Blending FFHQ faces with KaoKore pre-modern Japanese art faces
source: https://github.com/rois-codh/kaokore https://arxiv.org/abs/2002.08595from projector.py
referenced:
https://github.com/justinpinkney/stylegan2/blob/master/blend_models.py
https://github.com/XingruiWang/Animefy/blob/master/blend.py
Model blend:
- model trained for 500,000 iteration on FFHQ
- model further trained for additional 180,000 iteration on KaoKore
Latent space is found by projection on blended model,
But wwould it have been better if projection of FFHQ trained StyleGAN2?