Re:This X Does Not Exist (2019)
by gwern on Sunday November 22, 2020 @12:59PM
(#25179444)
To be fair to TWDNE, just generating random faces is among the least interesting use-cases, and one could do better than TWDNE with some more work.
For example, here is a demonstration of how the new model averaging/transfer tricks let you convert a face from human to anime to fox to pony to furry in any direction by any percentage (summing to 100%): https://twitter.com/arfafax/status/1296084902928986113
(And since you can encode any pre-existing photograph into the human face model, you can do things like "what if Barack Obama but only 60% human and 20% fox and 10% anime and 10% My Little Pony?" Impossible-sounding, yet, there you have it.)
The pony/furry models are the same ones used for http://thisfursonadoesnotexist.com/ and https://thisponydoesnotexist.net/ , incidentally. (TFDNE/TPDNE came out much nicer than TWDNE because of better cropping/aligning, longer training runs on entire TPU pods rather than a few GPUs, and transfer learning.)
For example, here is a demonstration of how the new model averaging/transfer tricks let you convert a face from human to anime to fox to pony to furry in any direction by any percentage (summing to 100%): https://twitter.com/arfafax/status/1296084902928986113 (And since you can encode any pre-existing photograph into the human face model, you can do things like "what if Barack Obama but only 60% human and 20% fox and 10% anime and 10% My Little Pony?" Impossible-sounding, yet, there you have it.)
The pony/furry models are the same ones used for http://thisfursonadoesnotexist.com/ and https://thisponydoesnotexist.net/ , incidentally. (TFDNE/TPDNE came out much nicer than TWDNE because of better cropping/aligning, longer training runs on entire TPU pods rather than a few GPUs, and transfer learning.)
We're also still working on BigGAN for generating entire anime images. An older model: https://www.gwern.net/Faces#danbooru2019e621-256px-biggan Samples from a current training run: https://cdn.discordapp.com/attachments/693736350561861633/78... https://cdn.discordapp.com/attachments/693736350561861633/78... We had mostly given up, but recently discovered that random cropping was the reason our runs were taking forever and were pretty bad; fixed cropping like top cropping works much better.