WORDS → COMPOSITIONS → NEURAL WEIGHTS
Describe something.
Build a world around it.
A small language model assembles new block sculptures. A neural engine makes them solid, playable, and visible.
Move the gold player with WASD / arrows. Space jumps. Q / E rotates.
A creative model on your device.
One-time download of about 280 MB, plus the browser runtime. Saved locally when storage allows. Requires WebGPU. Your prompts stay on this device.
For unfamiliar objects, describe their parts, sizes and relationships. Generation takes seconds; the world keeps playing.
The starting arch and stairs are an authored demo. Walk around while the model loads.
In this world
Loading…
Refer to these names when editing. Unmentioned groups should stay in place; use Undo if the model gets it wrong.
Scene recipe · inspect or edit directly
Recipes create new compositions from boxes and group operations. They compile to neural weights. Use attach for touching parts and place for relative group positions. This editor works without the language-model download.
What this experiment can do
The model generates compositions from reusable operations. It was adapted with synthetic scene programs, without image or video training. It can miss details, misread references, or produce invalid recipes. Novel subjects are experimental. Visuals are block sculptures, with a limit of 48 parts and four camera angles.
Rendering, movement and box collisions run as neural networks. A conventional compiler converts the generated recipe into weights after each edit. This is a generative scene editor backed by a neural engine, not an unrestricted video world model.
Measured results & limitations ↗ · Generated successes & failures ↗ · Technical notes ↗ · Source ↓ · Model attribution ↗
