Atlas (World Labs)

Atlas is an early-access world model announced by World Labs on September 1, 2026. World Labs describes it as an "omni model" for camera-controlled image and video generation, sparse-view 3D reconstruction, video reframing, and simulation workflows. Its disclosed architecture combines multimodal sequence modeling, autoregressive generation, a diffusion process based on rectified flow, and a transformer backbone.^[1]

Atlas was not generally available at launch. World Labs offered access only to select partners through a request form. As of September 3, 2026, the company had not published model weights, code, a public Atlas API specification, pricing, a model card, a technical paper, or a model license.^[1]

This page covers the World Labs model. It is unrelated to Atlas (language model), the Atlas humanoid robot, ChatGPT Atlas, or the SWE-Atlas software-engineering benchmark.

Overview

Scroll sideways for more →

World Labs released an official launch video alongside the announcement. It presents generated camera moves, reconstructions, video reframing, and robotics-oriented scenes, but it is a demonstration rather than an independent test.^[2]

Architecture

World Labs characterizes Atlas as a "multimodal autoregressive diffusion transformer." This is a high-level description, not a complete architecture specification.^[1]

Multimodal spatial context

The detailed announcement names text, images, camera poses, and 3D depth maps as the modalities Atlas currently processes. Video is represented as a sequence of images rather than listed as a separate primitive. Each image and depth map is paired with an explicit camera pose, allowing the sequence to encode where an observation sits in 3D space. World Labs calls the resulting collection a spatial context.^[1]

This distinction matters for camera control. A text instruction such as "pan left" specifies a qualitative motion. A camera pose specifies a viewpoint geometrically. Atlas conditions generation on those poses while also using the visual and textual information already placed in context.^[1]

Autoregressive diffusion transformer

Atlas treats a task as a sequence of multimodal elements. It generates each new element one at a time, conditioned on earlier elements, following the broad logic of an autoregressive model. At the same time, World Labs describes each generated output as emerging through gradual denoising with a rectified-flow process. The announcement also calls Atlas a latent diffusion model.^[1]

These parts are compatible rather than contradictory: autoregression describes how elements are ordered and conditioned, while diffusion describes how an individual continuous output is generated. The transformer supplies the shared sequence-modeling backbone. Earlier research established transformers as attention-based sequence architectures, diffusion transformers as transformer backbones for latent diffusion, and rectified flow as a learned distribution-transport method.^[5][6][7]

World Labs says Atlas can benefit from techniques used in language-model serving and diffusion systems, naming KV caching, cache-aware routing, disaggregated serving, diffusion distillation, classifier-free guidance, shifted noise schedules, and VAE design. The announcement does not say that the early-access system implements every item in that list. It also does not disclose parameter count, context length, network depth, latent representation, training compute, or sampling settings.^[1]

The company says Atlas was pretrained from scratch on a large, diverse multimodal corpus and that larger training runs improved its capabilities. It does not describe the corpus composition, filtering, licensing, or data provenance, so the training claim cannot be independently audited from the launch materials.^[1]

Camera-controlled generation

Atlas accepts one or more reference images plus specified camera positions and angles, then generates views along a chosen path. The launch page's videos use between one and six input images and manually designed camera paths. World Labs reports that Atlas can produce up to one minute of video at 1440p and shows one such example made from a small number of reference images.^[1]

The one-minute duration and 1440p resolution are vendor-reported demonstration scope. The announcement does not provide latency, hardware, cost, success-rate, or quality-distribution measurements for such generations, so those figures should not be read as an independently verified service guarantee.^[1]

When a requested view exposes content outside the reference images, Atlas must generate a plausible continuation. World Labs explicitly says the model "imagines" such areas and gives examples including the back of a robot, a lawn outside the original frame, and connective doorways or hallways between unrelated images. These regions are inferred or synthesized from learned priors. They are not measured observations of the source scene.^[1]

