I understand what you're saying, although it could be argued that at least for some types of image tasks one would prefer something like an SVG output with layers to make it easier to edit.

For music, I think it's partly an academic question of "can we do it" rather than trying to maximize immediate practical usefulness. There's already quite a bit of work on symbolic music generation (mostly MIDI), a lot of it quite competent, especially in more constrained domains like NES chiptunes or classical piano, so a full text-to-audio pipeline probably seemed a more interesting research problem.

And for a lot of use cases, where people might truly not care too much about tweaking the output to their liking, the generated audio might be good enough; the examples were pretty plausible to my ear, if somewhat lo-fi sounding (probably because it's operating at 24kHz, compared to the more standard 44-48kHz).

In the future a more hybrid approach probably makes sense for at least some applications, where MIDI is generated along with some way of specifying the timbre for each instrument (hopefully something better than general MIDI, though even that would be fun; not sure if it's been done). I'm sure that in the near we'll see a lot more work in the DAW and plugin space to have these kind of things built-in, but in a way that they can be edited by the user.