Did Apple say anything about what training data they used for their generative image models?