Question regarding the difference in fade in with the official implementation about style-based-gan-pytorch HOT 6 CLOSED

I think it is not very meaningful. It is just the side effect of changing initial size 4x4 to 8x8.

from style-based-gan-pytorch.

This implementation also alternates fade in and stabilizing stages.

from style-based-gan-pytorch.

Ah I see. You took the min of 1 and the weight. Closing this now.

from style-based-gan-pytorch.

On a second thought, in the ProGAN paper, it seems that they are doing stabilizing and then fading in. Here since we have alpha = min(1, 1 / args.phase * (used_sample + 1)) and used_sample grows from 0 up, wouldn't it reverse the order (i.e. fade in first and then stabilize) ? I think this kind of makes a difference at the very first level of resolution. Thanks!

from style-based-gan-pytorch.

Yes, it's different. But after initial resolution it will be same.

from style-based-gan-pytorch.

Thanks! If I understand correctly, we start at 8x8, and at this time we fade in with the 4x4 right? My concern is that since the 4x4 is not yet learned (as we started from 8x8), would the fade-in at this stage be advantageous at all?

from style-based-gan-pytorch.