Why train a policy in imagination (Dreamer)

Latent world-model agents like Dreamer train the policy mostly on rollouts 'imagined' in the learned latent model rather than on real environment steps. What is the main benefit?