We propose EVODiff, a novel information-theoretically grounded framework that optimizes conditional variance in diffusion models' generative process, achieving significant gains in both efficiency and image quality without prior trajectories.