World Labs turns visual prompts into explorable 3D worlds

World Labs develops AI that generates and reconstructs spatial environments. Its public Marble application turns prompts and visual inputs into persistent, explorable 3D worlds, while the World API lets developers add that capability to their software. Atlas is a newer model entering selective early access. The first buying decision is whether you need a creative workspace, a programmable service or a research partnership.

01 / ProductMarble is a spatial creation tool with a developer counterpart

World Labs’ Marble announcement describes generation from text, images, video and coarse 3D layouts, followed by editing, expansion and composition. The result is a scene a viewer can move through. That is a different output from a fixed image or a video clip whose camera path is already chosen. It gives a creative team a space in which to explore framing and alternatives.

The World API announcement takes the same broad capability into software workflows. Generation is asynchronous: an application submits a request and later receives a result that it can show or use downstream. An application developer therefore needs a job-management flow, rather than assuming world generation behaves like an instantaneous image URL.

The model documentation distinguishes Marble 1.1, a Plus option with variable cost for larger worlds, and an older Draft model for faster iteration. Separately, the September 2026 Atlas announcement describes a broader model entering early access with selected partners. A Marble subscription should not be presented as guaranteed access to Atlas merely because both belong to World Labs.

02 / AudienceWho benefits from an explorable scene

A creative team can use Marble when it needs to explore a location concept from several viewpoints before investing in a more detailed production asset. Film previsualisation, immersive storytelling and early environment design are plausible situations. The value comes from being able to revisit the same spatial concept, rather than repeatedly asking for unrelated still images and trying to reconcile them afterwards.

A software team has a different job: create an interface that lets users request, inspect and retain generated worlds. The World API is the relevant route for that integration. It should be budgeted and evaluated as a service dependency, including how the application reports progress, handles a failed request and limits expensive repeated generations.

The tool is less suitable as the sole source of truth for dimensions, construction details or physical behaviour. A generated room may be visually persuasive while inventing unseen geometry. Our Luma blueprint is an adjacent comparison when the desired deliverable is generated visual media. The deciding question is whether the project needs a navigable spatial asset, a rendered sequence or both, rather than which company has the broadest AI description.

03 / WorkflowA proposed scene-development workflow

Consider a proposed workflow for a small studio designing an imaginary visitor centre for a film concept. The studio owns its reference material and chooses a plan permitting the intended commercial work. This is an editorial example, not a commissioned World Labs case study or a test we performed. The deliverable is a scene for creative review, with no claim that it is suitable for construction or real-world navigation.

Start with one visual direction and an explicit review question: can the director find a convincing entrance view and an interior camera position within the same concept? Keep the first brief small enough to judge. Generating a complete fictional city before agreeing the visual language creates more material to discard and makes comparisons less meaningful.

The Marble overview describes a Create workspace and Studio tools for editing, composition and recording. In the proposed workflow, review the first result from several viewpoints before spending credits on expansion. Look for inconsistent geometry, implausible transitions or details that only work from the initial angle. Those checks are suggested creative criteria, not defects we observed in a generated sample.

Use an inexpensive draft to explore alternatives, then choose a model and export path for the selected concept. The export guide distinguishes splat exports from high-quality textured meshes and notes their plan requirements. Decide the downstream application early: an attractive browser preview does not establish that the asset will behave as expected in the studio’s existing tools.

Keep the original references, the generation settings and the chosen output together. That lets another team member understand which details were supplied and which were generated. If a later request changes the layout, compare it against the agreed concept rather than accepting a visually impressive but incompatible replacement. The useful outcome is a reviewable scene and a list of remaining production work, not a claim that generation has finished every stage of asset creation.

04 / PricingApp subscriptions and API usage have separate economics

The rendered Marble pricing page, checked on 22 September 2026, lists monthly plans and action-based credits. Commercial rights and high-resolution textured mesh exports appear from Pro. The table below records the displayed monthly prices, rather than an annual-equivalent figure. Generation counts are advertised as “up to” amounts because actions and model choices affect consumption.

The billing guide says subscription credits do not roll over, while purchased top-ups carry over and expire after one year. Marble and the World API have separate balances, even when the same email is used. A team building an application should not buy a creative subscription on the assumption it includes developer usage.

The API pricing guide lists US$1 per 1,250 credits, with a US$5 minimum purchase. As an illustrative calculation, a 1,580-credit standard text-to-world request costs US$1.264 at that rate, before further actions. It also warns that admitted requests can create an overage beyond the prepaid balance; disabling auto-refill is not a hard spending cap. An application needs its own request limits and usage monitoring.

05 / DistinctionsPersistent worlds change how a team can iterate

The distinction is persistence of a spatial object across review decisions. When a scene can be revisited from different angles, a director can discuss the relationship between viewpoints rather than approve isolated frames. That can be valuable even if substantial manual work remains before production. The benefit should be tested against a real review task, not inferred from the amount of generated detail.

World Labs’ documentation also makes an interface limitation clear: some advanced creation and editing features are not supported on mobile, and desktop is recommended for the full workflow. A producer may be able to review a shared result on one device while an artist needs another environment to create or modify it. Establish that handoff before promising an all-mobile production process.

The NVIDIA blueprint provides context for teams whose next step involves AI infrastructure or simulation tooling. A generated world is an input to a larger workflow, and its appearance does not validate the physics or behaviour of that workflow. The downstream system still needs the representations and properties required by its own task. Visual similarity and operational correctness answer different questions.

06 / QuestionsExport, rights and model access need explicit checks

The first question is output suitability. A Gaussian splat, a collider mesh and a textured mesh are different artifacts. The export documentation places basic exports on paid plans and high-quality textured mesh export with commercial rights on Pro. Confirm the format, performance and editing behaviour in the actual downstream application before producing a large library of assets.

The second is entitlement. The live pricing page and billing documentation do not describe every feature gate identically: the rendered page currently places several creation and editing capabilities in Free, while the billing guide describes some under higher tiers. The common, consequential boundary for the proposed commercial workflow is Pro’s commercial rights and high-quality mesh export. Check the live account and terms for a specific feature rather than treating every help-page example as a universal plan promise.

The third is model availability. Atlas is entering selective early access. Its research claims should not be used to describe what every Marble user receives today. An external team depending on a new model should obtain explicit access and confirm its supported interface before basing a delivery schedule on it.

Finally, budget for rejected attempts as well as accepted assets. An application can generate a technically successful world that does not satisfy the user’s creative brief. Track cost per accepted result, review time and downstream repair work in a small evaluation. This public-source article does not establish generation quality, latency or productivity improvements from hands-on testing.

07 / DecisionChoose the output and access route first

World Labs is most compelling when a project benefits from moving through a persistent scene and iterating on it. Start with a small visual brief, decide how the output will be used and select the corresponding app or API route. That gives pricing, export tests and model availability a concrete meaning, while keeping early-access research separate from the product a team can use now.