Hi @harskish , thank you for sharing your work. I have read you paper and have been debugging your code and there is one question I'll like to get some clarity on.
I have been working with the stylegan2, using what is sometimes referred to as w+ latent space (Abdal et al., 2020). That is, the style vector is 18x512 dimensions, with a different style vector for each generator block.
If I understand your code correctly, it uses 1x512 style vectors. If necessary, they are repeated, as e.g. here:
I do see code in your repository, where there is a check on the length of the style vector, for one global vector, style mixing, and the 18 style vectors.
However, I don't see a way of configuring the scripts (e.g. interactive.py) such that they would sample a 18x512 style vector.
Unless I'm missing something and this can be done. Is the reasoning behind this that running a PCA on a 18*512 dimensional vector didn't produce stable results?
The reason I'm asking is because I have fine-tuned a stylegan2, such that it uses "w+" style vectors, but I don't see a way to use your approach to investigate it. Any suggestions welcome!
Hi @harskish , thank you for sharing your work. I have read you paper and have been debugging your code and there is one question I'll like to get some clarity on.
I have been working with the stylegan2, using what is sometimes referred to as w+ latent space (Abdal et al., 2020). That is, the style vector is 18x512 dimensions, with a different style vector for each generator block.
If I understand your code correctly, it uses 1x512 style vectors. If necessary, they are repeated, as e.g. here:
I do see code in your repository, where there is a check on the length of the style vector, for one global vector, style mixing, and the 18 style vectors.
However, I don't see a way of configuring the scripts (e.g. interactive.py) such that they would sample a 18x512 style vector.
Unless I'm missing something and this can be done. Is the reasoning behind this that running a PCA on a 18*512 dimensional vector didn't produce stable results?
The reason I'm asking is because I have fine-tuned a stylegan2, such that it uses "w+" style vectors, but I don't see a way to use your approach to investigate it. Any suggestions welcome!