World Labs

World Labs is a spatial intelligence research and product company founded by Fei-Fei Li, Justin Johnson, Ben Mildenhall, and Christoph Lassner. Its first product, Marble, generates navigable, editable 3D environments from prompts and visual inputs. Its broader program connects world generation with camera control, reconstruction, and simulation of how actions change a physical scene.[1][5][13]

Li dates World Labs’ founding to early 2024. It publicly emerged that September with $230 million in backing. WIRED’s launch account describes Li recruiting Johnson and investor Martin Casado suggesting Lassner and Mildenhall; Mildenhall left Google to join. The company’s initial roadmap put spatial-model development first, augmented reality next, and robotics later.[2][3]

In her November 2025 essay, “From Words to Worlds”, Li argues that language alone leaves substantial gaps in spatial understanding. Her intended world models are generative, multimodal, and interactive: they create environments, combine different kinds of observations, and predict subsequent states conditioned on actions. This agenda extends beyond navigating or editing a generated scene to modeling the physical consequences of interaction.[2][5]

Marble entered limited-access beta on September 16, 2025, and general availability on November 12. The beta emphasized environments rather than isolated people, animals, or central objects. At general launch, Marble accepted text, images, video, and coarse 3D layouts, supported editing, expansion, and composition, and exported Gaussian splats, meshes, or videos.[4][5]

A single image shows only part of a space; Marble must invent details outside that view. Multiple images or video give users more control over what different parts of the generated world look like. Its experimental Chisel mode adds direct control over layout: boxes, planes, or imported assets establish coarse structure, while text specifies appearance. As the company puts it, “Chisel decouples structure from style.”[5]

Gaussian splats represent scenes using colored, semitransparent ellipsoids with position, scale, orientation, color, and opacity. Marble’s mesh exports serve different purposes: collider meshes support coarse physics simulation, while higher-quality meshes aim to preserve visual detail for conventional 3D tools. Video exports provide camera control; enhanced videos can add motion and remove artifacts while following the generated scene’s structure.[9][5]

World Labs used Marble-generated environments in Marble’s own launch film. According to its production case study, a16z New Media helped shape the film, while Pixomondo and World Labs integrated Gaussian splats into Unreal Engine for an LED volume stage. Virtual production supervisor Nhan Le’s team adjusted scale, depth, and lighting; director Mark Helsel and cinematographer Cameron Nowy worked with the environments on set. Generated scenery thus entered an established production pipeline with further preparation for filming.[10]

Before Marble’s general launch, World Labs introduced the RTFM research preview on October 16, 2025. It explored a different way to produce consistent views: an autoregressive diffusion transformer trained on video generated frames without building an explicit 3D scene. Input-frame activations in its KV cache implicitly represented the world. World Labs reported interactive inference on a single H100 GPU.[6]

RTFM associated stored frames with camera positions and orientations. For each requested viewpoint, it retrieved nearby posed frames as context—a technique the team called “context juggling.” This spatial memory avoided attending to an ever-growing history of frames and helped preserve previously visited places. The preview supported viewpoint exploration; dynamic worlds and further interaction were proposed extensions.[6]

The World API, launched January 21, 2026, exposed Marble generation to other applications. Generation runs asynchronously: software submits a request and receives a navigable world after processing. Inputs include text, single or multiple images, panoramas, and video. Results can be rendered in browsers or exported into downstream tools, extending world creation beyond Marble’s own interface.[7]

World Labs also built Spark, an open-source Gaussian-splat renderer that integrates with Three.js alongside conventional meshes. According to the company’s account of its origins, Spark began internally because existing renderers imposed constraints on multiple-object rendering, animation, framework choice, or browser compatibility. The team developed a general-purpose Three.js/WebGL2 library, shaped by Marble’s need to compose and display generated worlds.[8][9]

Spark 2.0, introduced April 14, 2026, added viewpoint-dependent level of detail and streaming. A hierarchy of splats supplies coarse or fine detail according to the camera view. Its .RAD format lets the browser fetch scene chunks progressively, while a fixed GPU memory pool swaps chunks in and out. These mechanisms reduce the need to download, store, and render all the detail of a large world at once.[9]

World Labs announced its acquisition of SceniX on July 21, 2026. The robotics team brought a real-to-sim-to-real (R2S2R) engine, described the following week, that turns physical tasks into reusable simulations for training and evaluating robot policies. Financial terms were not disclosed.[12][13]

The R2S2R workflow captures robots, sensors, surroundings, objects, and task demonstrations, then reconstructs simulations preserving task-relevant observations and dynamics. Controlled variations in object configurations, appearance, physical properties, robot states, and viewpoints provide training experience. Matched interactions and policy rollouts in simulation and reality test whether the simulated world reproduces behavior and supports useful evaluation decisions.[13]

World Labs reported early policies trained entirely in simulation transferring to physical robots on tasks involving deformable cables, articulated boxes, rigid objects, and cluttered pick-and-place. It also reported that simulation preserved relative policy rankings on an ALOHA bimanual cube-handover task. Zero real-world policy-training data still allowed real-world capture for task reconstruction. These results concern the company’s demonstrated tasks rather than arbitrary robots or environments.[13]

On September 1, 2026, World Labs introduced Atlas, a multimodal autoregressive diffusion transformer pretrained from scratch. It processes text, images, camera poses, and depth maps, with video represented as sequences of images. Images and depth maps are grounded by camera poses in a shared spatial context; outputs are generated one at a time, conditioned on that context. Atlas can jointly generate views and geometry, producing images, video, point clouds, or Gaussian splats.[14]

Additional observations constrain Atlas’s reconstruction, while unseen regions can be plausibly filled in. In one company demonstration, a single garden photograph yielded an imagined surrounding scene; adding photographs of the cottage and main house progressively constrained those buildings. For video reframing, engineers and researchers captured examples with ordinary phones and compact mounting equipment, then generated new viewpoints from three to five camera views.[14]

Atlas entered early access with select partners, with plans to power future versions of Marble and other products. World Labs reported camera-controlled video generation of up to one minute at 1440p and benchmark advantages over selected video and reconstruction models. Its camera-following comparison gave Atlas native camera inputs while describing camera motions to other models through text prompts.[14]ⁿ¹

World Labs announced $1 billion in new funding on February 18, 2026, naming AMD, Autodesk, Emerson Collective, Fidelity Management & Research Company, NVIDIA, and Sea among its investors. Salesforce Ventures lists the company as private and Li as cofounder and CEO. Lassner announced that he stepped away from his role while continuing to support World Labs as an advisor.[11][15][16]