This is great! Now that the model can be used in PyTorch, I've starting playing with @AydaoAI's anime StyleGAN directly guided by CLIP. Starting slow by searching for Asuka by name in the latent space.
Okay, my fork is here: github.com/nagolinc/stylegan… And an example notebook is here: github.com/nagolinc/notebook…

9:24 PM · Jan 24, 2021

6
5
2
41
Doesn't get the face right, but it's going somewhere. 😄
1
0
0
2
Can it generate "a dog"? Not yet, at least.
0
0
1
6
Replying to @advadnoun @AydaoAI
got this for Kermit the Frog
1
0
0
11
wtf
0
0
0
2
Replying to @advadnoun @AydaoAI
hatsune miku
0
0
0
3
Replying to @advadnoun @AydaoAI
Are you doing some form of gradient descent? Didn’t think that’d work too well, sweet if it does
1
0
0
2
Yeah, this is maximizing CLIP's text & image encodings cosine similarity by learning the latent vector given to the StyleGAN network.
1
0
0
3