We can use the preview images. Once selected, we prompt the model on what else in the available library goes well with what's in the room. We could probably do this once for each model and cache it to save on compute.