Yuhao
Yuhao
Basicly, no. You need to align faces before swapping them.
Yes, we aligned profile/side faces before swapping. If you are fimiliar with face recognition, alignment only relocates faces according to the locations of eyes, the nose and two polar points...
No need, they have been aligned already.
CelebA-HQ-Mask has the masks for you. If you are interested in swapping faces on your own data, please refer to this issue #26 for more details.
Can you reconstruct your data (w/o swapping) to see if glasses appear? If so, the problem may be caused by improper GAN inversion results, and the low level latent codes...
Thanks for your attention. But sorry, we don't have a plan to implement the code on colab.
see issue #32
I didn't encounter this problem before, but https://animadversio.github.io/tech_note/debugging-stylegan2 does give a Linux solution, you may want to check it.
Firstly, can you do reconstruction?
To reconstruct the input face. You can see more details about reconstruction by `refine` function in `inference.py`. If you cannot reconstruct the input, the problem may happens in encoder or...