Look at a photograph of a room. You can admire the wallpaper, judge the furniture, perhaps suspect the owner of having too many mirrors. But try stepping sideways. The photograph has nothing to say. What sits behind the chair? Where does the hallway lead? World Labs has made those missing answers its business: turning visual references into spaces people can move through, change and use.
- Make a place: Marble creates persistent, navigable 3D environments from visual and written inputs.
- Keep the craft: Customers add art direction, gameplay and production refinement.
- Follow the compute: AMD’s pending acquisition connects spatial AI research to its hardware.
The immediate appeal is creative control. Once a scene exists in three dimensions, a camera can visit it repeatedly. The larger ambition is physical intelligence: machines that understand places, objects and actions. Between those two propositions sits an unusually revealing company, with a public creation tool, an open-source renderer and research reaching into robotics.
A photograph gets a floor plan
World Labs’ founding group combines Fei-Fei Li, Justin Johnson, Christoph Lassner and Ben Mildenhall. Their shared territory is machine learning, computer vision and graphics. Li’s earlier ImageNet work helped establish modern visual AI. In her account of founding World Labs in early 2024, she describes a conviction that language alone cannot equip models to reason about the physical world.
The company’s first public product, Marble, launched in November 2025. Give it text, an image, multiple views or video, and it generates an explorable environment. Users can edit and expand worlds, then export them for downstream work. January’s World API made generation programmable, letting developers request environments from within their own software rather than make each one by hand.
Ten places the cameras could no longer visit
The best explanation comes from a television production. For Harlan Coben’s Final Twist, Promise needed ten historic crime-scene environments. The locations had changed or vanished. Surviving pictures were sometimes dim, low-resolution or confined to one viewpoint. Going back with a camera was no longer an option.
In the published production account, the team first restored archival imagery. Marble turned those references into Gaussian splat environments, a 3D representation made from colored, semi-transparent elements. Artists expanded missing areas, corrected inconsistencies, refined geometry and materials, and prepared the results for a 16K LED volume. Coben could then perform against reconstructed surroundings that responded to camera movement.
That sequence is the useful bit to copy. Improve the references. Generate a foundation. Inspect it. Spend skilled effort on the shots that matter. A subscription buys access to generation; a broadcast-ready environment also requires people, equipment and judgment. Confusing those two costs would make a splendid sales pitch and a miserable production schedule.
- 01 Restore
the reference - 02 Generate
the space - 03 Refine
with artists - 04 Shoot
on the stage
The room stays put. You change your mind.
Magnific’s 3D Scenes offers another practical example. Its Marble-powered environments let creators compose objects and choose camera positions inside a persistent space. A campaign team can change the framing without asking an image generator to reinvent the whole room. The customer buys control over a scene, with generation supplying the surroundings.
Marble’s experimental Chisel mode applies a related idea: rough boxes, planes or imported assets establish the layout; text supplies the look. The company’s neat formulation is “Chisel decouples structure from style.” Anyone who has tried describing the exact position of a sofa to an image model will appreciate the distinction.
“Chisel decouples structure from style.”World Labs, Marble launch
At the listed monthly prices, Standard costs $20, Pro $35 and Max $95; a free plan offers limited credits. Pro adds commercial rights and high-resolution textured mesh export. Generation, expansion and exports consume credits, so the headline price does not describe every project’s capacity. API credits are purchased separately. The business combines subscriptions, usage charges and custom enterprise plans.
Three people, two months, one browser
WithLore’s ICARE gives the proposition a different shape. A primarily three-person team built its explorable dream world in roughly two months, combining Marble environments with Spark 2.0 and custom Three.js systems. Characters, props and gameplay still came from a wider production stack. WithLore estimates its previous approach would have taken closer to a year. That is the studio’s comparison, rather than a controlled benchmark.
WithLore’s reported ICARE build. Generated environments, custom gameplay, human direction.
Spark explains why distribution matters as much as generation. World Labs built its own renderer after existing options imposed limits, including problems displaying multiple splat objects together. Large environments then exposed download and memory constraints. Spark 2.0 streams changing levels of detail instead of demanding the entire world at once. Developers can copy the MIT-licensed renderer and adapt it.
Why the laboratory wants a chipmaker
The research now extends beyond scenery. July’s SceniX acquisition brought robotics simulation into World Labs. Its real-to-sim-to-real work reconstructs physical tasks, varies simulated conditions and compares outcomes against hardware. A convincing robot video is a beginning; reliable performance amid shifting objects and changing light is the harder test.
September’s Atlas announcement brought generation, reconstruction and simulation into one multimodal model, entering early access with selected partners. Its missing views can be plausible inventions. More reference images reduce that freedom. For imaginative environments, filling gaps is useful. For an exact building survey or validated robotic task, plausibility needs evidence and inspection.
The competitive field includes DeepMind’s Genie, with interactive generated worlds, and NVIDIA Cosmos for physical AI. World Labs’ visible foothold is a creation workflow with exportable environments and browser infrastructure. Its product examples make the case more clearly than an abstract contest over which model understands reality.
Funding gave the research room to develop: $230 million announced in 2024, followed by $1 billion in February 2026, including Autodesk’s $200 million investment. Then came AMD’s approximately $8.2 billion all-stock acquisition agreement. Announced September 28, it remains subject to approvals, with closing expected by year-end. Li’s stated rationale is closer coordination with hardware after a technical partnership that began in 2025. The picture acquired depth; the laboratory now wants a closer hand in the machinery that draws it.