Rare and extreme events are hard to predict because they occur infrequently and may be absent from available data. Here, authors introduce a method that uses statistical knowledge to generate plausible unseen extremes even when none appear in the training data.