Paper Overview
- Field: Computer Vision (CV)
- Authors: Jing He, Kaixin Ding, Xingye Tian, Guibao Shen, Wenhang Ge, Xin Tao, Pengfei Wan, Ying-Cong Chen
- Date: 2026-10-08
- arXiv: 2610.12382
- Static regions: aligned with geometric world priors via semantic-guided masked reprojection, enabling more reliable static-consistency evaluation. An auxiliary camera-motion reward suppresses near-static degenerate solutions.
- Dynamic subjects: a strong vision-language model (VLM) serves as a dynamic world prior, with a VLM-as-judge reward built from sample-specific checklists evaluating dynamics, physical plausibility, shape, and texture consistency.
Summary
Faithful visual world simulation requires generated videos to maintain 4D world consistency, encompassing both static and dynamic consistency. Static consistency requires coherent 3D structure in static environments across viewpoints, while dynamic consistency requires plausible subject motion and consistent appearance over time.
Geometry-aware post-training offers a promising way to improve world consistency. However, existing methods often rely on a static-scene assumption. Even those that accommodate dynamic scenes struggle to provide reliable static-consistency feedback, while dynamic consistency is often overlooked or inadequately assessed.
The WorldAlign Approach
WorldAlign is a decoupled 4D reward framework that semantically separates static regions and dynamic subjects, aligning each with world priors suited to their assumptions:
Results
On two pretrained image-to-video models, Wan2.1 and Wan2.2, WorldAlign jointly improves static and dynamic consistency without suppressing overall or subject motion.
---
*Auto-collected on 2026-10-12.*