World Labs
17 Intel
World Labs is a spatial intelligence company building AI systems that perceive, generate, and let people interact with three-dimensional environments. It was founded in 2024 by Stanford professor and AI researcher Fei-Fei Li together with Justin Johnson, Ben Mildenhall, and Christoph Lassner, leaders in computer vision and graphics research. The company has raised roughly $1.23 billion and released Marble, a tool for generating persistent, explorable 3D worlds, followed by the Atlas model for creating detailed simulated environments from a single image.
Is this you? Sign in with X to claim this profile.
Intel
World Labs introduces Atlas, a multimodal world model pre-trained to take camera movement as an input, generating video with precise trajectory control and reconstructing explicit 3D from one or many images. It reports beating top open-source reconstruction models and emitting RGB plus depth for rob
World Labs ships two Marble models. Version 1.1 improves lighting and contrast with a stated reduction in visual artifacts, while 1.1-Plus is built for scale and larger, more complex environments. Both are selectable in the product and live in the API.
image-blaster, built by a World Labs team member and published on GitHub, chains Marble, Claude skills and fal into one pipeline that turns a single image into an explorable 3DGS environment plus object meshes, physics-ready objects and ambient SFX, with Three.js, Unity and Unreal outputs.
Three World Labs papers on generating 3D from 2D priors: World Tracing predicts layered per-pixel depth with diffusion to recover occluded surfaces from a single image, Modality Forcing adapts text-to-image models to reason jointly over text, images and depth, and Flex4DHuman lifts monocular video t
World Labs open-sources Spark 2.0, a web renderer built to stream scenes of over 100 million Gaussian splats on ordinary devices, aimed at making high-fidelity 3DGS worlds viewable in a browser rather than a native client.
Marble now takes a rough scene laid out in any 3D application and generates a finished world from that geometry, moving world-model output from prompt roulette to a layout you control.
World Labs breaks down the ICARE build: seven worlds designed and iterated in Marble, streamed as millions of Gaussian splats in the browser through Spark 2.0, with Three.js handling gameplay and a Blender integration connected to the live runtime.
World Labs argues 3D is to spatial work what code is to software: a durable, externalized artifact both humans and machines can inspect, edit, rerun and pipe into existing tools. The consequence is a preference for world models that output structured 3D over ones that output pixels.
Marble is generally available. The world model takes text, images, video or coarse 3D layouts as input, supports interactive editing, expansion and combination of worlds, and exports Gaussian splats, meshes or video. Marble Labs adds case studies and documentation.
World Labs upgraded its 3D world generation to produce bigger, stylistically varied environments with cleaner geometry, and opened Marble as a limited beta. Worlds can be exported as Gaussian splats and dropped into Three.js scenes through the open-source Spark renderer.
How Spark 2.0 streams huge Gaussian splat scenes on the web: a level-of-detail system that adapts splat density to the viewpoint and fetches data on demand, built on THREE.js and WebGL2 so it runs on phones and headsets rather than WebGPU desktops only.
Fei-Fei Li argues that language models remain ungrounded in physical space, and that spatial intelligence, perceiving and reasoning about 3D worlds, is the next capability frontier behind robotics, simulation and immersive creation. The founding thesis behind World Labs' world models.
World Labs splits the overloaded term 'world model' into distinct functions, mapping renderers, simulators, and planners onto the classic POMDP loop of action, state, and observation. Useful for judging what a given spatial or video model actually commits to.
World Labs puts Marble behind a public API. Requests run asynchronously from text, images, panoramas, multi-view sets or video and return navigable 3D scenes carrying layout, depth and lighting, ready to render in the browser or drop into simulation, games and robotics pipelines.
World Labs introduces Atlas, an omni world model pretrained on text, images, video and 3D. Every input is grounded at a 3D position in a shared spatial context, giving pixel-perfect camera control, minute-long 1440p generation, and reconstruction it claims outperforms specialized 3D models.
World Labs details the real-to-sim-to-real engine it gained with the SceniX acquisition: turn one physical robot task into an interactive simulation, then systematically vary appearance, clutter, physics and viewpoint to manufacture the breadth of experience a policy needs to generalize.
World Labs' RTFM renders explorable scenes by generating video frame by frame as you interact, at interactive framerates on a single H100, with a world that persists rather than resetting. The post also works through the token and context economics that make world models expensive to serve.