I needed an illustrated map for a website. Same aerial view of a coastal old town, but in the soft illustration style the rest of the site already used. Simple enough, I thought: take a photo, ask ChatGPT to turn it into an illustration, done.
It wasn't done.
The whack-a-mole phase
The first output looked like an illustration, but it wasn't my city. The model kept the rough shape and invented the rest: a clock tower that doesn't exist, monuments in the wrong place, a skyline that felt like a generic postcard instead of the real place.
So I did what everyone does. I asked for fixes.
"Move that marker." "That building is wrong." "Remove the tower."
And every fix quietly broke something else. Correct the tower and the harbor shifts. Fix the harbor and the park regrows in the wrong spot. Each prompt solved one problem and reopened two.
The problem wasn't the wording. I was asking the model to restyle and remember at the same time, and it couldn't hold the map still while it painted.
Prompting better actually helped
Before giving up I tried doing the boring part properly. A handful of rules that made a real difference:
- Start from the goal. Say what the image is for — cover, social, docs. The use case narrows the style.
- Describe the concept, not just the objects. What it should communicate, before what should appear.
- Name the visual style. Minimal, editorial, realistic, illustration. Vagueness gets you AI-default mush.
- Say what you don't want. No text, no people, no invented landmarks, no "AI look."
- Be specific about the change — and about what must stay. "Change X, keep everything else identical" is a very different instruction than "change X."
- Keep text to a minimum. Five words max, if any.
- Point at references you trust. Styles like Apple, Stripe, Linear, Vercel.
- Iterate one thing at a time. One change per turn. Batch edits are where regressions hide.
That last rule is the antidote to whack-a-mole, and it worked. The second illustration was closer.
But "closer" still wasn't the style I wanted, and the buildings kept drifting from the real ones. I was polishing a source that was never going to get me there. It's the same wall I hit the first time I used AI for real design work — the tool is capable, but it can only work with what you hand it.
You can't prompt your way out of a bad reference image.
The revelation: change the source, not the prompt
The unlock had nothing to do with prompting. I stopped feeding the model a flat photo and gave it a better starting point: Google Earth in 3D.
This is the part that's specific to maps and real places: they have a ground truth you can capture. Google Earth hands you that truth for free — real geometry, in 3D, from any angle.
Obvious in hindsight, and it changed everything. A 3D view gives the model real geometry — actual building volumes, the real coastline, the real park on the hill — instead of a single flat angle it has to guess depth from. The model no longer had to invent structure. It only had to restyle what was already correct.
Two settings mattered to get a clean plate:
- Rendering quality: Highest. Sharp buildings, no mushy textures for the model to misread.
- Turn off markers and labels. No pins, no place names baked into the image — just the city.
Anchor it to reality
With a faithful 3D capture as input, the prompt got shorter and firmer. The core instruction was about preservation, not creativity:
Keep every building, street and coastline exactly as they are. Do not invent monuments, towers or figures. Only change the rendering style to a soft hand-drawn illustration.
That's the whole trick with landmarks. Left alone, the model fills gaps with plausible-looking invention. Given real geometry and an explicit "don't invent," it stops hallucinating and just paints.
The first illustration from the Google Earth plate was immediately right.
Faithful to the real city, in exactly the editorial illustration style the site used. The thing I'd been chasing for a dozen prompts showed up on the first attempt — because the input finally deserved it.
Markers and the path, in the right order
The map still needed points of interest and a route between them. My instinct was to ask the AI to add them. Wrong instinct — that's asking it to invent placement, and invented placement is exactly what I'd spent all morning fighting.
Instead I added the markers and drew the path in Google Earth first, on the real 3D view.
Then I ran the same restyle prompt. The illustration inherited the exact positions and the route, now rendered in the same hand-drawn style as everything else.
Same lesson, one more time: put the information into the source, and let the model only change how it looks.
What I actually learned
Most of my early effort went into the wrong half of the problem. I kept refining instructions for a model that was doing its best with a weak input. It's a close cousin of the trap where you keep pushing an AI that was never going to get there on its own — the answer is rarely one more prompt.
The best prompt is a better reference image.
When an AI image keeps missing, the reflex is to write a cleverer prompt. Usually the faster fix is upstream: give it a source that already contains the truth — real geometry, real positions — and let the prompt do the one job it's genuinely good at, which is style.
Maps make this obvious, because the truth is right there to capture. But the same move works for anything that has to stay faithful to something real — a product, a floor plan, an actual place: find the closest truthful source and start from it, instead of asking the model to invent from a weak one.
Fix the input first. The prompt gets easy after that.
Top comments (0)