Geo-Align: Video Generation Alignment via Metric Geometry Reward

Geo-Align is a reinforcement learning framework for camera-controlled video re-rendering.

It uses a metric 3D estimator to recover camera trajectories and apply scale-aware perceptual rewards.

The method avoids relying on paired real-world multi-view training data.

It improves both camera-motion controllability and visual quality, especially in out-of-distribution scenes.