We propose LeMiCa, a training-free acceleration method that improves content consistency in diffusion-based video generation by minimizing worst-case global errors through lexicographic minimax path optimization