A fast version for feature conformal prediction.

We show that discrete representation of images improve unconditional and compositional generation

Steering given distributions towards ideal distributions, where fairness and accuracy are not at a tradeoff.