Normalizing Flow Penalization

Question

I am looking to fit a normalizing flow, specifically a Masked Autoregressive Flow model. However, this model leads to high variance on lower dimensional, less complex data. I am using a neural network to parameterize the scale and shift components of the MAF. I think we need to smooth the loss surface, i.e., the log likelihood of the normalizing flow. Is there any literature outlining good ideas or what would be a good starting point? More generally, are there any papers that have a solid theoretical foundation to defining a penalty that can smooth something like a normalizing flow?