The meaning of noise regarding “z” can be confusing. So let’s avoid this term for a moment and simply say what distribution we want to use for z in our sampling. Some people use uniform distribution and then some people use the normal distribution. DNN can be simply viewed as a function estimator. From this perspective, it can carry out extra conversions if needed subjects to the model complexity and whether it is trained probably. I think you can start with the one that you believe will model the latent factor better and start iterations from there.