fable51-worlds: three generated places, and a verification stage that makes a wrong elevation fail a test

Every world writes its own renderer, so a style cannot travel between them

The claim that style is as open as subject is true, but the mechanism behind it has a cost.

The same street can come back as a cel-shaded anime plate, a photographic reconstruction or a night scene, and the reason given is that each world's renderer is written for it rather than chosen from a list. So there is no style picker. A renderer is authored per world.

That is the right choice for quality, since a shared style system would be the thing forcing every world into the lowest common denominator of what one renderer can do. It also means the three properties the project claims for code worlds are paid for at world granularity rather than project granularity.

Composability is stated inside a world, not across worlds. A townhouse generator, a roof kit and a street-plot layout engine are modules with contracts, and a district is a few hundred lines calling them. Put a second district in the same world and it reuses those. Put it in a different world and it does not.

The editability claim is the strongest of the three and the least world-specific. Making the shopfront recesses deeper is described as a diff rather than a re-roll, so the world changes where you asked and nowhere else, and the change survives into every later render. That property comes from the world being code, and it holds regardless of how many renderers exist.

The head-to-head pins both models to one camera route and one shot timing

The comparison design is the most careful thing on the page, and it is worth reading as methodology rather than as marketing.

Where a competing model was given the same brief, its world was filmed to the Fable walkthrough's camera route and shot timing. The two then run side by side, unedited, with the competing build on the left and Claude Fable 5.1 on the right.

Fixing the camera is what makes the comparison say something. Without it, a difference between two videos could be a difference in cinematography rather than in the world. With it locked, the only variable left is the geometry, the materials, the lighting and the dressing.

It also means the competing build is doing the same job rather than something easier. It is a code-generated world, not a rendered clip, which is why there is a Codex source and viewer directory linked alongside the competing footage for the Kyoto world.

The durations differ per world because the walkthroughs do: just under a minute for Union Square, a little over fifty seconds for Higashiyama, and under half a minute for the trench run. The trench run is also the shortest because it is a sequence of action beats rather than a walkable district, and it is the one world briefed by a film clip instead of a photograph or a sentence.

Builders never grade their own work, and every world ships a QA report

The verification stage has a rule attached to it, and the rule is the reason to believe the rest.

Playwright drives the real application, screenshots fixed viewpoints, and diffs them against photographs taken from the same spot. Then independent reviewer agents file reports, and those reports drive the next fix cycle. The four reviewer roles named are architect, geographer, technical artist and interaction.

What is prohibited is builders grading their own work. Each world ships its own quality assurance report with defects included, which is the part that makes the claim checkable from outside.

The verifiable property follows from this arrangement. Every elevation in the Kyoto world is described as an independent survey query rather than something an agent estimated and moved on from. A walker drives the real movement code over the whole route rather than replaying a recorded path. And a wrong number is called a failing test rather than a matter of taste.

That last phrase is the load-bearing one. Most generated 3D work fails by being plausible, and plausibility is precisely what a photograph diff can catch and a human looking at a flythrough cannot. Whether the diffs are strict enough to be useful is a separate question, and one the repository invites you to answer by running the world.

MIT covers the code; the geometry carries three different licences

The licensing section is more careful than a single licence line, and the split matters.

Code and generated assets are MIT. Geometry is derived from three sources under three different terms: OpenStreetMap under its share-alike database licence, a national elevation dataset in the public domain, and a national elevation service. Reference photographs are not redistributed at all, with provenance recorded per world instead.

So the MIT grant does not extend to the map data, which is where a share-alike obligation lives. That is worth reading as what it is: a clean code licence with the data terms stated separately rather than papered over.

The trademark paragraph is separate again and narrower than the licence one. Brand names and logos identify the real businesses at their real locations and belong to their owners, which is the situation in the Union Square world where the walkthrough passes a real store. Rather than blurring those signs, the world keeps them and attributes them.

Two of the three worlds are real places with real businesses in them, so that paragraph is doing actual work rather than boilerplate. It is also the paragraph that runs out of room, since it begins the unaffiliated-homage note for the third world and stops partway through a sentence about film footage.

The top level is three worlds and three comparison builds, and nothing else

The repository has nine entries: a gitignore, a licence file, a readme, and six directories. There is no package manifest, no lockfile, no test directory at the root, no continuous integration configuration, and no build script visible at this level.

The six directories are three worlds and three matching comparison builds. A world directory and a world-plus-a-suffix directory sit side by side for San Francisco, for Kyoto, and for the trench run. The suffix used in the names refers to the competing model, and the page links each one as the competing build alongside the Fable build.

