There are already efficient models that can generate single images in real time as fast as you can type the prompt. I wonder if you could tweak this approach to instead generate a page full of static thumbnails on submitting a query (using different seeds for each result), and then upscale the desired image when clicked.
It won’t be the same image once you upscale it. Moreover, while the fast generators do exist, they’re lower quality. That doesn’t mean just resolution—it also means comprehension and prompt understanding, so upon running the same prompt through a better AI you’ll get something entirely different.
