Atlas: The First Multimodal World Model by Fei-Fei Li's Startup
Atlas is a groundbreaking multimodal world model that generates images and videos with pixel-level camera control and can reconstruct them into 3D scenes.
Atlas: The First Multimodal World Model by Fei-Fei Li's Startup
Meet Atlas – the first multimodal world model in history. It generates images and videos with pixel-level camera control and can reconstruct them into 3D scenes in explicit point cloud and 3D Gaussian splats format. It is also trained for space-time simulations.
Key Features:
- Video Duration: Up to 1 minute at 1440p.
- Reference Images: Up to 6.
Architecture
Atlas employs a multimodal diffusion transformer, essentially a hybrid of LLM architecture (with autoregression + caching) and diffusion video models (for denoising), further combined with the idea of spatial context. This means that each contextual element has an explicit 3D attachment through the camera position.
Benchmarks
It claims superiority over specialized models in camera-controlled generations (MiniMax, Gemini Omni Flash, FLUX, etc.) and in 3D reconstruction (Pi3X, VGGT, Depth Anything 3, etc.).
Applications
The engine is so precise that it can be used to train robots. Atlas can build an explicit 3D scene from video, where robot movement can be simulated and trained (this is called the Real-to-Sim-to-Real approach). The model generates RGB and depth from the perspective of the robot's onboard cameras and can even create a simulation where object manipulation is possible.
Conclusion
This release is an impressive achievement in the field of AI and modeling.
Why it matters
AnalysisAtlas opens new possibilities for creating 3D scenes and training robots, which could significantly impact the development of AI technologies. This is an important step towards integrating AI into the real world.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Google Gemini Pro (Model with Infinite Context)
The flagship working model from Google DeepMind featuring a record context window of over 2 million tokens. Capable of analyzing entire books, lengthy videos, and massive codebases in a single query.
Native Multimodal Reasoning
AI architectures capable of performing deep logical reasoning over code, architecture diagrams, UI screenshots, and video within a unified attention space.
MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.
JSON — Structured Data for Programs, APIs, and AI Agents
A complete hands-on guide to JSON: syntax specifications, serialization via parse and stringify, REST API consumption, JSON Schema validation, and structured outputs for LLMs.