And in fact- I think something that would be really helpful is if these models shipped with another model that detected how similar an output image is to a spec

And in fact- I think something that would be really helpful is if these models shipped with another model that detected how similar an output image is to a specific image in the training set You could get this score alongside the results if you wanted to generate images but know which ones probably shouldn't be used commercially, or you could choose to filter your results by some similarity threshold