Two different jobs for a neural network
Most of the AI that came before the current wave of image and video tools was discriminative. A spam filter looks at an email and outputs a label: spam or not spam. An image classifier looks at a photo and outputs a category: cat or dog. These models learn a boundary between classes in existing data. They never had to produce anything new, they only had to sort what already existed.
Generative models flip the task entirely. Instead of learning to separate categories, they learn the underlying distribution of the data itself, meaning they learn what makes a photo of a cat look like a photo of a cat at a statistical level: the textures, the shapes, the way light falls on fur. Once a model has learned that distribution, it can sample from it to produce a brand new image that never existed in the training set but is statistically consistent with what a cat photo looks like.