World Labs Unveils Atlas With Native Camera Control and 3D Reconstruction

World Labs introduced Atlas on September 1, 2026, presenting it as a multimodal world model that generates camera-controlled images and videos, reconstructs scenes in explicit 3D, and reframes recorded events from new viewpoints.

The most consequential change is that Atlas accepts camera geometry as a native input. Instead of asking a video generator to interpret words such as “pan,” “truck,” or “crane,” a user can specify camera positions and trajectories within a shared three-dimensional context. World Labs says Atlas can produce videos lasting up to one minute at 1440p from a small set of reference images and a manually designed camera path.

Atlas is not yet a generally available product. It is entering early access with selected partners, and World Labs has not announced public pricing, a model identifier, an API release date, or broad access conditions.

1. Atlas Unifies Several Spatial Tasks in One Model

World Labs describes Atlas as an “omni model” pretrained from scratch. Its architecture combines an autoregressive transformer with latent diffusion: multimodal elements are processed as a sequence, while images and other continuous outputs are generated through a rectified-flow denoising process.

The model currently works with text, images, camera poses, and 3D depth maps. Videos are represented as sequences of images. Each image and depth map can be associated with an explicit camera pose, grounding the inputs in what World Labs calls a spatial context.

That design allows one model to handle several tasks that are often separated across different systems:

• generating images and videos from reference images;
• controlling the viewpoint and camera path of generated frames;
• synthesizing novel views of an existing scene;
• estimating depth and reconstructing explicit 3D geometry;
• reframing synchronized video footage from new angles;
• generating sensor observations for simulated robots;
• creating conventional images and 360-degree panoramas from prompts.

World Labs markets Atlas as a first-of-its-kind multimodal world model, but that priority claim has not been independently established. The company itself described its earlier Marble system as a multimodal world model in November 2025. The more specific technical claim behind Atlas is that generation, reconstruction, camera conditioning, depth, and temporal simulation have been combined within one pretrained architecture.

2. Native Camera Geometry Changes How Shots Are Controlled

Atlas can take between one and six reference images for the camera-controlled video examples shown at launch. Users place those images in a spatial context and specify where the virtual camera should move. The model then generates views along that trajectory while attempting to preserve the scene’s appearance and geometry.

This is materially different from text-only camera prompting. A phrase such as “slowly orbit the subject” leaves a model to infer the orbit’s radius, direction, elevation, speed, and framing. Native camera input can encode those choices directly, giving filmmakers and visualization teams a more precise control surface.

World Labs calls the result “pixel-perfect camera control.” That wording should be treated as a company description rather than an independently measured accuracy guarantee. The published evaluation tests whether human raters believe Atlas follows an intended camera path better than comparison models; it does not establish zero pixel-level error.

The spatial context also supports composition. World Labs demonstrates two otherwise unrelated reference images placed at different positions in 3D space. Atlas generates intermediate spaces—such as doorways, corridors, and connecting rooms—to form a continuous world between them.

That generative behavior is useful for world-building, but it also establishes an important boundary: continuity does not necessarily mean reconstruction accuracy. When an input does not reveal part of a scene, Atlas invents a plausible completion based on its learned priors.

3. Sparse Images Can Become Point Clouds and Gaussian Splats

Atlas reconstructs scenes from one or more posed images and can use more than 100 images in the same spatial context. World Labs says two or three views are typically enough for a faithful reconstruction, while additional views reduce how much unseen content the model must generate.

A single-image result is necessarily a mixture of observation and synthesis. In one company example, Atlas preserves the visible portion of a garden from a ground-level photograph but invents surrounding buildings and terrain. Adding photographs of a nearby cottage and main house progressively constrains the resulting scene.

Atlas can return novel-view images, depth information, point clouds, and 3D Gaussian splats. From one image, it generates additional views while estimating their geometry; from a video, it predicts per-frame depth and combines the frames into a reconstruction. Regions never captured by a camera are filled by the model.

Gaussian splats provide a renderable representation rather than only a collection of generated frames. They can be viewed from changing positions and integrated into downstream 3D workflows. Atlas uses the same broad representation as Marble, which should make the model’s output compatible with future versions of World Labs’ existing product.

