The prompt was "A cute otter in a rainbow whirlpool holding shells, watercolor"

Seems like the otter should be holding shells, the way a normal human parses it.

The tool showed the otter 'in holding-shells', which are shells that hold otters apparently. Also some random shells strewn about, as the technique is sensitive to spurious detail sprouting up from single words.

Until the tool permits some kind of syntactic diagramming or so forth, we'll not be able to control for this.

Just the other day here, I saw a picture of a fork and some plastic mushrooms. The prompt was 'plastic eating mushrooms' which was ambiguous even to humans. The tool chose to illustrate the subclass of mushrooms 'eating-mushrooms' (as opposed to poison mushrooms or decorative mushrooms I suppose) made of plastic.

When we're playing around this can seem whimsical and artistic. But a graphic designer might want some semblance of control over the process.

Not sure how a solution would work.