World Labs Unveils 'Atlas,' a Spatial Intelligence-Omni World Model

On September 1, 2026, World Labs unveiled Atlas, a new omni world model targeting spatial intelligence. Atlas is a model pre-trained from scratch to natively process text, images, video, and 3D data, combining all inputs into a shared spatial context to generate and simulate worlds while maintaining 3D consistency.

Architecture and Key Features

Atlas adopts a multimodal autoregressive diffusion transformer structure. Each image forms context based on its specific position within 3D space, and the model uses this as a condition to generate the next frame. Key features include Camera-Controlled Generation, which accepts precise camera geometry as input to generate 1440p videos up to 1 minute long; 3D Reconstruction, which reconstructs high-quality 3D scenes from sparse input images; and World Simulation, which implements 'bullet time' effects from ordinary smartphone footage or generates Real-to-Sim data for robot simulation.

Performance and Benchmarks

Atlas demonstrated superior performance compared to specialized models. In camera-controlled generation evaluations, it recorded higher preference rates than the latest video models such as Gemini Omni Flash and FLUX 3. This advantage is particularly pronounced with complex camera trajectories. In 3D reconstruction benchmarks (AbsRel), it showed lower error rates than top open-source models like Pi3X and VGGT-Ω 1B, achieving an average AbsRel of 8.6 (×10⁻³).

Scalability and Outlook

Atlas was pre-trained on large-scale multimodal data and exhibits scaling characteristics where new capabilities are unlocked as model size and training compute increase. World Labs expects this trend to continue, leading to dramatic improvements in the performance of future world models. Currently, Atlas is available in early access and is expected to serve as the core engine powering World Labs' product lineup, including Marble.