(a) requiring zero raw k-space training data, (b) being generalizable different contrasts and forward models by design, and (c) offering a built-in explanatory framework. 3/n