genforce / idinvert

[ECCV 2020] In-Domain GAN Inversion for Real Image Editing
https://genforce.github.io/idinvert/
MIT License
461 stars 65 forks source link

How to obtain projected direction given pretained boundaries? #33

Closed fungtion closed 4 years ago

fungtion commented 4 years ago

I tried to get age/eyeglass boundry using the method discribed in the paper "interfacegan", e.g. _n = n_age - (n_age^T n_eyeglass) neyeglass with the boundary file age.npy and glass.npy provided, but when I manipute age/eyeglasses attribute, the images hardly changed as expected. If there is any difference when manipulating w and wp?

ShenYujun commented 4 years ago

This phenomenon is discussed in InterFaceGAN, which is that conditional manipulation does work well on StyleGAN W space. The reason is that all boundaries in the W space are already orthogonal. You can verify this by measuring the similarity between n and n_age you've mentioned. They should be very close to each other.

fungtion commented 4 years ago

Thanks. Is that means better distanglement in w space leads to the orthogonal boundaries?

ShenYujun commented 4 years ago

Yes. But the disentanglement in W space is not perfect. That is why the entanglement between "eyeglasses" and "age" can still be observed. How to find more disentangled boundaries is still worth exploring.

fungtion commented 4 years ago

I see, thank you.