The model appears to have overfit a bit (seed 535 is blatantly Vivillion: https://twitter.com/ak92501/status/1492560750103121920 )
Which is interesting because GANs tend not to overfit.
Traditional GANs typically have a different problem called mode collapse, in which the generator learns one (or a few) outputs that the discriminator likes. Conditional GANs are given a condition as input and have more chance of overfitting, as the loss function is a combination of the GAN output and a supervised loss based on the condition.
Re:Pokemon GAN
by Starlevel001 on Monday February 14, 2022 @07:12PM
(#30340579)
Vivillon just has 20 different forms, so it makes sense it would be more common.
Re:Pokemon GAN
by Starlevel001 on Monday February 14, 2022 @06:47PM
(#30340288)
0 is closer to Sivally than Arceus.
Some manual seeds:
4552 is a corrupted Talonflame
2518 is another Vivillion, and so is 3330, and so is 2216, and so is 3373, and so is 3151 (looks like all Vivillion forms were put in)
3129 is horrific, it's like Glaceon has been stretched into a Wurmple and put on top of a Dracovish head and an unknown body
1742 is a mutilated Slurpuff
2410 is a Tapu mixed with... Corsola-G?
3805 is obviously Magmortar with some dragon head
3546 is clearly Reshiram (with genie arms?), and 1878 is mega Tyranitar (both down to the colours)
2194 is Gible x Joltik
1576 is radioactive Alcremie
2000 is straight up Roselia
- 0 as Arceus + Shaymin S
- 1 as Lickilicky + Espurr
- 10 as Kirlia + Porygon 2, don't know where the colors are from
- 20 as Shellos + Froslass
- 30 as Roggenrolla/Oddish/Poliwag
- 42 as Garchomp + Trevenant
- I see the Kyurem in 60 but no idea where the tail is from
- 102 as Dragonite + Hitmonlee