Looks cool! One piece of feedback I have is that I personally wouldn't want to autocomplete prompts as shown in the image you have there. I'd want something that "fixes" my prompts so that they're understood by the model. Like say if I put "Afro-Asian samurai carrying two swords in a cyberpunk environment", for it to say how I could re-word it to generate something closer to what I actually want.
That's a really neat idea, since most of the time people already know what they want to generate. I'd have to think a little more about how I'd get the training data for that. LMK if you have any ideas.
Cool! Based on your description, I thought I can give the generator an image and it'll return the prompt. Would that be possible?