We propose SING, a method that leverages natural gradient variational inference for accurate and efficient inference and drift estimation in latent SDE models.