search

yuval-alaluf / hyperstyle

Official Implementation for "HyperStyle: StyleGAN Inversion with HyperNetworks for Real Image Editing" (CVPR 2022) https://arxiv.org/abs/2111.15666
https://yuval-alaluf.github.io/hyperstyle/
MIT License
1.01k stars 115 forks source link

face editting boundary #49

Closed yuxu915 closed 2 years ago

yuxu915 commented 2 years ago

Hi, thanks for your brilliant work! When I run face editing in HyperStyle, I find that the 3 boundaries have different shape, specifically, the boundary shape of 'pose' is [1, 18, 512], while 'smile' and 'age' are [1, 512]. I just wonder why is that? Besides, I would like to try other attributes, could you please provide other boundaries? Thank you.

yuval-alaluf commented 2 years ago

There is no difference in the behavior. If you have a boundary that is [1,512], the editing will simply broadcast to the dimension of the input latent which is [18, 512] so everything should work fine. If you want other boundaries they can be downloaded from: https://github.com/mit-han-lab/anycost-gan You can follow the instructions in their README.

yuxu915 commented 2 years ago

Thanks!