WorldAlign: Decoupled 4D Reward for World-Consistent Video Generation
WorldAlign introduces a decoupled 4D reward framework to improve video generation consistency by separating static regions from dynamic subjects. It aligns static geometry using geometric reprojection and uses a Vision-Language Model (VLM) as a prior for dynamic subjects, assessing dynamicity and plausibility via sample-specific checklists. This allows online post-training without human preference annotations, resulting in better static and dynamic consistency.