Video world models typically simulate an environment from the perspective of a single acting agent. That actor-centric design creates a persistent problem: once an object leaves the agent's field of view, the model loses direct evidence of how it changes, and often fails to preserve its state and dynamics when it reappears.
Researchers at KAIST CVLab propose World Observer, a framework that decouples observing from acting. Instead of relying on one viewpoint, it jointly generates a perspective actor for the agent-centric view alongside one or more panoramic observers that watch selected regions of the world. Objects that exit the actor's view can therefore keep evolving visually inside an observer, so their updated states are reflected when they re-enter.
To keep the actor and observers geometrically aligned, the method warps both from a shared panoramic source, establishing explicit correspondence. An Observer Sink of high-resolution perspective references is used to restore fine appearance details upon re-entry.
Because the observers are decoupled from the actor, they can be placed freely across the scene, extended to multiple locations for broader coverage, and driven by control signals to steer how out-of-view regions evolve.
The team also introduces world-space metrics and a benchmark spanning real and synthetic scenes to evaluate out-of-view evolution. According to the paper, World Observer substantially improves out-of-view dynamics while remaining competitive in visual fidelity, camera control, and 3D adherence.