Yes, training data continues to be the only real-world reference for GANs and the output is as good as the training set. There’s no escaping this fact. The ingenuity of these models is from a different perspective.

Compared to traditional ML models, learning for GANs is very different since the direct input here is just random noise. Training data is purely used in the feedback loop of learning through the discriminator, hence GANs take in broader learnings from this data. This enables them to get ‘creative’ and produce completely new output, though still inspired by the data fed in. And, when you feed in representative data, even in low volumes, thats where GANs display their mastery and do way more than most other models.