Using stories is a great idea, but I'm not sure how doable it is to generate images that suggest a story (certainly, harder than cute variations on robots); in addition, those stories have to be memorable i.e. make sense, as stories. I think that's approaching a hard problem, maybe even hard AI.
But I like the idea. Perhaps the image problem could be met by combining it with a bio (giving a story); and the "make sense" problem could be addressed by a story grammar (following the hero's myth as a template, with recursive and optional parts), written using templates consisting of canned pseudo-english sentences with gaps filled by a set of names, objects, places that play the various roles in the "hero grammar": the key, the sword, the grail, the shadow, the mentor, the ordinary world, the special word, various thresholds - perhaps some word generation for place names. If the story made sense, as a journey, it might be memorable, even stirring in an awkward way, despite all the grammar/template/presets...
Of course, maybe I'm wrong: "<color> <warriortype> with <injury>" is already fairly rich. Even, extending it to incongruous occupations (surgeon, nurse, motorcyclist, developer). Perhaps, like theatresports, just starting points of a place, occupation and problem is enough to suggest a story to the user?