Spatial reconstruction and 3D output

World Labs presents Atlas as both a generator and a computer vision reconstruction system. It can accept a variable number of views, from one image to more than one hundred, and produce novel-view images together with explicit depth or geometry. The company says two or three images often yield faithful reconstructions, but this is an author-reported claim illustrated by selected examples rather than an independent reconstruction study.^[1]

For a single image, Atlas generates additional views and estimates their geometry before converting the result into a 3D representation. For video input, it performs depth estimation on individual frames and combines the estimates into a point cloud. The point cloud can then be completed and converted to a 3D Gaussian splatting scene for rendering.^[1]

Gaussian splatting represents a scene with explicit 3D Gaussian primitives and can support high-quality, real-time novel-view rendering. The original 3D Gaussian Splatting work optimized those primitives from calibrated images.^[8] Atlas uses Gaussian splats as an output representation, but World Labs has not disclosed enough implementation detail to assume its conversion pipeline is identical to the original method.

The word "reconstruction" needs a boundary here. Geometry visible in the inputs can be estimated from image evidence, while geometry absent from every camera is generated. World Labs states that Atlas fills unseen regions in both its single-image and video workflows. A completed splat may therefore be visually coherent without being a factual measurement of hidden surfaces, dimensions, or objects.^[1]

Space-time generation and video reframing

World Labs describes Atlas as a space-time simulator because it models both scene layout and change across image sequences. One launch demonstration uses footage from three to five ordinary phones or action cameras to freeze an event and synthesize views from new angles, creating a "bullet time" effect without a large synchronized camera array.^[1][2]

This demonstration shows multiview conditioning and temporal reframing. It does not, by itself, establish exact recovery of motion, occluded geometry, or physical dynamics. The launch material provides no quantitative temporal-consistency or physics-accuracy benchmark for these examples.^[1]

Real-to-Sim workflows

Atlas is also presented as one component in Real-to-Sim workflows for robotics. In the navigation examples, World Labs captured two large spaces with phone video, used 24 frames from each space for reconstruction, simulated robot paths, and generated RGB and depth observations from body-mounted camera viewpoints. The examples join an explicit environment representation with sensor-view generation from the same model.^[1]

For manipulation, the company says a few recordings can help build simulations of how objects move and interact. The resulting setups can vary objects, positions, robot motion, lighting, and background to create training or testing conditions. The Atlas post shows rigid, articulated, and deformable-object examples, but it does not report component-level measurements for contact physics, forces, deformation, or sim-to-real policy transfer.^[1]

World Labs had separately described a broader Real-to-Sim-to-Real engine in July 2026. That system reconstructs physical tasks, varies scenes and physics, and supports policy training and evaluation.^[4] The post predates Atlas and does not attribute all of its results to Atlas. Those broader robot-policy results therefore should not be treated as Atlas-specific evidence.

Image generation

Although World Labs says world modeling is the primary focus, Atlas also generates images and 360-degree panoramas from text or image prompts. The company demonstrates varied visual styles, prompt following, and rendered text. No standardized image-generation benchmark is included in the launch material.^[1]

Reported evaluations

All quantitative Atlas results available at launch were designed and reported by World Labs. Human raters in the camera study were third parties, but the evaluation was not an independent publication or reproduction.^[1]

Camera-path following

World Labs evaluated one input image followed by a target path containing one to three cinematic camera motions. Atlas received the path in its native camera format. Comparison video models did not accept camera data, so they received a text description using terms such as pan, truck, and crane. Third-party raters then selected the output that better followed the intended path.^[1]

Scroll sideways for more →

The figures above are vendor-reported shares from the official chart, where 50 percent represents an even split. World Labs also notes that more sophisticated prompting or multimodal prompts might improve some baselines. Because Atlas receives an explicit camera path while the baselines receive text, the experiment tests camera control through each model's available interface rather than conditioning all systems on equivalent data.^[1]

