本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

fable51-worlds

Worlds as code. A prompt in, a world you can walk out.

Claude Fable 5.1 agent swarms take a brief - a sentence, a photograph, a clip - then research the place, model it, render it, and check it. What ships is a plain Three.js app that opens in a browser.

No game engine. No proprietary 3D tiles. No downloaded meshes. Every building, storefront, sign, tree and traffic light is generated by code that lives in this repo.

---

⚔️ Head-to-head: Claude Fable 5.1 vs GPT-6 Astra

Same brief, same test conditions, two models. Each built Union Square from scratch, then the Astra world was filmed along the Fable walkthrough’s camera route and the two are shown side by side, unedited.

▶ Watch the head-to-head · 59 s · left GPT-6 Astra, right Claude Fable 5.1 · Fable 5.1 build · GPT-6 Astra build

---

Worlds

🌉 Union Square, San Francisco

▶ Watch the walkthrough · 59 s · 1920×1080 · aerial, plaza, Nintendo, lower level, Apple

⛩️ Higashiyama, Kyoto

▶ Watch the walkthrough · 54 s · 1920×1080 · seven scenes, Gion to Kiyomizu-dera at sunset

More worlds coming.

---

Any input → a world

Text, video or image. The brief names the subject and the style, and the same pipeline runs behind all three.

Input	A brief looks like	The world that comes back
📝Text	”Hanamikoji, Kyoto - as a hand-painted anime background”	A named place in a named style: real geometry on surveyed ground, with the look written to order
🎞️Video	a thirty-second walk-and-talk from a film	The set behind the shot, continued past the edges of frame - step off the camera path and walk the rest of it
🖼️Image	one photograph, any angle	The place in the frame as geometry: turn around, change the hour, light it differently

Style is as open as subject. The same street can come back as a cel-shaded anime plate, a photographic reconstruction or a night scene, because each world’s renderer is written for it rather than chosen from a list.

---

Roadmap

A world is the substrate. What stands on it is the point.

✅	Interactive scenes	Real places, walkable in a browser - this repo
🔜	Animation and video	Motion as code: shot lists, camera language and continuity held across a sequence minutes long, cut from the world itself rather than generated frame by frame - anime, and documentary camera-matched to the real location
🔜	Agent environments	Worlds as training and evaluation environments, where the same code that builds the place also supplies the reward and the verification signal - and, played straight, an RPG
🔜	4D, embodiment, robotics	Scenes that change over time, and simulation where the ground truth comes free with the geometry

---

How these are made

Every stage is in the repo, and every stage can be re-run:

1. Reconnaissance - parallel research agents pull map geometry, elevation, transit and street specs, and a storefront census, each fact carrying its source and a confidence level.
2. Asset generation - scripts emit the kit a world needs: façade modules, street furniture, vehicles, vegetation, fixtures. Some worlds ship no binary assets at all and draw every texture at start-up.
3. Runtime - a pure Three.js app assembles terrain, streets, façades, props, crowds and traffic from JSON specs.
4. Verification - Playwright drives the real app, screenshots fixed viewpoints, and diffs them against photographs taken from the same spot. Independent reviewer agents - architect, geographer, technical artist, interaction - file reports that drive the next fix cycle. Builders never grade their own work, and every world ships its own QA report, defects included.

---

Why code

A world written as code is not the same kind of object as a world generated as pixels or held in a latent space.

Verifiable	You can check it by running it. Every elevation in Higashiyama is an independent survey query; a walker drives the real movement code over the whole route; Playwright diffs fixed viewpoints against photographs taken from the same spot. A wrong number is a failing test, not a matter of taste.
Compositional	The parts are reusable and legible. A townhouse generator, a roof kit, a street-plot layout engine - each is a module with a contract, and a district is a few hundred lines that calls them.
Editable	”Make the shopfront recesses deeper” is a diff, not a re-roll. The world changes exactly where you asked and nowhere else, and the change survives into every later render.

License

Code and generated assets: MIT. Geometry is derived from OpenStreetMap (ODbL), USGS 3DEP (public domain) and the GSI elevation service. Reference photographs are not redistributed here - provenance is recorded per world. Brand names and logos identify the real businesses at their real locations and belong to their owners.