Fable 5.1 World Modeling

Original Article ↗ Hacker News Discussion ↗

AI-generated 3D city scenes are getting good enough to recreate real places like San Francisco’s Union Square directly from OpenStreetMap and terrain data, with agents writing Three.js code and even scripting NPC traffic and pedestrians. Commenters are impressed by the fidelity and potential for games, film pre‑production, and rapid prototyping, but debate whether this qualifies as true “world modeling,” how cost‑effective it is at ~$33 per scene, and how useful current AI-generated meshes and textures are beyond demos.

Demo and Implementation

  • Fable 5.1 was used to auto-generate an interactive 3D model of San Francisco’s Union Square in Three.js, with walkable streets, readable storefronts, traffic, cable cars, and interior spaces.
  • Geometry comes from OpenStreetMap and USGS 3DEP; the model writes code to convert this data into a Three.js scene.
  • NPC pedestrians and traffic logic are implemented in custom code; links to Pedestrians.ts and Traffic.ts show pathing and behavior systems.

Cost, Tokens, and Workflow

  • Reported run: ~2 hours, heavy subagent usage, self-QA loop, ~8M tokens, ≈$33 in API cost.
  • Generated in a single shot with a very long, carefully structured prompt that includes orchestration instructions.
  • Some commenters want more detail on duds, reliability, and babysitting; this remains largely unspecified.

Use Cases and Future Directions

  • Interest in using this approach for open-world and AR-style games, real economies, and benchmarking multimodal coding capabilities.
  • Others are experimenting similarly (e.g., RTS game, ski resort scene) and see major productivity gains for hobbyists and indie devs.
  • A Kyoto “hand-painted” style world has also been added; there is curiosity about scaling to whole cities and using cheaper models like GML-5.3-flash.

Debate Over “World Model” Terminology

  • Several commenters argue the title is misleading: “world model” is commonly used for deep ML models that learn physics/sensory structure, not code-generated 3D maps.
  • Others adopt a broader definition: anything that can simulate a world qualifies.

3D Asset Quality, Topology, and Tooling

  • Concerns: AI-generated meshes often have messy topology, excessive poly counts, and weak UV/texturing; many tools are “demo-only” for real games.
  • Counterpoint: code-generated geometry via primitives/CSG yields clean topology “by construction,” though texturing is still a gap.
  • Tools mentioned: Opus 5 (cheaper, but struggles at higher fidelity), Meshy, Tripo, Hunyuan3D-2, OSM2World; some claim recent “smart low poly” features now produce game-usable assets, others remain unconvinced without wireframe evidence.
  • Techniques discussed: using LLMs for low-poly silhouettes plus baked textures, using diffusion models for detailed assets, “socketing” structures and fitting detailed assets via PCA and bounding boxes, instancing and culling for performance.

Skepticism and Comparisons

  • Some see this as an impressive but not yet practically useful “AI made this” demo; others stress that this replaces thousands of manual modeling hours.
  • $33 per world is debated: expensive versus free GIS/OSM pipelines, but potentially cheap relative to human labor.
  • Additional critiques: inaccurate visual overlays in the repo, confusion about hosting an online demo, and difficulty using similar methods for precise architectural models from DXF.