Model do not works well for Asian faces.

[stereomatchingkiss]

As the title mentioned, I use the pretrained model on Asian faces, the faces after transformed, color of their skins become white, I guess this maybe caused by data imbalance. Do you have any suggestions if I want to retrain the model on dataset mostly composed with Asian faces.

Should I retrain the model of pSp Encoder and FFHQ StyleGAN too? Thanks

[yuval-alaluf]

You could try fine-tuning the FFHQ generator on Asian faces and then perform layer blending in order to get a stronger generator for your domain. Then you may be able to get better results (possibly without needing to train anything else).

[stereomatchingkiss]

Thanks for your helps :).

Do you mean download the model of FFHQ StyleGAN of this project(StyleGAN model pretrained on FFHQ taken from rosinality with 1024x1024 output resolution.), and finetune it with the codes of rosinality project?

[yuval-alaluf]

Yea exactly and then perform layer blending using the link I provided above.

[stereomatchingkiss]

I find out the generator weights are come from the Nvidia project, and their license do not allowed commercial use, if I use the pretrained model of that project, I guess the end product wouldn't be able to use for commercial purpose?

If this is true(can't use for commercial purpose), I would like to replaced the generator trained by this stylegan2 project, any suggestions?

Thanks

[yuval-alaluf]

I am no expert on licensing and I am not really the right person to talk to about using your work for commercial use. The repo you linked looks good and should be compatible with my code with a few changes.

[stereomatchingkiss]

Thanks, training the model, 1024 is too big for my machine, I train the generator with 512 only. Is it ok if I replace the generator by this new generator, which train with 512 input size? Thanks

[yuval-alaluf]

Yea. It should be fine to change the generator from 1024 to 512.