> In fact how do you know DALL-E actually created them, and did just regurgitate some it was trained with?
I think DALL-E is not released so researchers are unable to take it apart yet, but this question was already researched a lot in the context of other generative models and so far they really did generalise (assuming a well trained model, the can overfit).
I think DALL-E is not released so researchers are unable to take it apart yet, but this question was already researched a lot in the context of other generative models and so far they really did generalise (assuming a well trained model, the can overfit).