search

grey-eye / talking-heads

Our implementation of "Few-Shot Adversarial Learning of Realistic Neural Talking Head Models" (Egor Zakharov et al.)
GNU General Public License v3.0
593 stars 109 forks source link

Architecture of NN is basically BigGAN? #3

Closed aman-tiwari closed 5 years ago

aman-tiwari commented 5 years ago

From the paper, they allude to the architecture of the neural network being very similar to BigGAN. Would it be worth taking the placement of the self attention layers, resblock and normalization layers from it? (see pages 17-19: https://arxiv.org/pdf/1809.11096.pdf , the AdaIN layers could go where the batchnorm+linear is used)

castelo-software commented 5 years ago

Aha... I don't know why I didn't notice that the ResUp layers had two inputs before. But yeah, I guess that it makes sense to use the second input for the embedded vector and for the AdaIN layers to apply its style to the image inside the ResDown blocks! Good point!!

For the rest, I took the general structure of the three networks from BigGAN, so they should already look quite similar.

aman-tiwari commented 5 years ago

Did it help with results?