LESSON wm.3.2 · Paper Study · 140 min

Fei-Fei Li, Marble, and spatial intelligence

Spatial intelligence requires models to understand and generate navigable, revisitable, editable 3D environments rather than a video from one prescribed camera path. Marble accepts

DIRECT ANSWER · VERIFIED SOURCES ·

How does World Labs spatial intelligence differ from video generation and 3D reconstruction?

Spatial intelligence requires models to understand and generate navigable, revisitable, editable 3D environments rather than a video from one prescribed camera path. Marble accepts text, images, panoramas, multi-view input, or video and creates spatially persistent worlds that can be explored on the web or exported. Unlike reconstruction, which recovers an observed scene, a generative world can invent and extend unobserved space, adding geometric and factual uncertainty.

Key takeaways

  • Video generation models temporal imagery; persistent worlds must also maintain geometry and object state across viewpoints.
  • 3D reconstruction prioritizes faithful recovery, while generative worlds also complete, create, and edit.
  • Spatial consistency needs closed camera loops, revisited locations, and geometric measurements—not screenshots alone.

Boundary and caveat

Navigability does not imply a complete physical dynamics model. Marble scene generation and an action-interactive robotics world model remain different capability levels; persistence, geometry, and simulation fitness require task-specific evaluation.

Primary sources

Learning objectives

  1. Explain what problem “Fei-Fei Li, Marble, and spatial intelligence” solves without hiding behind terminology.
  2. Trace the variables and causal links across persistent 3D, multimodal creation, navigation.
  3. Complete “Compare video generation, 3D reconstruction, and generative worlds” and judge the result with evidence rather than intuition.

Core concepts

persistent 3D

persistent 3D requires geometry and object state to survive viewpoint changes. Navigability, revisiting, and editing are stronger than local video realism and need independent closed-loop trajectory tests.

multimodal creation

multimodal creation is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.

navigation

navigation requires geometry and object state to survive viewpoint changes. Navigability, revisiting, and editing are stronger than local video realism and need independent closed-loop trajectory tests.

editing

editing maps text, images, multiview input, or video into an explorable world and may extend or modify it. Consistent conditioning does not guarantee correct geometry; test spatial closure, edit locality, and uncertainty in unobserved regions.

and the World API

and the World API requires geometry and object state to survive viewpoint changes. Navigability, revisiting, and editing are stronger than local video realism and need independent closed-loop trajectory tests.

Build and verify

Compare video generation, 3D reconstruction, and generative worlds

  • Predict: write the expected output, trend, or failure before running code.
  • Build: implement only the minimum components needed to answer the question.
  • Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
  • Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.

Open the complete interactive lesson