|
TLDR
There are two ways to make an AI visual: diffusion paints pixels, code writes a renderer. Fable 5 showed the code path can build 3D worlds and games that diffusion simply cannot.
• Fable 5 built playable 3D games and navigable, photorealistic worlds from single prompts, entirely in code.
• No diffusion was involved: every mesh, texture, and light was generated by a program at runtime.
• For 3D, interactive, parametric, and reproducible work, the code path does what diffusion cannot.
• The model is offline now, but the approach is portable: any capable coding model can do a version of it.
|
|
•••
|
|
There are two fundamentally different ways to make an image with AI, and most of us only use one of them. The first is diffusion. You give a model a description and it paints pixels directly, the way an image or video model does. It is what we mean, almost always, when we say AI image generation. The second is stranger, and it is the one Claude Fable 5 demonstrated at a startling level before it was switched off last week: the model does not paint the image. It writes the code that generates it.
The distinction sounds technical, but it changes everything about what is possible. In the days Fable 5 was live, people did not just ask it for pictures. They asked it for worlds. And it built them, entirely in code, with no diffusion anywhere in the process.
The model itself is gone now, pulled by the order we covered last issue. But what it proved is not gone, and it is portable to other tools. So it is worth understanding this other path to a visual, what it does that diffusion cannot, and where it fits in your work.
|
|
What It Built
Worlds, not pictures.
|
|
The demonstrations were hard to believe. From a two-sentence prompt, Fable 5 produced a full racing game with four tracks, character models, music, and menus, in about fifteen minutes. From a short description, a navigable recreation of Yosemite Valley with procedural forests and water. One developer had it build a four-kilometer open world that runs in a browser, where every mesh, every texture, and every light is generated by code at startup, with no image or model files anywhere in the project. Another rebuilt a photorealistic Mediterranean village, its plaster and cobblestones and sea all generated procedurally.
Notice what these have in common. None of them is a picture the model painted. Each is a program the model wrote, which then draws the scene, in real time, in three dimensions, that you can move through. The model reasoned about geometry, light, materials, and structure, and expressed all of it as code.
|
|
Built on Existing Tools
The stack was already there.
|
|
Here is the part worth underscoring, because it changes how you read those demos. Fable 5 did not invent a new way to make graphics. It wrote code for tools that already existed and that creators have used for years: Three.js, WebGL2, and WebGPU, the standard graphics stack of the modern browser. People have long built epic real-time scenes, simulations, and videos on exactly these tools, work that rivals what offline render engines produce.
That is the real significance, and the reason this outlives one model. The capability was never locked inside Fable. It lives in a mature, openly available stack that runs in any browser, and that has been quietly reaching render-engine quality for a while. What Fable changed was the doorway: it let someone describe a scene in a sentence and have the code written for them, instead of writing it by hand. The ceiling was already high. The barrier to reaching it is what dropped.
|
DIFFUSION VS CODE
Diffusion paints a flat image from a description, and the result is fixed. Code writes a program that renders the visual, and the result is in true 3D, navigable, interactive, adjustable by changing a number, and reproducible from a seed. Two different paths, good at two different things.
|
|
Why It Challenges Diffusion
What code can do that pixels cannot.
|
|
For a whole class of work, the code path does things a diffusion model structurally cannot. It produces real three-dimensional space you can move through, not a flat image that only looks 3D. It produces interactivity, a playable game rather than a clip of one. It gives you precision and control, because every element is defined explicitly rather than coaxed out of a prompt. And it is infinitely adjustable: change one parameter and the whole scene shifts, the terrain reshapes, the season changes, because it is generated by rules, not baked into pixels.
There is also reproducibility, which diffusion has always struggled with. A code-generated world rebuilds itself identically from a single seed, every time. For anyone who has fought to get a diffusion model to produce the same thing twice, that alone is a different way of working. This is what people mean when they say the code path challenges diffusion. Not everywhere, but in 3D, interactive, parametric, and reproducible work, it offers something the painted image cannot.
|
|
Where Diffusion Still Wins
The honest balance.
|
|
This is not a story about one approach replacing the other. For most of what creators make day to day, diffusion is still the right tool and will stay so. A single photorealistic portrait, an organic subject, a painterly illustration, a quick one-off image with a particular feel: diffusion produces these beautifully and instantly, and writing code to render a convincing human face is far harder than prompting one.
The two paths are complements, not rivals. Diffusion is for the painted image, the mood, the single frame, the organic and the photographic. Code is for the system, the world, the interactive, the precise, the parametric. The creators who get the most out of this moment are the ones who know which path a given job wants, and reach for the right one.
|
|
The Catch
Read the demos honestly.
|
|
A note of restraint is warranted. The viral demos were astonishing, but they were demos, often the most polished run someone chose to share, and independent verification of the most ambitious ones is still thin. Consistency at production scale is unproven, the big generations consumed real compute and cost, and the model that produced them most impressively is, for now, offline. Treat what you saw as a glimpse of what the code path can do, not a promise that it does it reliably every time.
But the glimpse is the point. The approach does not depend on one model. Any capable coding model can write a renderer, generate a procedural scene, or build a small interactive world, and that capability is only going to deepen. The specific tool was switched off. The path it revealed is wide open.
|
|
The Bigger Picture
Stop prompting pictures. Start describing systems.
|
|
Step back and the real shift is in what you are describing. With diffusion, you describe a picture and hope. With code, you describe a system, the rules, the parameters, the behavior, and the system makes the visual, reliably and adjustably. That is a different kind of creative act, closer to designing a machine than painting a frame, and it has been quietly available this whole time to anyone willing to think that way.
So the next time you reach for a diffusion model out of habit, pause and ask which path the work actually wants. If you need a single beautiful image, prompt for it. But if you need a world, a system, something in 3D, something you can turn the dials on, consider describing the machine instead. The tools to build it from a sentence are here, they are spreading across models, and the creators who learn to think in systems will be making things the rest are still only prompting for.
|
|
•••
I am putting together a worldbuilding pack, a set of starting prompts for the code path: navigable 3D scenes, procedural landscapes, small interactive worlds, and parametric graphics you can adjust by changing a number, each one a prompt you can paste into a capable coding model, no coding background needed.
Want it when it ships? Reply with send me the worldbuilding pack and I will get it to you.
|
|
A QUESTION FOR YOU
Would you rather prompt an image, or describe the system that makes it?
Reply and tell me which path tempts you. For some work the code path changes everything; for some it is needless complexity, and I am curious where you land.
If this was useful, forward it to a creator who assumes every AI visual comes from a diffusion model.
|
|
Until next time,
Luxe Prompting
|
|
Luxe Prompting
AI Image Generation for Creators
|
|
