KAIST's World Observer Decouples Observer From Actor to Fix Out-of-View Blind Spots in Video World Models

A KAIST AI-led team (with Sony Sejong Seoul contributors) posted World Observer, which attacks the actor-centric bias in video world models by jointly generating a perspective actor view and one or more panoramic observers that watch regions the agent can't see. Observers are geometrically grounded via panoramic warping and an 'observer sink' of high-resolution references lets the model restore fine appearance when objects re-enter view. The authors report substantially better out-of-view dynamics while remaining competitive on visual fidelity, camera control and 3D adherence.