Improving T2 distribution estimation using deep generative priors