[taylorius]

This is true, though machine learning models will learn typical scene structures, alongside the rendering parameters. It is also common to use more than one source image (from a different viewpoint, a bit like photogrammetry).