World Labs / Marble

Fei-Fei Li 2025 · Paper

One-line summary — Marble, from Fei-Fei Li’s startup World Labs, generates persistent, navigable 3D worlds — delivered as Gaussian-splat scenes — from image, video, or text prompts, positioning “Spatial Intelligence” as the next foundation-model capability after language.

Problem

Existing routes to a 3D scene sit at two extremes. Reconstruction methods (NeRF, 3D Gaussian Splatting) require tens of posed input images of a real place and only reproduce what was observed. Video generators (Sora, GAIA-1) create compelling imagery from a single prompt, but their output is a 2D pixel stream with no persistent underlying scene — look away and back, and the world may have changed. World Labs argues that world models should “reconstruct, generate, and simulate 3D worlds” and let humans and agents interact with them, and builds Marble as a generative model whose output is an explicit, persistent 3D world.

Method & architecture

Marble is a commercial system; no paper, equations, or training details are public. What follows is the system-level pipeline as described in World Labs’ announcement (“Marble: A Multimodal World Model”, Nov 12, 2025):

Because the underlying model is undisclosed, no equations can be given here; treat this note as a product/system description.

Results

No quantitative benchmarks are published. Publicly demonstrated results are qualitative: navigable, viewpoint-consistent Gaussian-splat worlds generated from single images, text, multi-image, and video prompts across diverse scene types and artistic styles; structure-controlled generation via Chisel; edited, expanded, and composed worlds; and mesh/video exports (nearly all videos in the announcement post were rendered directly from Marble worlds). See the World Labs blog for the demonstrations; claims cannot be independently verified.

Why it matters for SLAM

Marble targets the same output artifact SLAM produces — a consistent, renderable 3D scene — but obtains it by generation from priors rather than by measurement, which sharpens the question of when sensor-based mapping is actually necessary. Its export formats (Gaussian splats, collider meshes) are exactly the representations modern SLAM and robotics simulation consume, suggesting concrete hybrid uses: filling unobserved regions of a SLAM map with plausible geometry, providing map priors for exploration, and supplying photorealistic 3D environments for testing SLAM and navigation stacks. As a commercial product from one of AI’s most prominent researchers, it also signals that Spatial AI has moved from research vision to product race.