Stage II —Geometry-aware preference optimization
ground-truth camera trajectory
Input
GT poses  ·  input video
Video-DiTafter Stage I
sample 5 seeds
Seed 1
COLMAPSfM
recovered camera trajectory, accurate
small pose error preferred
Seed 2
COLMAPSfM
recovered camera trajectory, drifting
large pose error
rollouts ranked by recovered camera-pose accuracy  →  Flow-DPO