Sparse-view reconstruction

For sparse-view 3D reconstruction, each system receives images with camera poses and predicts one 3D point for every input pixel. World Labs compares Atlas with Pi3X using poses, pi-cubed, VGGT-Omega 1B, Depth Anything 3, and MapAnything across DTU, ETH3D, KITTI, NRGBD, 7-Scenes, Tanks and Temples, and ScanNet.^[1]

The company reports that Atlas has the lowest average point-map absolute-relative error among the listed methods, at 25.3 thousandths. World Labs says it reproduced the baselines under a common protocol. The chart does not establish superiority on every reconstruction dataset, output type, or metric, and the result has not been independently reproduced in the cited sources.^[1]

Relationship to Marble

Marble is an existing World Labs product, not another name for Atlas. World Labs made Marble generally available in November 2025 with support for text, image, video, and coarse-3D inputs, interactive world editing, and export to Gaussian splats, meshes, or video.^[3]

The Atlas announcement says Atlas will power future versions of Marble and other World Labs products. It also notes that Atlas and Marble use Gaussian splats as a scene representation, which can ease integration. The future-tense announcement does not establish that the Marble service available in September 2026 already used Atlas.^[1][3]

Access, reproducibility, and limitations

Atlas entered early access with select partners. A request form was public, but the model itself was not offered as a generally available self-service product at launch.^[1]

The lack of a technical paper, weights, code, model card, public API contract, and model license limits independent inspection. The published description is not sufficient to reproduce the architecture, training procedure, benchmarks, or serving stack. The launch page also omits generation cost, latency, training-data composition, safety evaluations, and data-governance details.^[1]

Its demonstrated outputs have further interpretive limits:

• Visual consistency does not guarantee metric geometry, correct scale, collision geometry, material properties, or physical accuracy.
• Hidden parts of a scene are synthesized rather than observed, so they may be plausible but factually wrong.
• The camera benchmark gives Atlas native geometric controls and gives comparison systems text, making it informative about available camera-control interfaces but not a modality-matched comparison.
• The reported one-minute 1440p result is a company example without a published distribution of quality, speed, or failure cases.
• The Atlas Real-to-Sim examples do not constitute a published closed-loop robot-policy or sim-to-real transfer benchmark.
• The cited sources contain no independent reproduction of Atlas results.

World Labs' own taxonomy warns that generated scenes can look convincing while containing wrong scale, self-intersections, or other structural errors that would make physics unreliable.^[9] Atlas is intended to span generation, reconstruction, and simulation, but its early-access announcement does not show that these broader reliability problems are solved.

References

1. ^World Labs Team, "Atlas: A World Model for Spatial Intelligence," September 1, 2026. worldlabs.ai/...atlas
2. ^World Labs, "Introducing Atlas; A Foundation Model for Spatial Intelligence," YouTube, 2026. youtube.com/watch
3. ^World Labs Team, "Marble: A Multimodal World Model," November 12, 2025. worldlabs.ai/...marble-world-model
4. ^World Labs Team, "Building Worlds That Train Robots," July 28, 2026. worldlabs.ai/...real-to-sim-to-real
5. ^Ashish Vaswani et al., "Attention Is All You Need," arXiv:1706.03762, 2017. arxiv.org/...1706.03762
6. ^William Peebles and Saining Xie, "Scalable Diffusion Models with Transformers," arXiv:2212.09748, 2022. arxiv.org/...2212.09748
7. ^Xingchao Liu, Chengyue Gong, and Qiang Liu, "Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow," arXiv:2209.03003, 2022. arxiv.org/...2209.03003
8. ^Bernhard Kerbl et al., "3D Gaussian Splatting for Real-Time Radiance Field Rendering," ACM Transactions on Graphics 42, no. 4, 2023. repo-sam.inria.fr/...3d-gaussian-splatting
9. ^World Labs Team, "A Functional Taxonomy of World Models," June 3, 2026. worldlabs.ai/...taxonomy-of-world-models