Works pretty good already, but needs a little bit more cleaning up of the dataset. Still, OK for now :-)

Trained with 164 screencaps.

The model tends to often produce portrait shots by default, so using terms like “wide shot” or “full body view” help in this case^^