CausalAlign: Causal Alignment of Video Generation Models for Consistent Ego-Motion

Researchers present CausalAlign, a method for causally aligning diffusion-based video generation models to fix camera ego-motion drift over long horizons, according to the paper's abstract. The work targets a known failure in video generation models, whose camera ego-motion often drifts despite producing visually plausible output, and treats consistent ego-motion as essential for maintaining geometry and trajectories across frames.

Diffusion-based video generation models VGMs can produce visually plausible videos, yet their camera ego-motion often drifts over long horizons. Camera ego-motion determines how the viewpoint changes between frames and is essential for maintaining consistent geometry and trajectories. To address t