We present a theoretical framework which unifies posterior and end-to-end guidance for flow/diffusion models.

We learn a generative model of the Pareto set that can be conditioned on subjective preferences, without retraining, for online multi-objective optimization tasks on discrete/mixed spaces.