search

rolux / stylegan2encoder

StyleGAN2 - Official TensorFlow Implementation
http://arxiv.org/abs/1912.04958
Other
380 stars 71 forks source link

real image mix #5

Closed rainsoulsrx closed 4 years ago

rainsoulsrx commented 4 years ago

hi, thank you for your work, I have a question when doing style mixing between two real images, In stylgan1, I chose range(8,18), which performs well, but in stylegan2, I use the same param, the results turn out to be bad. I wonder, the scale in stylegan 1 and 2 is represent different meaning?

dimitry12 commented 4 years ago

@rainsoulsrx I suspect, semantic content of style-layers changed in StyleGAN2. Were you able to find the range which works for StyleGAN2?

b4nn3d commented 4 years ago

can you share an how to for create and mix images from npy files?

rolux commented 4 years ago

Style mixing works as demonstrated in the original StyleGAN2 repo. Nothing has changed there. See this comment for orientation.

If, however, you're mixing latents from encoder/projector results, this may not work as expected. In case the distance of your latents to w_avg becomes too large, they lose their "semantic content" altogether. For more, read the thread linked above.

shahik commented 4 years ago

@rainsoulsrx Plz share the code for npy files style-mixing ?