Static research summary generated from local editorial content.
| Attribute | Value |
|---|---|
| Topic | Video World Models |
| Summary | How video world models learn physics, temporal consistency, and interactive simulation from large-scale video, from Sora and Genie to Cosmos and V-JEPA. |
| Related Models | 9 |
| Citations | 3 |
Editorial body section preserved directly in static HTML.
Video world models learn environment dynamics directly from video sequences, treating video as a rich source of temporal, spatial, and physical structure. Instead of only modeling rewards or compact control states, they model how scenes evolve over time and how actions or prompts alter those futures.
Editorial body section preserved directly in static HTML.
Video captures object permanence, motion, occlusion, contact, and scene continuity at internet scale. That makes it one of the most information-dense training signals available for world models. As a result, video-based systems are increasingly central to robotics simulation, autonomous driving, embodied AI, and interactive environment generation.
Editorial body section preserved directly in static HTML.
The field now spans diffusion-based generators, autoregressive token models, JEPA-style predictive latent models, and action-conditioned interactive simulators. Sora and Cosmos emphasize generative realism, V-JEPA emphasizes predictive abstraction, and Genie-style systems push toward controllable environments rather than passive video synthesis.
Editorial body section preserved directly in static HTML.
Video world models are strongest when they need to produce realistic future frames, model plausible physical events, and create large quantities of synthetic training data. They are especially useful in driving simulation, environment generation, and embodied pretraining, where realism and temporal continuity matter.
Editorial body section preserved directly in static HTML.
The biggest open problems are long-horizon coherence, stable memory, action grounding, and evaluation. A model can generate visually plausible clips while still failing at causal control, object permanence, or physical consistency over longer interactions.
| Model | Lab | Category | Index v1.1 |
|---|---|---|---|
| Sora | OpenAI | Generative World Model | 63/100 |
| Genie 2 | Google DeepMind | Generative World Model | 79/100 |
| Genie 3 | Google DeepMind | Generative World Model | 89/100 |
| NVIDIA Cosmos | NVIDIA | Foundation World Model | 87/100 |
| V-JEPA | Meta | Self-Supervised World Model | 70/100 |
| V-JEPA 2 | Meta | Self-Supervised World Model | 87/100 |
| UniSim | Google DeepMind | Generative World Model | 72/100 |
| DIAMOND | Microsoft Research / University of Geneva | Model-Based RL | 64/100 |
| Pandora | Tsinghua University / ByteDance | Generative World Model | 52/100 |
FAQ answers rendered directly into static HTML for extractable responses.
Not always. A video generator can produce plausible sequences without being a strong interactive world model. Video world models become more useful when they preserve causality, memory, and controllability over future states.
Sora, Genie 3, NVIDIA Cosmos, V-JEPA 2, and UniSim each represent different parts of the frontier, from generative realism to predictive abstraction and interactive simulation.
Short extractable summary preserved directly in static HTML.
Editorial provenance and refresh policy preserved directly in static HTML.
Published by world-models.io editorial board.
Lead editor Bernard Grenat.
This research page curates topic explanations, linked models, and citations grounded in primary research sources.
Each editorial page is assembled from primary sources, normalized into extractable summaries, checked for factual drift, and reviewed before publication or major refreshes. Last reviewed: 2026-06-21.
Pages are refreshed when a new paper, benchmark, release, architecture update, or stronger primary source materially changes the answer a reader or AI system should retrieve.
Each page links back to relevant primary sources and keeps a stable canonical URL so readers can verify claims, trace context, and reference the most up-to-date version. See the editorial policy.
Primary research citations embedded in static HTML.