Atlas is an omni world model developed by World Labs, designed to achieve spatial intelligence through the generation, reconstruction, and simulation of 3D environments. Built as a multimodal autoregressive diffusion transformer, Atlas moves beyond traditional text-to-video paradigms by natively integrating 2D images, text, and 3D spatial geometry into a unified "spatial context."

Core Methodology

Atlas is fundamentally defined by its architecture as a multimodal autoregressive diffusion transformer, which leverages the following technical components:

• Spatial Contextualization: Unlike models that process inputs in isolation, Atlas grounds every input image and depth map in a precise 3D position in space. This creates a shared spatial context that enables the model to interpolate between unrelated image pairs and maintain 3D consistency across long-range generations.
• Autoregressive Transformer Backbone: The model treats multimodal data as sequences. By processing inputs autoregressively, the transformer maintains temporal and spatial consistency, predicting the next element in a sequence while conditioned on previously encoded spatial context. This architecture allows the model to leverage LLM-style optimization techniques like KV-caching.
• Rectified Flow Diffusion: Atlas utilizes a rectified flow diffusion mechanism to generate high-fidelity continuous outputs (images and videos). By gradually denoising latent representations, it provides high-resolution, stable visual outputs. The model uses this to perform complex tasks like filling in occluded regions—guessing missing geometric details based on its internalized world knowledge.
• Native Camera Geometry: A defining feature of Atlas is its use of explicit camera poses as a native input type. Rather than relying on fuzzy natural language prompts for camera movement, the model accepts precise camera geometry, allowing for pixel-perfect control over trajectory, panning, and zooming.

Capabilities and Applications

Atlas demonstrates superior performance across three primary domains:

1. Camera-Controlled Generation: Atlas can output up to 1 minute of video at 1440p. Because it natively understands camera geometry, it enables users to hand-design exact camera paths that remain consistent with the input images, even when extrapolating parts of the scene not present in the original frames.
2. Spatial Reconstruction: The model performs novel view synthesis from sparse inputs. It converts 2D images into 3D representations, such as point clouds or 3D Gaussian splats, allowing for immersive, navigable environments reconstructed from as few as one or two images.
3. Space-Time Simulation (Real-to-Sim): Atlas serves as a simulator for robotics and VFX. It models physical interactions and time-evolution, enabling it to reframe captured video from impossible angles ("bullet time") or simulate various robot sensors (RGB and depth) as they navigate reconstructed environments.

Performance and Scaling

Atlas is designed to improve with computational scale, following empirical trends where increased training compute yields proportional gains in capability. In comparative benchmarks, Atlas outperforms specialized state-of-the-art models in both camera-conditioned video generation and 3D reconstruction. Its ability to generate explicit 3D outputs (Gaussian splats) allows it to bridge the gap between creative visual generation and practical applications in gaming, robotics, and design.