World Labs Opens Atlas 3D World Model to All Developers

World Labs has moved Atlas, its spatial world model, from closed partner access to general, self-serve availability. Atlas generates spatially consistent scenes and camera-controlled video from photos or simple geometry. The launch includes two browser apps, Viewfinder and Chisel, plus a documented API for agents and custom pipelines.

The company introduced Atlas as a research preview in September, claiming it could generate up to a minute of 1440p video from a small set of photographs while giving users control over the camera. Access was limited to unnamed partners, and the announcement included no academic paper, model card, dedicated pricing, or public API. The new release addresses access and integration, while the documentation and pricing gaps remain.

Pixels and blocks feed the same model

Viewfinder and Chisel use the same underlying model but accept different kinds of input:

• Viewfinder starts with images. A single image can generate a plausible surrounding scene, while as many as 64 photographs can reconstruct a real environment. Generations include color and per-pixel depth, which estimates the distance between the camera and each visible surface.
• Chisel starts with geometry. Users assemble planes, blocks, and other basic shapes into walls, furniture, and structural elements, then style the scene with text prompts or reference images. Its agentic workflow interprets instructions and applies a sequence of revisions to the 3D layout.

Both applications export keyframes, video clips, and 3D Gaussian splats. Gaussian splats encode a scene as collections of colored, semi-transparent 3D ellipsoids, allowing compatible viewers to render and stream detailed environments efficiently.

The pipeline gets endpoints

Atlas exposes its core workflows through operations listed in the API reference:

The published GET example retrieves an existing asset by ID and authenticates through the WLT-Api-Key header:

curl --request GET \
  --url 'https://api.atlas-beta.worldlabs.ai/api/v2/assets/asset_01j...' \
  --header "WLT-Api-Key: $WLT_API_KEY"

The splats2Mesh operation is especially useful for production pipelines. Gaussian splats render well in dedicated viewers, but polygon meshes are easier to edit, simulate, and import into tools such as Blender and Unreal Engine. A hosted converter can replace the separate reconstruction tools commonly used for that step.

From partner preview to self-serve access

1. Atlas has its own front door. The model now lives on a dedicated Atlas site instead of being accessible only through Marble Studio.
2. The API is available without a partner agreement. Developers can obtain access directly and connect the model to internal tools, automated workflows, or agents. The documentation and API hostnames retain beta labels.
3. A real-time variant is available as a research preview. World Labs streams generated worlds from cloud GPUs to a browser at up to 15 frames per second while the user controls the camera. The company says this variant is tuned for low-latency inference while preserving novel-view synthesis, the ability to render camera angles absent from the source images.

Atlas pricing still has blanks

World Labs has not published a dedicated Atlas pricing page. Marble Studio provides a rough baseline: since October 30, a refined Viewfinder splat generation has taken about 20 minutes and consumed 5,000 credits, equivalent to $5 under that schedule. Results can be downloaded as SPZ or PLY files. Marble plans range from a free tier with 7,000 credits to a $95 monthly plan with 120,000 credits and commercial rights. Atlas workloads may use a different schedule, so these figures do not establish its production cost.

Benchmark claims await replication

World Labs reports that human raters preferred Atlas over competing video models in 75% to 94% of trials. The company also says Atlas achieved lower 3D reconstruction error than reproduced open-source systems across six benchmark datasets.

Several comparisons used different test conditions for Atlas and competing methods, which favored the company’s model. World Labs has not accompanied the results with an academic paper or model card, and independent researchers have yet to reproduce them. The reported figures therefore remain vendor evidence until external evaluations confirm the results.

Fewer handoffs in 3D pipelines

Typical 3D generation systems chain together a text-to-image model, depth estimator, NeRF or splat trainer, mesh extractor, and texture baker. World Labs says Atlas was pretrained from scratch to combine video generation and 3D reconstruction, allowing one hosted service to produce consistent views, camera motion, depth, splats, and mesh conversions.

Where the API fits first

1. VFX and game previsualization: Teams can block out a scene in Chisel, export a splat, and use it as a reference in a digital content creation tool.
2. Architectural visualization and virtual tours: A small photo set can become a navigable reconstruction of a room or property.
3. Synthetic data for robotics and computer vision: The API can generate RGBD sequences, which combine color and depth, with controlled camera trajectories for training policies or vision models.
4. Agent-driven scene building: An LLM-based agent can invoke Atlas, inspect rendered outputs, revise geometry or prompts, and repeat the process programmatically.

Production questions remain measurable

Teams can now test Atlas against their own scenes and measure reconstruction accuracy, generation latency, mesh quality, prompt sensitivity, and credit consumption. Independent evidence remains limited, refined splat exports require meaningful time and credits, and Chisel’s results depend on the quality of its prompts and reference images. Self-serve access makes those constraints testable without waiting for a partner slot.