z_dim for 512x512 model

[ajbrock]

z_dim for 512x512 model

[ssnl]

@ajbrock any update on this?

[ajbrock]

Hi Tongzhou,

This looks like a typo or a mismatch somewhere between the paper and that particular model setup--we'll take a look and see about updating. FWIW the model is quite robust to the choice of z dimension (in the paper we mention that we are able to train with very low dimensionalities, on the order of 32 dims IIRC) so I wouldn't read too much into this; there's no explicit design choice to e.g. make the z-dim smaller for 512x512 than 256x256. BigGAN-Deep uses a 128-dim z for all resolutions, for example.

Andy