BigGAN seems to be the best for supervised GANs, while StyleGAN seems to be the best for unsupervised GANs. Has anyone tried adding labels (class information) to the latent space in a StyleGAN generator? I wonder if StyleGAN could be used with ImageNet.
5
11
1
80
Replying to @jm_alexia
I did this for binary attributes (CelebA-HQ dataset), and this seems to work for most attributes (smile, gender, hair color, glasses, etc.)

Apr 23, 2019 · 12:35 PM UTC

1
4
This tweet is unavailable
I was conditioning on a bit vector of all 40 CelebA attributes. The code is straightforward: 1) save -labels.npy file with binary attributes in a bit vector while creating CelebA-HQ TFRecord files 2) train StyleGAN with 'cond' option.
1
1
5