Questions about stage 1 & 2 of Progressive Fine-Tuning about dualstylegan HOT 4 CLOSED

1. Could I use the pretrained model you provided (showing below) on several different dataset that I collected them myself.

Yes. You don't need to pretrain another model.
This model is pretrained on FFHQ, independent of the extrinsic styles, thus it can be applied to any extrinsic styles.

2. I only find out stage 2 codes in pretrain_dualstylegan.py. Could you showd me codes of Stage 1 in pretrain_dualstylegan.py?

Stage 1 initializes the extrinsic style path, which I directly put in the model initialization function __init__():

3. I am not understand stage1 very well from paper. In stage 1, you trained color transform blocks with style mixing FFHQ images?
   The purple hair in following figure you showed it on purpose. In most cases, it is regular hair color (black, yellow, ...) when we train model in stage1, right?

Stage I only initializes the extrinsic style path to make the initialized model can naturally follow the original StyleGAN to perform color transfer.
Yes. the purple hair is just for easy visualization of color transfer. In most cases, the sampled faces have regular hair color.

4. In stage 2, you mentioned you sampled z2~ from {z2,E(g(z2))}. Could I know why you do it this way instead of only z2 or E(g(z2)).

Because we want the extrinsic style path to work for both the randomly sampled latent code or latent code embeded from style image. These two kinds of latent codes have different distributions.

from dualstylegan.

Thank you for quick and clear reply.
Based on your reply, Stage 1 doesn't involve any training, am I right?

from dualstylegan.

Yes. No training is required.

from dualstylegan.

Thank you for answering

from dualstylegan.