Couldn’t the model be trained to output 360 panoramic images and “just” cut to a smaller format for useable images. Then the share of things in the frame / things in the world would be a bit better.
maybe a bit, but 360 panoramic image still doesn't really address things out of direct line of sight. e.g. in a house, go through a door to a room, come back it's a different room :P
or even if you have a vast line of sight but a guy goes behind a tree
The hair changing to a helmet in the motorcycle clip for example, is just the head passing behind the other person. So not clear this approach would improve the issue much at all.
or even if you have a vast line of sight but a guy goes behind a tree