In score-based generative models, the score is the gradient of the log data density. It guides the sampling process to generate realistic, high-fidelity samples.