TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 18—20 · FACTUAL · EVOLVING
The perspective can lie
Separate regions of one image can follow different cameras, so trace the lines.
The dilemma
A generated render arrives, polished and atmospheric. Every surface looks real. Do you trust the space it shows — its depth, its heights, its room sizes — or do you spend ten minutes with a ruler on lines the client will never look at?
The choices
Trust it, because it looks right and time is short. Test it: trace the edges of floors, lintels and parapets, extend them, and see whether lines that are parallel in the world meet at one vanishing point on one horizon. Or generate the render with a control input that fixes the perspective from your own model, so the test is built in.
The consequence
Trust a broken perspective and the space you present does not exist. The room reads deeper than it could be, the ceiling higher, the street wider. Someone will measure it later — a client, a contractor, a jury with a scale rule — and the error becomes yours. Test it and you lose ten minutes and gain a checking habit. Build it in and you give up some freedom for a scene that holds together.
The case
A final-year student presents a generated interior of a double-height library. The juror asks for the section. Drawn against the render, the gallery's floor line vanishes to a point a metre above where the ground-floor window heads vanish. The two storeys were painted in two different rooms. The student had not traced a single line.
Take it to crit
Handed a slick render, did you test its perspective construction before trusting the space? Show the traced lines.
How it works
Diffusion models learn what perspective looks like locally — edges that recede, shadows that lengthen — but they have no single camera model for the whole image. So separate regions of one image can follow different cameras. Work such as ControlVP adds vanishing-point constraints to generation. That is evidence that the failure is common, and a sign that it is being engineered away. EVOLVING for that reason.
What this idea builds on
What this idea opens up
Sources
- Okumura, Shiohara & Yamasaki, 2025 — ControlVP (arXiv:2512.07504)
- TEC.U
- practical-applications
- Diffusion — How AI Paints from Noise
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.