That layout has a real implication for anyone who wants to run one of these. There is nothing at the root to install. Each world directory has to carry its own way of being served, which is consistent with the description of each as a plain application that opens in a browser, and with the total absence of a build tool at the repository root.

It also means the repository is three deliverables and three exhibits rather than a library with three examples. The comparison material sits at the same level as the thing being compared, which is the correct elevation for a portfolio and a poor one for a dependency.

The absence of root-level tests is consistent with the testing story rather than at odds with it: the checks described are per-world and run inside each world directory, driving that world's application with a browser driver.

One world was briefed by a film clip, and the trademark note stops mid-sentence

The third world is the odd one out in three separate ways, and each is a deliberate choice rather than an inconsistency.

Its input is not a place. It is a clip from a film, and the page spells out what that means for the pipeline: build the set behind the shot and continue it past the edges of frame, so you can step off the camera path and fly the rest of it. The brief itself is checked in as a separate document, and the comparison has its own method document beside the footage.

Its output is smaller. The two real-place walkthroughs are 1920 by 1080; this one is 1280 by 720. It is also structured as beats rather than as a district: approach, S-foils, interception, dive, trench run, torpedoes, fireball, shockwave.

And it is the one that needs a trademark note. The page states the world is an unaffiliated homage, and names the relevant trademarks and their owner. Then the sentence about film footage begins and the page ends.

What is written so far is only the negative half of that sentence, so a reader is left with an unambiguous statement that this is a homage and an unfinished statement about footage. Anyone about to reuse this world should confirm the position on included footage rather than infer it from the truncated clause.

One of five roadmap rows is shipped, and the rest are dated as coming

The roadmap table has five rows and one check mark. Interactive scenes, described as real places walkable in a browser, is the row that points at this repository.

The other four are all marked as coming, and each is a substantial increase in scope rather than an increment on the first.

Animation and video asks for motion as code, with shot lists, camera language and continuity held across a sequence minutes long, cut from the world itself rather than generated frame by frame. Two variants are named: anime, and documentary camera-matched to the real location.

Agent environments asks for worlds as training and evaluation environments, where the same code that builds a place also supplies the reward and the verification signal, and, played straight, an RPG.

Then two rows about measurement and physics: scenes that change over time and simulation where the ground truth arrives with the geometry, and a row about measuring how world quality changes with more inference time and compute.

That last row is the one worth noting for its own sake. Asking how output quality scales with inference budget is a research question rather than a feature, and putting it on the roadmap says the project intends to be evaluated on that axis. The ground-truth claim in the simulation row follows from the same premise as the verification stage: if geometry is code and elevations are survey queries, then the correct answer is in the repository.

Some worlds ship no binary assets and draw every texture at start-up

The asset generation stage is the one with the most interesting variation, because it does not do the same thing twice.

Scripts emit the kit a world needs: facade modules, street furniture, vehicles, vegetation and fixtures. Then the note: some worlds ship no binary assets at all and draw every texture at start-up.

A repository that refuses downloaded meshes and then also refuses bundled textures has pushed the constraint all the way through. If textures are drawn rather than shipped, they are code, which means they are diffable, which means the editability claim extends to surfaces as well as to geometry.

The runtime stage is what consumes all of it. A pure application assembles terrain, streets, facades, props, crowds and traffic from JSON specifications. So the pipeline is: research produces facts with sources, scripts produce modules, JSON carries them to the application, and the application draws them.

Reconnaissance is the stage that makes the verification claim possible. Parallel research agents pull map geometry, elevation, transit and street specifications, and a storefront census, and each fact carries its source and a confidence level.

That confidence level is the piece that connects everything. A survey query with a stated confidence can be checked against the source it names, which is what turns a generated number into something a reviewer agent can rule on.

Editorial conclusion

fable51-worlds is a portfolio of three finished worlds rather than a tool, and its most transferable idea is the verification stage: screenshot fixed viewpoints, diff them against photographs taken from the same spot, and let reviewer agents that did not build the work file the defects. That discipline is what lets a wrong elevation become a failing test rather than a matter of taste. Two things to hold in mind. Each world writes its own renderer, so a style does not travel between worlds and a fourth place costs a fourth renderer. And the MIT licence covers code only, with the geometry arriving under three separate data licences, one of which is share-alike.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. PhiloLabs/fable51-worlds on GitHub
4. README