Closed pnbao closed 3 years ago
Faces should be pre-aligned before fed into the encoder. Please refer to this link.
Our encoder is based on a StyleGAN model trained on FF-HQ with 256 resolution. The official model is with 1024 resolution. So, they may mismatch. Furthermore, we use a StyleGAN generator that uses different w
s for different layers. This is a little bit different from the original StyleGAN generator that uses the same w
for all layers.
Thank you for your great work!!
I have tested the model with some random images from internet, the encoding results and manipulation results is not really good as compared to other images included in your examples folders.
How can I improve the model for better encoding images like those?
For the interpolation, can we adjust the copied style range to interpolate coarse style, middle style and fine style?
By the way, the model trained based on StyleGAN, so can the encoded latent can be fed into a StyleGAN generator?