Trained on Emoji, Pokemon, or Simpsons? Answer: All of them but with the datasets swapped out occasionally. Cycling datasets is a fun way to explore different types of blended models.
Replying to @Norod78
The alchemical brewer of intellectual properties can summon great beasts, but it comes at a cost. These elixirs will make you vulnerable to a dark magic of powerful entities. The demons of copyright law.
Replying to @jonathanfly
Yeah cycling datasets can lead to some quite unusual creations. A long while ago I had a StyleGAN1 model that was gwern's anime one with a little fine art portraits that was... interesting.

Jan 15, 2020 · 2:44 AM UTC

Cycling datasets can also be a good way to make unique images beyond "push-button" style mixing. I'm working on a model that's "abstract artwork of cars" based on @MichaelFriese10's abstract art one, which was finetuned from the LSUN Car dataset. I'm training it *back* to cars.
This tweet is unavailable
Yes that's right. The results get a bit more horrific if you keep going