Comment by dfdydx

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.