I think the future is more like saying to the computer what you actually want. Like "delete the apple from the picture".
Photoshop already has a primitive form of this particular action, called Content Aware Fill, where you draw a rough selection around an object you want to delete, and it will fill a plausible background in its place. Works pretty well.
https://helpx.adobe.com/content/dam/help/en/photoshop/using/...
There are certainly some uses of PS that things like caption-to-image GANs will replace. Analogously, last night I spent quite a while with the expressive, high-bandwidth, direct-manipulation interface of a box cutter cutting cardboard, for something I'd much rather be doing with a laser cutter cutting shapes generated from a constrained-optimization algorithm. At least this time I managed to only cut the cardboard.
Texture synthesis is a really interesting tool for both purposes. There are a lot of amazing papers in recent SIGGRAPHs with new algorithms for this kind of thing.
You were attacking interactive image editing by typing strings of text:
> in command line ImageMagick you say "x=100 y=200". This is not a trivial difference
But I don't think anyone really thinks that's a good idea.
I agree that direct manipulation is a great benefit, and I don't think there is any plausible alternative for designating positions and especially paths in pixel images the way incremental search provides a better alternative for designating positions in text. Vim in particular is distinguished from vi by supporting more direct-manipulation and noun-verb interaction flows, and vi differed from command-line editors similarly by supporting more immediate feedback and pseudo-direct-manipulation interaction: you would change a word by typing fscwbad ^[ rather than s/strange/bad/^M, for example. I think this is an improvement.
I'm continuously kind of appalled that we're carrying around these high-bandwidth multitouch devices all day and apparently the best way we've figured out to take advantage of this situation is one-finger scrolling of lists of pre-existing options, tap-selecting, and a simulated onscreen keyboard, plus the occasional pinch-zoom. I feel like even pictures under glass is about 10 years overdue for some new UI paradigms to take advantage of the new possibilities of the medium.
I wonder if there's a way to describe and interactively refine descriptions of areas of interest in a symbolic way—as Lisp trees if not Unix strings—maybe using existing neural networks. Certainly a speculative idea, of course.