TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Studio Practice · AGE 18—20 · FACTUAL · EVOLVING
Comparing drawings
Overlay the two issues yourself, because a vision model's summary of changes is a lead list.
When to use
A new issue of a drawing arrives and you must know exactly what moved before anything downstream moves with it.
The method
One: if you have vector files, compare geometry — overlay the two issues, or run a drawing-comparison tool; geometric comparison is deterministic once the two files are correctly registered and normalised — same origin, same units, same layer states — and is only as good as that registration. Two: if you only have scans or rasters, overlay at the same scale and flick between them, or run an image difference; expect noise. Three: a vision model's "summary of changes" is a lead list, never a finding — ask it for suspected changes, then check every one yourself against both sheets. Four: record each confirmed change in the revision register, clouded on the drawing, with the issue date. Five: anything the model did not flag is not therefore unchanged.
Watch for this
A multimodal model describing two plans as identical when a door swung the other way. Current benchmarks built from real drawings show these models counting doors, windows and rooms unreliably on plans; a door that moved is harder still.
Try it
Take two issues of one plan. Ask a vision model what changed. Then overlay them yourself. Score the model: found, missed, invented. Keep the score; it is your own measurement of a moving capability.
Prove it
Show one change register where every entry was confirmed against the two sheets, and say which entries the machine found and which it missed.
How it works
Drawings are a notation, and notation carries meaning in small marks — a swing arc, a hatch, a level tag. Image models trained mostly on photographs learn appearance, not notation, so they see a plan the way they see a picture of a plan. Vector comparison does not have this problem because it never looks; it subtracts coordinates. Its own failure is upstream: a shifted origin, a unit mismatch or a layer left off makes everything look moved, or nothing. That is why the method routes geometry to arithmetic and gives the model only the job it is good at: producing a list of places to look. EVOLVING because drawing-trained models are an active research direction and the benchmarks are re-run on every new model.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.