Explicit representations of scene state (like object positions and tracking maps) extracted from generated frames to maintain consistency.