I am wondering why dlatents_avg (latent_avg in the )aren't used in the stylegan2 network. The tensorflow version of the Stylegan2 when tracking it and using truncation with it gives a much better quality to the results.
I see the PyTorch version also supports it with the truncation and truncation_latent parameters in the forward call of the generator. I have tried implementing it myself, but am actually getting slightly worse results.
Hello, and thanks for a great work!
I am wondering why dlatents_avg (latent_avg in the )aren't used in the stylegan2 network. The tensorflow version of the Stylegan2 when tracking it and using truncation with it gives a much better quality to the results.
I see the PyTorch version also supports it with the truncation and truncation_latent parameters in the forward call of the generator. I have tried implementing it myself, but am actually getting slightly worse results.
Also, are you planning to add support for it?
All the best, Ivan