Retain synthesized output structure from different labels.

[ExponentialML]

This is less of an issue and more of a discussion. First of all, great work. I've trained a few models and it works great, but was wondering if a certain functionality exists.

Is it possible to do motion re-targeting with this repo, similar to First Order Motion Model??

I know that you can create labels on a single photo from a video, train it, then drive the synthesized image with that video sequence. The problem is, what if someone wanted to drive the synthesized result with something that has a completely different structure from what it was trained on? The synthesized video will conform to the user created labels, thus distorting it and making it look strange.

Let me explain using your starfish example. The training pair is a labelled set and a starfish. If I labeled a starfish that was bigger than the one it was trained on, the synthesized result would conform to the bigger one, thus distorting the one it was trained on. This is the case scenario you don't want.

Does this functionality exist, or are there any plans of this implementation. Thanks!