Atlas: The One Model That Maps, Builds, and Plays in Any World

RoboticsSpatial IntelligenceAI & Machine Learning

Imagine a single AI that can paint a city, rebuild a room, and let a robot navigate it—all from a handful of photos.

TL;DR
• Atlas is an omni world model that handles text, images, video, and 3D.
• It excels in camera‑controlled generation, 3D reconstruction, and space‑time simulation.
• Performance scales with compute, promising continual improvement.

Summary
Atlas, World Labs' new omni world model, unifies text, images, video, and 3D into a single spatial context, enabling generation, reconstruction, and simulation tasks with high fidelity and precise camera control. Built as a multimodal autoregressive diffusion transformer, the model scales with compute and demonstrates superior performance on camera‑conditioned generation and sparse‑view 3D reconstruction benchmarks.
• Deploy Atlas only after a formal impact assessment to identify privacy and safety risks.
• Publish clear documentation of the spatial context handling and diffusion steps to allow external audit of algorithmic decisions.
• Implement a user‑feedback channel that logs reconstruction errors and unintended content for rapid redress.
• Restrict training data to sources with documented consent and enforce encryption of depth maps and camera poses.
• Conduct regular third‑party reviews of scaling experiments to verify that performance gains do not compromise ethical safeguards.

Atlas generates pixel‑perfect camera‑controlled video from one to six reference images, supporting custom paths up to one minute at 1440p. By grounding each input frame in explicit 3D pose, it can extrapolate unseen scene regions while maintaining geometric consistency, allowing creators to direct shots rather than rely on vague textual prompts. The system also blends unrelated images in a shared spatial context, producing seamless interpolations that showcase its world knowledge.

For spatial reconstruction, Atlas requires only a few sparse images to produce accurate novel‑view synthesis, outperforming specialized 3D models. Adding more views reduces imaginative filling, yielding faithful recreations of real environments such as Stanford's Main Quad. The model outputs both 2D frames and explicit 3D representations, including point clouds and Gaussian splats, which render efficiently on‑device and integrate with World Labs' Marble platform.

In space‑time simulation mode, Atlas reframes video captured by ordinary smartphones into bullet‑time perspectives and generates synthetic sensor data for robotics. Real‑to‑Sim pipelines benefit from synchronized RGB and depth streams, enabling scalable training of navigation and manipulation policies without expensive capture rigs. The architecture’s diffusion‑based denoising and transformer backbone support rapid adaptation to new tasks while preserving a unified spatial reasoning core.

Benchmarks confirm Atlas leads recent video models in adhering to complex camera trajectories and surpasses top open‑source reconstruction systems on standard datasets. Scaling experiments show consistent capability gains with increased compute, suggesting future iterations will further enhance fidelity and control. Early access is limited to select partners, with a call for responsible integration and ongoing ethical oversight.

Will Atlas redefine how we create and interact with virtual worlds, or will it simply extend existing tools?