//=time() ?>
Done! Pushed into a rgb branch, in case things go awry. If anyone has a good RGBA dataset to help me test this out I can then merge it into main (@duskvirkus
@dvsch @makeitrad1) https://t.co/UJdIb8xHzu
Images can be generated as of now, like this one with a scratch RGBA network https://t.co/Sw6AREO1DU
The Tree of Life w/@midjourney. Inspiration taken from The Fountain by @DarrenAronofsky
My last credit, but I think I'm on the right track here. Will come back later...
If you don't resize the octaves, then you get vastly different results (same order as before). Tied to this, other layers become interesting, so perhaps more iterations is warranted when not resizing the octaves (will depend on the model and what is being sought out to display).
Discriminator Dreaming using @Buntworthy's #StyleGAN3 Wikiart model, with the following layers: b16.conv1, b8.conv0, b4.conv, and b4.fc, respectively. Starting from Perlin noise (seed 0), 20 it each, resizing all 5 octaves back to 1024 res, lr=0.01; D has learned quite a bit!
@aicrumb Do you know exactly which line of code ot exits, or is it not always the same line?
Discriminator Dreaming with layers D.b8.conv0, D.b8.conv1, D.b4.conv and D.b4.fc, respectively. Were the spirals curly hair all along, overrepresented in the MetFaces dataset? https://t.co/YQuYOaM4Pb
@hexorcismos Es el primer tick, 4kimg. Si se ve raro, de ahi empiezan a salir dobleces/patrones. Pongo 8, 16, 32 y 64 kimg: