17 hours of training later and still no convergence in sight.
It's either giant bug-eyed monsters breaking through the wall or just the typical StyleGAN2 wrinkles of a model that needs a lot more training.
20 more hours of training and still no discernible structures.
Speculative titles:
"Goethe's Corpse Contemplating His Torn-out Heart"
"God Designs the Banana"
"The Bug-Eyed Monsters' Barbecue"
With a little bit of good will I can recognize a giant killer robot and Armageddon by meteor storm and flying monster slugs here. Still no humans in sight.
Trying to continue the training with a different γ value. So far it is still trapped in slug-face territory. I fear the data is just too dissimilar and unaligned to ever reach a "normal" look.
For a moment I had my hopes up, but at about 3 am the whole model collapsed.
It's almost as if I have never been away.
https://twitter.com/quasimondo/status/878539629628149760?s=20
The latent wormholes in these examples are almost acceptable.
@Quasimondo Here I was thinking these were cut scenes from Nausicäa
The third image here (cbc8d… jpg) is the best so far in the following regard:
I’ve noticed with GPT-2 that it’s really bad at typesetting basics, like smart quotes or non-smart-quotes, single quotes or double quotes etc. (Maybe that’s on me for not sorting that out in my corpus.) And some of the early ones you posted feel like they were rended quite mismatchedly— more like collage than the work of one hand. That’s how GPT-2’s text feels to me too, tbh.
Server run by the main developers of the project 
It is not focused on any particular niche interest - everyone is welcome as long as you follow our code of conduct!
The model is still in the proto-stage of the training, so this will have to do for now.