Ideogram 4
Ideogram 4 state of the art open weights
In the last post an agent chained two Hugging Face Spaces to turn prompts into 3D monuments. This time we pushed one step further: what if the generated assets were not the destination, but the vocabulary of something bigger?
The result is Infinite London: every visit procedurally collapses a brand-new Victorian city out of 36 AI-generated 3D tiles. Streets connect, garden squares open up, Big Ben spawns rarely on the skyline, and the world materializes forever at the horizon as you roam it with WASD.
Text-to-3D is slow: minutes per asset on a shared GPU. You cannot generate a city on demand. But you do not need to. The trick is an old one from game development: Wave Function Collapse (WFC), an algorithm that assembles modular tiles into endless coherent arrangements by matching edge sockets, the way a sudoku solver fills a grid: every cell starts as a superposition of all possible tiles, and observing one collapses its neighbors.
WFC has always had one bottleneck: someone has to make the tileset. That someone used to be a 3D artist. Now it is a prompt loop:
ideogram-ai/ideogram4
generated each tile as an isometric game-kit specimen on black. One prompt per tile.microsoft/TRELLIS.2
reconstructed each image into a textured GLB, which we then simplified ~12× for
instanced rendering in the browser.Both Spaces were driven end-to-end by an agent reading their
agents.md specs.
*Every Gradio Space on the Hub has an Agents button: copy the agents.md instructions straight into your coding agent.* The split of
labor is the point: **36 expensive generations, cached once. Infinite cities, free,
in the browser, forever.** The marginal cost of a new world is a random seed.
The 36 tiles follow a simple socket grammar. Each edge is either open paving or street, and WFC only places tiles whose edges agree:
Every tile is an inspectable GLB on the Hub. Here is the black-cab stand, fresh out of the pipeline:
The first attempt at a tileset failed in an instructive way. Prompting for picturesque scenes produces beautiful images that reconstruct into incoherent worlds: backdrop walls become floating planes, vignette bases tear the grid, and every cell boundary is a visible seam.
The fix was to prompt like a modular-kit artist instead of a photographer: flat slab grounds, routes running flush edge to edge, solid freestanding buildings complete from all four sides, perfect square footprints, no backdrops. Same models, same pipeline. The difference between chaos and a city was entirely in the design language of 36 prompts.
The browser side is plain three.js, no build step. A few details that make it feel alive:
seed + chunk coords, so walking back always finds the
same streets. Far chunks are evicted; the world is infinite in every direction.Roam the city, press R for a new world, or paste an
agents.md link into your own
coding agent and generate a different city entirely. A new theme is 36 prompts away.
👉 mishig/infinite-london · sibling galleries: Paris · Japan · Egypt
Ideogram 4 state of the art open weights
High-fidelity 3D Generation from images
Explore an endless, procedurally generated 3D London
Explore interactive 3D models of Paris monuments
Explore 3D Egyptian monuments with interactive rotation
Explore 3D captures of Japan’s famous monuments