Stylegan2 trained on imagenet (1.2 million images, unconditional) at 512x512 resolution. Trained from scratch on a TPUv3-512 in just 45 hours. Training speed: 3k images per second. Tensorboard logs: tensorboard.dev/experiment/y…

8:18 PM · Mar 26, 2020

12
22
4
112
2,980
1
0
0
6
141
2
0
0
5
125
1
0
0
3
108
1
0
0
4
66
1
0
0
4
79
2
1
0
6
238
Here's the model, in pkl form: battle.shawwn.com/sdc/styleg… Have fun!
1
0
2
12
Latest pkl: battle.shawwn.com/sdc/styleg… Step 533504, roughly 2x more training. The samples look much more interesting.
2
0
1
17
Latest pkl: battle.shawwn.com/sdc/imagen… Step 1,394,688. The samples look interesting, but it's probably no longer improving. We think stylegan is underfitting imagenet, so we're going to try increasing parameter count (or switch to biggan).
2
0
0
10
Replying to @theshawwn
Awesome results and awesome compute power!
0
0
0
2