advimman
Finding mean for truncation trick
Can you please explain what this does in the notebook? Should this truncation be recomputed if one is to create more diversity of generation?
hi! To create more diversity you need to increase the value of the truncation argument in g_ema, which currently is set to 0.6 (please pay attention to the last cell in a notebook). You don't need to recompute truncation_latent.
Finding mean for truncation trick cell computes mean vector in W space and later every sampled noise vector in W space is attracted to the mean vector with power one minus truncation.
Thanks for the explanation. I noticed less diversity (than I expected) when I generated 1K images (but using a fixed truncation after it was computed in the block before-last in the notebook). I like the quality of the faces generated. It appears more tricky to project images into latent space. Assuming you guys are ahead of me on that, i would like to share/know your thoughts.
Hi! I am one of the authors who worked on this paper.
Could you give some more details on how you are doing the projection currently? We actually found that real world images invert very nicely into the latent space. The restored images have more fine details preserved such as earrings, facial hair, clothing patters, complex backgrounds. It is also possible to use multiple style vectors for each image so that different regions of the image correspond to different optimized style vectors.
-Kirill
We have used a number of approaches, a quick summary
- Similar to STYLEGAN2-ADA projection using VGG metric.
- VGG+L2+LPIPS+ID_loss (similar to pSp).
For now we working with FFHQ 256 for speed, but the reconstruction while decent is not as good as we expected. but really want to have the FFHQ1K working well. Send me email to [email protected] and I will share with you the results and we can discuss it more.
Of course I would love to see your projection results and algorithm if you can share. thanks
I see, as of right now we have been more focused on inverting images via optimization, and the encoder architecture is early stage, but I would love to see what you guys have managed to do so far!