It does require each feature to be part of the training dataset to compose the result. Even better, the state-of-the-art stochastic models can output identical data but you can completely ignore the original work’s license. Yeah science, bitch!