low-nfe

D2PO proposes a dynamic preference optimization framework that aligns diffusion sampling policies with perceptual quality using direct preference optimization, outperforming regression-based methods under low-NFE constraints.