Agree, you can even see on their input/predicted/target examples that the created/invented data is off enough from ground truth to be in some cases unsuitable for photo interpretation.