How to generate unseen attribute combinations in the paper?

[buaacyw]

Thanks for your work!

I'm curious about how to generate unseen attribute combinations as mentioned in the paper. Do you generate it by fixing one single latent code β and changing those attributes?

In my opinion, this shouldn't work: since the deformed coordinates (K(x) in the paper) depend only on β, if you have changed the canonical space by changing the value of those attributes but K(x) remains the same, the K(x) may not be feasible on the canonical area that has been changed.

So I think you may have used several latent codes β to generate unseen attribute combinations. But this requires the images corresponding to those β to be well aligned, right? So this brings some constraints to this work?

Thanks!