What Is Atlas? World Labs’ AI Model for Generating and Reconstructing 3D Worlds

World Labs presents Atlas as an “omni” world model for spatial intelligence: one system trained to work natively with text, images, video and 3D information. Its goal is broader than producing a plausible video clip. Atlas is designed to generate images and video from controlled viewpoints, reconstruct real scenes in 3D and model how a space might appear as the camera or scene changes. 9

The distinction matters because conventional image and video generators generally optimize for visual plausibility frame by frame. Atlas is positioned around spatial consistency: the relationship between objects, viewpoints and the underlying environment should remain coherent as the camera moves.

How Atlas works

World Labs describes Atlas as a multimodal autoregressive diffusion transformer pretrained from scratch. The model associates visual inputs with their positions in a shared spatial context, then predicts new frames, views or 3D outputs that fit that context. 9

A user can provide one or more reference images together with camera information or a designed camera path. Atlas can then render requested viewpoints, including parts of a scene that were not directly observed. Those unseen areas are inferred from the available evidence and the model’s learned visual priors, rather than recovered as verified physical measurements.

That design gives Atlas two related abilities:

• Camera-conditioned generation: Generate images or video while controlling where the virtual camera moves.
• Spatial reconstruction: Infer a 3D scene and render novel views from a limited set of images.

The same spatial representation is also intended to support simulation, such as changing lighting, motion, backgrounds or object configurations for downstream robotics work. 9

What Atlas can generate

According to World Labs, Atlas supports several generation modes:

• Camera-controlled images and video from one or more reference images.
• Text-conditioned images.
• 360-degree panoramas.
• Video output of up to one minute at 1440p.

The company describes the camera control as “pixel-perfect,” meaning that the requested camera trajectory is a central part of the generation process rather than an after-the-fact edit. These capabilities could allow creators to reframe footage, synthesize camera moves or produce shots from sparse visual references. 9

The practical value of the approach is control. A prompt can describe what should exist, while images and camera geometry provide evidence about where objects are and how the scene should be viewed. Atlas is therefore aimed at the space between generative video and conventional 3D production workflows.

3D reconstruction from a few images

World Labs says Atlas can reconstruct real-world scenes from one to dozens of input images. The company also says that two or three photographs can often produce faithful results, while the model can incorporate more than 100 views when the priority is fidelity rather than inference. 9

The output is intended to include both novel-view images and explicit 3D representations. In other words, Atlas is not limited to making a sequence that looks correct from one prescribed angle; it is designed to represent enough of the scene to generate additional views.

There is an important caveat. When a scene is only sparsely observed, unseen surfaces must be inferred. A reconstruction can look convincing while still being wrong about an occluded object, a measurement, a hidden room or the geometry of a surface. Atlas’s visual quality should therefore not be confused with a guaranteed survey-grade reconstruction.

What the benchmark claims show—and do not show

World Labs says Atlas outperforms state-of-the-art specialist 3D-reconstruction models on the company’s reconstruction evaluations. The launch material highlights quantitative results for camera-conditioned generation and 3D reconstruction, while also noting that no single benchmark captures the model’s full range of capabilities. 9

Those are meaningful company-reported results, but they are not the same as independent validation. The material available here does not include a complete reproducible evaluation package with all dataset protocols, per-model scores, uncertainty estimates, model weights and third-party replications. The performance lead should therefore be read as a claim from World Labs, not as a settled industry result.

Further testing is especially important for:

• Long, complex camera paths.
• Occluded or ambiguous parts of a scene.
• Metric accuracy and geometric measurement.
• Consistency across extended video sequences.
• Collision and physical-interaction fidelity.
• Transfer from generated simulation environments to real robots.

How Atlas relates to Marble

Marble is World Labs’ existing user-facing product for creating persistent, explorable 3D worlds. It accepts text, single or multiple images, video, panoramas and coarse 3D structures, then produces worlds that users can explore and work with. 6

Atlas is the next-generation model underneath that product strategy, rather than simply a renamed Marble feature. World Labs says Atlas will power future versions of Marble and other products. 9

This relationship gives the announcement a clear product direction: Atlas supplies a more general spatial model, while Marble remains the accessible environment for creating and viewing persistent worlds. Marble’s documentation describes world creation from text, images and video, with generated worlds available for exploration and downstream workflows. 6

The World API and developer access

World Labs introduced the World API as a public interface for generating explorable 3D worlds from text, images, panoramas, multi-view inputs and video. The API documentation describes a workflow in which an application submits a generation request and later retrieves the completed world. 83

That is different from exposing Atlas itself as a downloadable or real-time local model. The documented API focuses on asynchronous world generation and retrieval, and the available materials do not establish a public Atlas model endpoint, local weights or real-time inference option. 389

For developers, this means the current public product surface is Marble and its API-based world-generation workflow—not necessarily direct access to every Atlas capability described in the research announcement.

Intended applications

World Labs frames Atlas as a general-purpose spatial system with several potential markets:

Visual effects

Camera-controlled generation could help artists reframe footage, create virtual camera moves and produce shots from a small number of references. The main promise is controllable spatial variation rather than an isolated generated frame. 9

Video games

Marble already focuses on persistent, explorable environments. Atlas could extend that direction by giving game and world designers a model that generates or reconstructs scenes while maintaining viewpoint and spatial relationships. 69

Design and visualization

Architectural, industrial and creative teams could use text, images and other references to explore navigable spatial concepts. Viewpoint control may make it easier to inspect and iterate on a concept than a sequence of unrelated renders. 69

Robotics

World Labs also describes a “real-to-sim” use case: creating simulated environments from casual recordings, generating robot-perspective RGB and depth observations, and varying objects, lighting, motion and backgrounds for navigation and manipulation training. 9

That is an ambitious application, but generated visual realism alone does not prove that a scene has accurate physics, reliable dimensions or sufficient sim-to-real performance. Those claims require task-specific evaluation.

Access, pricing and remaining unknowns

Atlas is not presented in the launch material as broadly available. World Labs is collecting early-access requests, while Marble and the World API are the documented public options. 98

The launch information reviewed here does not disclose a standard Atlas price card, model size, training data, hardware configuration, inference latency, throughput or per-generation compute cost. 9

Those omissions matter for anyone evaluating Atlas as a production tool. A visually impressive demo does not reveal whether a workflow is affordable at scale, fast enough for interactive use or reproducible across many scenes.

The bottom line

Atlas is best understood as World Labs’ attempt to unify generative video, 3D reconstruction and spatial simulation in one multimodal model. Its distinguishing idea is to make camera position and 3D context first-class inputs, rather than treating every generated frame as an independent image.

The company’s reported capabilities are substantial: controlled image and video generation, reconstruction from sparse views, explicit 3D outputs and possible robotics simulation. But Atlas remains an early-access system in the material reviewed here. Until independent evaluations, pricing, latency and physical-accuracy data are available, its strongest claims should be treated as promising research and product direction—not as independently proven production performance.