NVIDIA Cosmos3-Super-Image2Video is a frontier-scale video generation model developed as part of the Cosmos 3 family of omnimodal world models. Released in June 2026, the model is designed to produce temporally coherent and high-fidelity video sequences starting from a single input image and a descriptive text prompt. Unlike purely visual generators, it is categorized as a "world model" for Physical AI, meaning it is engineered to understand motion, spatial layout, and causality to assist in the development of robotics and autonomous systems.
The model is built on a Mixture-of-Transformers (MoT) architecture that unifies reasoning and generation. It utilizes a "two-tower" system consisting of a 32B reasoner tower (based on a vision-language model) and a 32B generator tower (a diffusion-based transformer), resulting in a total of 64 billion parameters. The reasoner tower acts as the model's "brain," interpreting the scene's physical context and planning the motion before the generator tower synthesizes the final visual output.
Capabilities and Performance
Cosmos3-Super-Image2Video has been recognized for its performance in open-weight video synthesis, specifically in its ability to maintain object permanence and realistic physics over time. It supports generation across multiple resolution tiers—256p, 480p, and 720p—and is typically used for generating synthetic data to train embodied agents. The model is distributed under the OpenMDW-1.1 license, allowing for both research and commercial application.
To achieve the best results, it is recommended to provide a detailed text description that specifies the intended motion and interaction between objects. Since the reasoner tower interprets the prompt to establish a physical blueprint of the scene, explicit mentions of weight, contact, and temporal changes help the model ground the generated video in a plausible world state.