The intuition behind this should be clear to anyone who's ever looked at the fakes####.png files the base StyleGAN2 repo spits out during training, or scrubbed through the images on Tensorboard. You're seeing different snapshots of the model through time as it trains.

So, obviously, doing a weighted average of the two endpoints should allow you to smoothly transition between the different domains. This wouldn't work on two models both trained from random starting points though, because there isn't a clear path between the two endpoints.

Here's a version with fixed latents all the way through. Only the model weights are being changed.

@theshawwn tells me I need to plug my Patreon more, but asking people for money is scary. patreon.com/arfafax