How would you implement your solution technically though? There isn't a reference set of images in the model, it's just weights.