[gwern]

The original Karras et al 2018 paper did both cars and cats, which aren't faces. Worked very well, unsurprisingly. (ProGAN also did well on those, though it was the faces everyone paid attention to.) Look at the samples in the paper or the Google Drive dumps, or at the interpolation videos have posted on Twitter.

Aside from the original work, on Twitter, people have done Gothic cathedrals very well, graffiti very well, fonts very well, and WikiArt oil portraits not so well. On Danbooru2017 full anime images (linked in my thread), one person has... suggestive blobs but has only put 2-3 GPU-days into it and we aren't expecting much so early into training. skylion has been running StyleGAN on a whole-body anime character dataset he has, and the results overnight (on 4 Titans) are pretty impressive but he hasn't shared anything publicly yet.

[lucidrains]

Great job on the Danbooru training! I've been following you on twitter and machinelearning for the longest time haha

[gwern]

Thanks! The wait on training is killing me, though. I've been doing large minibatch training to try to fix the remaining issues in the anime face StyleGAN and it's frustrating having to wait days to see clear improvement. Checking GAN samples is so addictive and undermines my ability to focus & get anything else done. I'm also eager to get started on full Danbooru image training, which I intend to initialize from skylion's model - whenever that finishes training...

(Who says we aren't compute-limited these days?!)

[lucidrains]

Haha, having to work around the computation limits are welcoming! It feels like building web apps back in the late 90's again. These days we have so much memory and disk space at hand it doesn't even feel like a challenge anymore.

That is, until Graphcore delivers their IPU.

[gwern]

I forgot one failure case: a few hundred/thousand 128px pixel art Pokemon sprites. StyleGAN seems to just weakly memorize them and the interpolations are jerky garbage, indicating overfitting. (No GAN has worked well on it and IMO the dataset is too small & abstract to be usable.)