3D & spatial world modelspartial
Where video-generative models produce frames, the 3D/spatial branch produces places — persistent, navigable 3D representations that don’t drift when you look away, because the geometry exists independently of the rendering. The field’s center of gravity is 3D Gaussian splatting (3DGS) as the output representation.
World Labs’ Marble — the branch’s first real product
Marble (marble.worldlabs.ai) became generally available on November 12, 2025 — World Labs’ “frontier multimodal world model” and the first commercial product of the branch.
Inputs: text prompts; single or multiple images (with directional placement); video; and coarse 3D layouts via the experimental “Chisel” sculpting mode. (Video and multi-image inputs are gated to paid tiers.)
Outputs: persistent 3D worlds, primarily as 3D Gaussian splats — Marble’s
highest-fidelity representation, ~2M splats, exportable as .ply/.spz — plus triangle
meshes (low-fi collider meshes and high-quality visual meshes) and rendered video with
pixel-accurate camera control.
Pricing at launch (may have shifted since — Marble 1.1 shipped later): Free (4 generations), Standard $20/mo (12), Pro $35/mo (25 + commercial rights), Max $95/mo (75).
The “persistent” framing is the branch’s core argument against frame-by-frame generators: in Marble the world exists as geometry first, so consistency is structural rather than something a video model must remember.
RTFM — persistence without geometry
World Labs’ own counterpoint to Marble: RTFM (“Real-Time Frame Model,” announced Oct 16, 2025) is an autoregressive diffusion transformer with no explicit 3D representation — a “learned renderer” that keeps worlds persistent and 3D-consistent by using posed frames as spatial memory, at interactive frame rates on a single H100. One company, both answers to the persistence problem: geometry-first (Marble) and memory-first (RTFM). A World API (Jan 2026) exposes the text/image/pano/video → 3D-world pipeline programmatically.
The rest of the branch
Odyssey’s Explorer (Dec 2024) also outputs worlds as Gaussian splats, exportable to Unreal/Blender — see companies & labs for the business side. Niantic Spatial’s Large Geospatial Model (30B+ posed images) remains announced-but-unreleased. On the academic side, WorldScore’s leaderboard is instructive: 3D scene-generation methods (WonderWorld 72.69, LucidDreamer 70.40, and now Voyager 77.62 on the live board) beat every video model at static world-building — see evaluation. The unsolved trade: 3D methods score zero on dynamics. Detailed verification of the 3DGS generation papers (WonderJourney, CAT3D/4D, Bolt3D) is still pending.
Why we care
This is the branch Empirical AI’s own experiments live in: our prompt → panorama → depth → 3D Gaussian splat pipeline is a minimal, single-GPU version of exactly the Marble recipe — and it hits the same research problems (disocclusion, sky handling, parallax holes) at a thousandth of the compute.
← Video-generative world modelsAgents & robotics: world models for control →