arXiv 2606.20083 introduces Holo-World, a video world model that uses a single image to control camera, object, and weather as three independent axes. The standout: "environment state" (e.g., weather, lighting, time of day) is now an independent control axis, not a derived effect of camera/object changes.
The "three-axis" control: traditional video world models allow control over camera (e.g., pan, zoom) and object (e.g., move a person), but environment state (e.g., change from sunny to rainy) is usually a "side effect" of the other controls. Holo-World makes environment state a first-class control axis — the user can independently specify "this should be a rainy scene" without specifying any camera or object changes.
The technical details: Holo-World uses a "factorized state space" that separates camera, object, and environment into three independent latent variables. The model is trained on a large corpus of video with environment annotations (weather, time of day, season), and the factorized state space is learned end-to-end.
The benchmark: on the "environment control" benchmark, Holo-World hits 89% accuracy in matching the specified environment state (vs 42% for the previous SOTA). The video quality is also competitive — Holo-World scores 81.3 on VBench, on par with Sora 2.
The bigger takeaway: "factorized control" is the right architecture for video world models. The traditional "everything in one latent space" approach makes environment control hard, and the factorized approach is a major improvement. For the industry, this means "video world model" applications (game AI, simulation, content creation) will get significantly better controllability, opening up new use cases.