search

genforce / idinvert_pytorch

[ECCV 2020] In-Domain GAN Inversion for Real Image Editing (PyTorch code)
https://genforce.github.io/idinvert/
MIT License
410 stars 54 forks source link

about your own stylegan generator #18

Closed bingxumu closed 3 years ago

bingxumu commented 3 years ago

Hi! Thanks for the great work!

May I ask you some questions about your own stylegan generator 256*256?

  1. I found that your own generator's last layer of the MLP(mapping) is different from the original stylegan. You don't repeat the w code (1, 512) 14 times, and you directly learn the wp code with shape (14, 512). Will this improve the model? It seems that it makes image manipulating better?
  2. When you train the boundary for your wp code with shape (14, 512), do you apply the function of np.linalg.norm() for every (1, 512), or directly for (14, 512)? It looks like the former.
  3. How do you find manipulate_layers?

Thank you again for Genforce's kind share! It's really cool !!!!

zhujiapeng commented 3 years ago
  1. It neither improves the model nor decreases the model. Yes, it makes better image editing results.
  2. Yes, the former one.
  3. https://github.com/genforce/idinvert/issues/12#issuecomment-671313664
bingxumu commented 3 years ago

Thanks for the reply!