Cosmos3-Super-Text2Image is a high-fidelity image generation model developed by NVIDIA as part of the Cosmos 3 suite of omnimodal world models. Built specifically for Physical AI and world simulation, this 64-billion-parameter model is a post-trained specialization of the Cosmos 3 Super foundation. It is designed to generate highly detailed, physics-consistent images from text descriptions, serving as a foundational block for creating synthetic data and simulating real-world environments.
The model utilizes a Mixture-of-Transformers (MoT) architecture, which uniquely unifies reasoning and generation tasks. This design pairs an autoregressive reasoning transformer (32B parameters) with a diffusion-based generation transformer (32B parameters) through shared multimodal attention. This combination allows the model to better understand complex spatial-temporal relationships and physical plausibility compared to standalone diffusion models.
Agentic Capabilities and Prompting
A defining feature of this model is its agentic upsampling workflow. Rather than processing raw user prompts directly, the system uses its internal reasoning component to refine and expand short descriptions into detailed, high-quality instructions. NVIDIA provides a dedicated agentic upsampling package alongside the model to optimize prompt alignment and visual quality.
For optimal results, users are encouraged to provide descriptive prompts that specify scene lighting, camera angles, and physical materials. The model's training on large-scale synthetic and curated physical datasets allows it to interpret these nuances and produce outputs that adhere to the laws of physics, making it suitable for training robotic policies and autonomous systems in digital twin environments.