search

anvoynov / GANLatentDiscovery

The authors official implementation of Unsupervised Discovery of Interpretable Directions in the GAN Latent Space
416 stars 52 forks source link

discover directions in full w space #9

Closed eps696 closed 4 years ago

eps696 commented 4 years ago

is it possible to discover directions in full w space (e.g. [18,512], one [512] latent per layer), instead of single tiled/repeated [512]? that provided way better results for the task of encoding ("projecting") foreign images into StyleGAN latent space, so should also work here.

anvoynov commented 4 years ago

Hi @eps696, have not tried that though I'm not quite sure it will work: the dimension seems to be too high and I suspect that the training would be too unstable.

eps696 commented 4 years ago

same feeling, thanks

anvoynov commented 4 years ago

@eps696 we also performed the rectification after the BigGAN's first linear layer. Though it worked, as far as I remember, it was worse compared to the original z.

eps696 commented 4 years ago

i'm just not sure it makes real sense to go w from z, as far as we keep the same power (dimensionality) of latents

anvoynov commented 4 years ago

it does as the w-space has different inner geometry.