Robert Luxemburg

Yes this both makes sense. I'll try and add it to the next release. There are going to be a few major changes, mostly for video encoding.

Yes, that would be a huge improvement. But isn't ResNet initialization included with pbaylies' [fork of the encoder](https://github.com/pbaylies/stylegan-encoder)? That one is well-maintained, and I'm not trying to duplicate it. It's...

Other than that, if you pull the latest changes, you can use the projector, which may be a better choice anyway.

@SimJeg: Ok, I've looked at this more closely ;) Three ResNet initializations below, from left to right: @pbaylies, StyleGAN V1; @quasimondo, [via twitter](https://twitter.com/quasimondo/status/1210984685226119173), using your (18, 512) ResNet above; **myself**,...

@SimJeg: I have trained a ResNet, and will post some results shortly. [resnet_18_20191231.h5](https://rolux.org/media/stylegan2/resnet_18_20191231.h5) (best test loss: 0.04438) If you get a `TypeError: Unexpected keyword argument passed to optimizer: learning_rate` you'll...

It took me a while to appreciate the fact (thanks to @pbaylies for the insight) that encoder output can have high visual quality, but bad semantics. The W(18, 512) projector,...

@pbaylies: I'm outputting into W(18, 512). It's SimJeg's code, with the `w_mix` arg passed in the last `get_batch` call. The initial advantage, averaged over 100 runs, is not exactly zero,...

@Quasimondo: To get a better sense of which layer does what, I used to render these style grids: https://youtu.be/hx51TqJ_adE Top row: style target, midpoint, style source, 0-3 mix (coarse styles,...

@Quasimondo: The changes you made to the ResNet training process sound interesting, I'll try to find out how much of that I can reproduce. I'm a still a bit reluctant...