TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Generative Mechanics · AGE 17—19 · FACTUAL · EVOLVING
Why 'left of' fails
The captions the models learned from rarely say where things are.
The idea
Ask for "a stair left of the door" and many image models give you a stair and a door in any arrangement. The captions the models learned from rarely say where things are. People write "a stair and a door", not "a stair to the left of a door". Spatial words are rare in training data, so the model has had little practice. That is one cause, and it has been measured: spatially rich captions closed some of the gap. It is not the whole cause. How the model binds words to objects matters too.
Why it matters
If part of the failure is a data gap, you can work around it today with a control image or a sketch, and expect that part to shrink tomorrow. If you think it is random, you can do neither.
See it in the studio
"Courtyard at the centre, kitchen to the east, verandah facing the street" comes back as a handsome house with the kitchen anywhere. The words that describe a plan are exactly the words the model has seen least.
Watch for this
Writing longer spatial descriptions to fix a spatial failure. Using the rare words more often does not make them common to the model. Draw it instead.
Try it
Prompt "a stair" ten times, then "a stair left of the door" ten times. Score how often the second set is actually left. You now have a number for your tool.
Prove it
Explain why "a stair left of the door" is harder for a model than "a stair", and give one reason the same failure can improve with better training data.
How it works
The SPRIGHT work re-captioned millions of images with spatial detail and measured better spatial accuracy after training on them. CoMPaSS and others continue that line of work. EVOLVING for exactly that reason: the gap has been measured and it is moving. SPATIAL TEST on this strand gives you the procedure to measure it yourself.
What this idea builds on
What this idea opens up
Sources
- Chatterjee et al., 2024 — Getting it Right / SPRIGHT (arXiv:2404.01197)
- TEC.U
- essential-ai-concepts
- CoMPaSS
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.