About the manipulation layer parameter.

[berylsheep-up]

Hello, thanks for the excellent work! I found that the boundaries learned by interfacegan contains the manipulation layer parameter. But i don't found explanation about the manipulation layer in the interfacegan paper. So I wonder is the manipulation layer learned by experience? Or the boundaries are learned specifically on the manipulation layer? Thanks a lot!

[zhujiapeng]

Once a StyleGAN is well-trained, you will find different w control different image semantics, you can refer the style-mixing part in original StyleGAN paper. In our paper, we first search all ws (i.e.,14 ws for 256x256 images) by layerwise manner. And then we just keep the ws pertained to a specific semantic and discard the rest.