Train an encoder just using latents

[lifematrix]

Hi, thanks for your excellent work. I have read and run your code, and it inspired me a lot.

I have a question, if it is possible to train an encoder that outputs latents only, i.e, without noise?
Maybe the quality of the reconstructed image is not perfect, but it is more convenient to do some image morphing
like Image2Stylegan. Does it work just setting the argument --disable-update-for noise?

Thank you very much.

Steven

[Bartzi]

Hi,

tanks for your kind words!

Yes, --disable-update-for is exactly the correct way to achieve what you want. If you say --disable-update-for noise you will train a model that should be able to predict a latent vector with results close to the results from the Image2Stylegan paper.

[lifematrix]

Hi, thanks for your nice reply.
I will try it later.