Yeah, this kind of thing was part of the subject of my PhD, first postdoc, and ongoing scientific work. The question is how to produce generative models and inverse-inference algorithms that are powerful enough to work in tens to hundreds of milliseconds in high dimensionality :-/