[marojejian]

Am i missing something, or is this just a case of "amortized inference", where you train a model (here a diffusion one), to infer something that was previously found via optimization procedure? (here NN parameters).