//=time() ?>
Normal sampling reveals some dataset issues though, like cutting the tops off
Current results with truncated sampling are pretty great
Morning update, probably won't get much better without cleaning the dataset and some tweaks
Fixed various issues, animegan quality improved a lot from earlier, only at 64x64 for now. Has the "blob" issue from StyleGAN now so I'll have to try adding weight demodulation next.
Trying to train the model to produce shoulders etc too but doesn't seem to be working out too well
Testing separation by layer in my GAN.
Fixing progressively more of the latent space:
And finally, fixing all but the most high level portion only affects the colors:
Meanwhile waiting on the 256x256 outputs to improve, not doing too well at the moment