This is a first step towards an updated API where we pass explicit graph rather
than a scene. This is because we can no longer deduct which graph to use since
it will depend on a context.
Will happen in several steps, so bisecting will not be such a pain.