generative model

A generative model is a statistical or neural model that attempts to learn (or approximate) the underlying probability distribution of data. A good generative model can typically:

1. Generate new samples consistent with that distribution
2. Assign likelihoods or densities to observed data points

Modern generative techniques include autoregressive models, variational autoencoders (VAEs), normalizing flows, diffusion and score-based models, and generative adversarial networks (GANs).

These generative models are trained using objectives designed to make their learned distribution approximate the true data distribution—for example, via maximum likelihood, adversarial losses, score matching, or diffusion-based estimation.

The end goal is that one can either sample from the model and evaluate or approximate densities or log-likelihoods.