For visual-effects and design teams, this could reduce the amount of dense photographic capture required to establish a navigable scene. The tradeoff is that generated geometry outside the observed views may look coherent without corresponding exactly to the original location.

4. Video Reframing and Robotics Extend Beyond Scene Generation

Atlas can reconstruct a recorded event from multiple synchronized cameras and render it from viewpoints at which no physical camera was present. World Labs demonstrates “bullet time” reframing with footage from three to five ordinary phones and action cameras mounted on tripods or clamps.

The system uses the multiple views to estimate the event’s spatial and temporal structure. An editor can then freeze the action or move a virtual camera around it. This offers a lighter capture setup than a conventional multiview studio, although the launch material does not disclose processing time, hardware requirements, or failure rates.

For robotics, Atlas combines reconstruction with generated sensor observations. In two navigation demonstrations, World Labs captured large environments using phone video, selected 24 frames for each reconstruction, and simulated robots moving along different paths. Atlas generated the RGB images and depth data that body-mounted cameras would observe from those paths.

The company also shows Real-to-Sim workflows for manipulation. A small number of recordings can be converted into a simulated task, after which developers can vary objects, positions, robot motion, lighting, or backgrounds. The demonstrations include rigid, articulated, and deformable objects.

These outputs could supply training and testing environments without repeatedly staging every variation on physical hardware. However, the Atlas announcement does not quantify whether the generated dynamics reproduce real-world forces, contact behavior, friction, or deformation accurately enough for a robot policy to transfer reliably.

5. The Benchmark Results Are Promising but Company-Reported

World Labs published evaluations for camera-conditioned generation and sparse-view 3D reconstruction. For camera control, each trial began with one input image and a trajectory containing one to three cinematic movements. Third-party human raters compared how well the outputs followed the requested path.

Atlas received the path through its native camera representation. Rival video models received text descriptions using conventional cinematic terms because they did not accept the same camera format. World Labs acknowledges that more elaborate prompting might improve some competing models, making this partly a comparison of control interfaces rather than a strictly identical input test.

For reconstruction, the company reports lower error than specialized open-source systems including Pi3X, π³, VGGT-Ω 1B, Depth Anything 3, and MapAnything under its reproduced evaluation protocol. The launch page does not provide an accompanying peer-reviewed paper or enough methodological detail to independently reproduce all of the reported results.

World Labs has also not disclosed Atlas’s parameter count, training-data composition, inference cost, generation speed, safety controls, or evaluation sample sizes in the launch material. Its demonstrations and benchmarks therefore establish the company’s reported capability, not independent validation across uncontrolled footage or production workloads.

Marble, by comparison, became generally available in November 2025 and can generate persistent 3D worlds from text, images, video, or coarse layouts, with exports including Gaussian splats, meshes, and videos. Atlas is intended to power future versions of Marble and other World Labs products. For most users, the announcement is therefore a technology preview rather than an immediate replacement for the currently available platform.

Frequently Asked Questions

Is Atlas publicly available?

No. World Labs says Atlas is in early access with selected partners. Interested organizations can submit an access request, but public eligibility rules have not been announced.

Can Atlas generate a 3D scene from one image?

Yes, but unseen areas are generated rather than recovered from evidence. Adding more images constrains the model and reduces the amount it must invent.

What 3D formats can Atlas produce?

The company says Atlas can generate depth maps, point clouds, and 3D Gaussian splats in addition to conventional image and video frames.

How long can Atlas-generated videos be?

World Labs demonstrates camera-controlled output lasting up to one minute at 1440p. It has not published generation-time or hardware requirements.

Is Atlas replacing Marble?

Not immediately. World Labs says Atlas will power future versions of Marble and other products, while Atlas itself remains in limited early access.

Sources

• Fei-Fei Li’s Atlas announcement on X
• World Labs: Atlas — A World Model for Spatial Intelligence
• World Labs: Marble — A Multimodal World Model
• World Labs Atlas early-access request
• SiliconANGLE: Fei-Fei Li’s World Labs debuts Atlas