We propose two approaches to uncertanty quantification in full wavform inversion that combines diffusion models (commonly used in machine learning) with Stein variational stochasic gradient method. The method scales well with the increase in dimension as long as the gradients can be computed efficiently.