Open nguyenquyem99dt opened 3 years ago
StyleMapGAN-Light is 2.5X smaller than the original version. Stylemap resizer accounts for a large portion of the network’s size, so we reduce the number of channels of feature maps in the stylemap resizer. The reconstruction image lacks some detail, but StyleMapGAN-Light still outperforms baselines, and FIDlerp is even better than the original version.
Please see appendix B in the supplementary paper.
Thank you for your explanation.
Because I have a limited memory GPU, I just can train model with small generator. But I don't enough understand about the difference between normal G and small G. Can you explain the main difference, please?