How to Train on a Single Image

Recently, there have been some styles of decoupling work. But the model cannot be generalized. I am trying to use pretrained methods to obtain the style space and laten code W of images outside of the ffhq dataset. I would like to know if this is feasible