I suspect that we'll end up with a split similar to amateur and professional photography, with generative models for the latter trained not on plain English but on something much more stringently structured, with ability to unambiguously specify many important but non-trivial parameters in the prompt, including for specific well-defined areas of the output etc. Probably with a GUI on top where you can literally highlight specific objects etc and adjust parameters, and it'll construct the query for the next iteration.