Research · AI · Spatial Computing · Procedural Generation · Research
Text-to-World Generation: Designed Worlds from a Sentence
Can a sentence compose a 3D world to an art director's standard? How H3M's worldgen system turns a prompt into a lit, finished space: looks stored as data, a catalogue of 1,576 reusable components, and an AI critic whose feedback became 45 art-direction rules.
· Ryan Brant
Generative AI can already produce 3D scenes from text. Most of them look generic. Designed spaces are different: they are lit, framed, and finished like a photograph. The question behind H3M's text-to-world research is simple to state and hard to meet: can a sentence compose a world to an art director's standard?
There is a second constraint. Worlds have to read as designed, not generated, and they have to stay light enough to run on a phone inside Portals.
Start from a sentence
A sentence either opens a prebuilt room, when it names one, or composes a new world when it doesn't. To set the bar, we wrote 120 scene prompts in four series, each one a brief an art director would recognise, from "a glamorous dinner before the guests arrive" to "pressure, preparation, and the moment before entry."
Light is the craft
What makes a room read as premium is light and grade. So in this system, light and grade are data, not code. There are sixteen looks so far, ten of them taken directly from the rooms' own lighting. A look can be applied to any set: the supper club's low tinted fill and flagged neutral key became a named look that other worlds can borrow.
Compose from parts
A room built in one piece teaches nothing. Built from named parts, every room grows a library that any future world can draw on.
Every piece in a room is a named, measured component with a role, tags, and a size. Across the 120 rooms, that produced a catalogue of 1,576 reusable components. The largest share is lighting (574 components), followed by architecture (358), decor (231), display (154), backdrops (83), sculpture (77), furniture (68), and textiles (31). A room's light rig is made of parts too.
The supper club, for example, is thirteen parts: a curved channel-stitched banquette 2.7 m across, an oversized pleated drum pendant, an etched smoked-mirror oval, and the lacquered shell that holds them.
Critique every clause
Taste has to scale. A critic model scores each render against its prompt, clause by clause, and the room is revised until the clauses pass.
Then we went a step further. The critic's reports were distilled into rules, each backed by the evidence behind it. 59 critiques became 45 art-direction rules, and 31 of them can check a plan in about 9 ms, before anything renders. Taste is taught offline and enforced at runtime.
Keep the world small
What the system keeps is a recipe of a few hundred bytes (the first spoken world was 403 bytes), which can be rebuilt at any of five levels of detail. Keep the words, not the geometry.
That same principle runs through Portals' persistence model, and since October 2026 text-to-world has been part of the Portals app's Composer: a prompt tray, a strip of the 120 rooms, and the world placed in AR as one item whose pieces can be selected, moved, turned, resized, and deleted.
What's next
- One library: room components, rebuilt library models, paint brushes, and looks in a single catalogue the composer picks from. - Placed like a set: pieces placed by the formula's set grammar and checked by its rules before a render. - Proven on a phone: frame rate, memory, and placement measured on device, not only in the browser. - Every scene to the bar: all 120 prompts are built; next, each is critiqued up to the target score, with levels of detail and at most 200K faces on mobile.
Every room teaches the library, and every sentence can draw on it.