genforce
Problem in styleGAN w space
When I use this code by
python edit.py -m stylegan_celebahq -b boundaries/stylegan_celebahq_gender_w_boundary.npy -n 5 -o results/stylegan_celebahq_gender_w_boundary -i -s W
It outputs scary results. Maybe the latent_codes that initialized is in z space, not in w space. When I change it into w space, it works well.
Another question: If I want to mantipulate real images, does it mean I need to train new attribute scores and hyperplane for specific GAN. Is there some available models and results? Thanks：）
Hello,
I'm having a similar problem. When I use generate_images.py with StyleGAN, setting anything other than 'Z' results in 'scary', non-facial synthesized images. For example, following call results in such images:
!python generate_data.py -m stylegan_ffhq -o data/stylegan_ffhq_out -s W -n 10
What can be the problem?
Hello,
I'm having a similar problem. When I use generate_images.py with StyleGAN, setting anything other than 'Z' results in 'scary', non-facial synthesized images. For example, following call results in such images:
!python generate_data.py -m stylegan_ffhq -o data/stylegan_ffhq_out -s W -n 10What can be the problem?
Maybe the reason is that the latent_codes is initialized in z space, but feed into w space. You could try to change the source code, by mapping the z code to w code, then feeding into the network.