Generative models are useful because we can sample vectors from them.
Discriminative models supposedely don't have this feature. But what's stopping us from sampling vectors that are "within the hypothesis"? Imagine we draw a circle around the cat vectors. And we can then randomly sample points within this circle to "generate" sample vectors of a cat.
So, by this technique: discriminative models can "generate" samples. I don't know if this technique is valid though