Closed xiaoqian-shen closed 2 years ago
Hi, thank you!
Yes, the current implementation uses conditioning only in DiscrEpilogue
: we decided to remove it from DiscriminatorBlock
, since it considerably increases the training time (~30% as far as I remember) and does not help that much (in terms of raw metrics at least). We also updated the paper accordingly a couple of weeks ago.
Thanks for your reply!
Hi, Great Work!
I have a question about the discriminator. The current discriminator is conditioned on δ_{i}^x only in DiscrEpilogue, but not the DiscriminatorBlock. Please kindly enlighten me if I misunderstand anything.