the code everywhere it calls z_plus_latent=True
and transforms zplus code to wplus code.
And all the training and testing processes should use wplus instead of zplus.
Thanks for your early reply, I will try to train model following your advice.