World Narrative Model for Highly Controllable Video Generation: A Paradigm Shift from Pixel Sampling to Physical World Orchestration

Modern video generation models, such as Sora or Kling, treat video creation as a black-box pixel distribution sampling problem. While these models can produce high-quality visuals, they are inherently uncontrollable. Creators cannot specify precise 3D geometry, joint angles, or camera trajectories, leading to a frustrating gacha loop where dozens or hundreds of generations are required to achieve a single usable shot. This lack of determinism and physical-level control makes these models unsuitable for professional filmmaking.

The World Narrative Model (WNM) introduces a paradigm shift by decoupling the generation process into two distinct stages: the Controller and the Renderer. The Controller, or WNM, interprets user intent to build a structured, editable 4D (3D+T) physical blueprint. This blueprint includes explicit scene layouts, asset placements, character skeleton motions, camera paths, and lighting configurations.

Instead of replacing existing foundation models, WNM uses them as frozen neural shaders. The structured physical parameters from the WNM act as a deterministic blueprint, guiding the foundation model to render the final video. This approach ensures that the output remains consistent, editable, and precisely aligned with the creator's vision.

Built on the WNM engine, the platform provides a professional-grade interface that supports both automatic generation and fine-grained manual refinement. Creators can use four dedicated control panels—Scene/Asset, Motion/Trajectory, Cinematography, and Lighting—to manipulate the 3D world directly. Because the underlying representation is a structured 4D world, changes to one attribute (like moving a character) propagate correctly to related entities (like shadows or held objects), maintaining physical and semantic coherence.