TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 18—21 · POSITIONAL · HELD
The model forgets the room
A consistent subject is not a held space, and the promenade is the test.
Our position
Today's image models can keep a subject looking consistent across separate images: the same chair, the same fluted timber, the same person. That is subject consistency, and it is real progress. It is not a room you can walk through. Ask to step past the door and the space changes. The window that was on the left is gone, the ceiling height changes, the corridor leads somewhere new. We hold that the test for a spatial designer is the promenade — Le Corbusier's promenade architecturale. Does one room still anchor the next as you move? A single striking view cannot pass that test. A consistent subject cannot pass it. Only a space that is held can.
Why we hold it
Architecture is experienced in sequence. A tool that produces beautiful, unconnected views is a tool for pictures of buildings, not for buildings. Naming the promenade as the bar keeps the profession's own standard in view when a demonstration of consistency is presented as a demonstration of space.
The strongest objection
We may be applying a world-state test to a system that was never built to hold a world. A single-image generator's job is to produce independent 2D samples. When it fails the promenade, that is not evidence that it is bad at its task; it shows that 2D subject consistency and persistent 3D scene state are different capabilities, and we are testing one tool for the other's job. "The model forgets the room" then says less about the model than about which tool we asked. The systems actually built to hold a scene — video models, world models — are the ones the test should be run on, and on those the bar is moving quickly. A designer can also supply the persistence the model lacks, by using a 3D massing as the control input for every view.
What would make us revise it
Two things, separately. If the objection holds — if the promenade is being asked of tools that never claimed it — the position narrows to "run the promenade on systems that claim scene state, and do not read a consistent subject as a held space", which keeps the test and drops the complaint. And a repeatable promenade test passed by such a system alone — a sequence of views through one interior, generated without a 3D control, in which adjacency, openings, heights and materials hold across every step, as judged by architects — makes this a history card. Reviewed every edition. We run the test ourselves.
Try it
Generate a view of a room. Then ask for the view from the doorway on the left of that room, then from the next room looking back through the door. Lay the three in sequence. Mark every element that changed. Then do it again with a massing model as control and compare.
Take it to crit
Tell subject consistency apart from true spatial persistence in your own sequence, and name the promenade as the harder test. Did your sequence pass it?
How it works
Subject consistency is achieved by conditioning on reference images or identity embeddings. It keeps the thing, not the scene. Holding a scene needs a representation of space that the model carries between frames. That is what HOLDING TIME and WORLD MODELS on this strand are about, and why they are harder.
Lineage
The promenade architecturale is Le Corbusier's phrase, from the Villa Savoye years. It means movement organised so that spaces and views unfold as you advance. Charles Correa later wrote about movement through space in the Indian context, and the two are sometimes confused — including in the Lab's own essay on spatial memory, which credits Correa. The term is Le Corbusier's. The test works under either name.
What this idea builds on
What this idea opens up
Sources
- The Architecture of Memory: Navigating… — the-architecture-of-memory-navigating
- “Promenade architecturale” is Le Corbusier’s term (c.1929, Villa Savoye), per Fondation Le Corbusier — not Charles Correa’s; today’s models demonstrate subject/character consistency, not navigable spatial coherence
- Fondation Le Corbusier, Villa Savoye
- OpenAI, Video generation models as world simulators
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.