WP6·Recommender for urban (re)design idea

From a gap to a picture.

An idea for the protocol's recommender: find what a street lacks, check it twice, picture it, and check the picture. Every edited image is an AI visualisation, not a design.

Start ↓All 47 pictures →
01 / 05

Where the gaps are.

Read from views along the street, using only tier-A features or segmentation.

–of usable views show this gap
02 / 05

Where to look first.

Candidates come only from WP5's priority buurten: little green, many low-income households.

124candidate views in 20 buurten
03 / 05

Confirm before you draw.

A larger model checked each gap. Two thirds of the "missing crossings" did not hold up.

04 / 05

Then picture it.

An open image model adds the missing feature and should leave the rest alone. Drag to compare.

05 / 05

And check the picture.

Both audit models read every edit blind. Making a convincing image is easy; choosing where it is justified is the hard part.

Ask the evidence

Answers that must cite their views.

For fifteen questions, a larger open model wrote short answers from the matched views only; a sentence that names no view withholds the answer (GeoAI-VLM's grounded answering). New questions here list the matching views, as this server has no GPU.

Details
How the edits were made and checked

Qwen-Image-2.1-Turbo (revision d65dbc9a), 8 steps, 1024 × 768, one fixed prompt per gap type plus the same instruction to keep everything else unchanged. Each edit was resized to 768 × 576, then audited blind by both models and segmented. Other features changed state 2.55 times per edit on average, against 0.68 expected from the small model's run-to-run instability; some are intended side effects (trees bring shade). Pixels changed in classes the edit should not touch: median 1.6%, 90th percentile 9.9%.