Some models classify. Others create. Generative models learn the underlying distribution of data and can sample from it, producing new examples that resemble the training set. They answer the question of how data is generated, not just how to label it.

Classical examples include naive Bayes and Gaussian mixture models. Modern approaches rely on neural networks: variational autoencoders, generative adversarial networks, and diffusion models. Each trades off sample quality, training stability, and computational cost.

Common generative models

• Variational autoencoders
• Generative adversarial networks
• Diffusion models
• Autoregressive models
• Normalizing flows

Evaluation is hard. Likelihood scores favour some families. Visual quality favours others. No single metric captures whether generated samples are useful, diverse, and faithful to the training distribution.