Separating explicit world state from appearance synthesis in video generation improves long-horizon consistency and enables direct control over predicted behavior without retraining the observation model.
Marionette is a world model for interactive games that separates world state prediction from appearance synthesis. Instead of directly generating pixels, it predicts explicit 3D skeletal poses and trajectories, uses a fixed geometric renderer to compute occlusion and geometry, then synthesizes realistic appearance on top. This makes long-horizon predictions more stable and controllable.