AI-built, SimCity-style NYC: what it teaches creatives about speed, style, and love for the craft
A single developer, Andy Coenen, set out to render New York City as an isometric pixel-art map-think classic SimCity vibes. The twist: he used a small stack of AI agents instead of placing pixels by hand. The result looks stunning at a glance (yes, even 432 Park Avenue shows up), but zoom all the way in and you'll see that familiar AI "goopiness."
That tension is the point. It shows where AI shines, where it still stumbles, and how creatives can keep their edge.
The project in a sentence
Coenen, a software engineer conducting AI research at Google DeepMind, aims to render all of NYC in an isometric pixel-art style. He leaned on AI tooling-Claude Code, Gemini CLI, Cursor-and says he "wound up writing almost no code." The map is generated tile-by-tile, then served in a way that lets you zoom without melting your machine.
How he built it (fast overview)
- Started with the idea of using Google's "Nano Banana" to turn satellite tiles into pixel art. Early runs struggled with consistency and hallucinations.
- Whiteboxed buildings using 3D CityGML data, but geometry didn't line up cleanly with top-down imagery.
- Switched to the Google Maps 3D Tiles API for tighter alignment and textures.
- Large models were too slow, costly, and stylistically jumpy. So he fine-tuned smaller: a Qwen/Image-Edit model on ~40 input/output pairs. It took ~4 hours and about $12.
- Built software to stream and zoom tiles efficiently. At city scale, the illusion works. At pixel scale, you can spot the AI artifacts.
The creative tension
Building this entire map by hand would take a lifetime. Automation turns that lifetime into weeks. But the tradeoff is obvious: AI gets you 80% there fast, then asks for judgment, taste, and cleanup.
Coenen's take is blunt: years of "dragging little boxes around on a screen" isn't creativity-it's grind. AI can end the grind. And there's a catch he embraces: "If you can push a button and get content, then that content is a commodity. Its value is next to zero. When hard parts become easy, the differentiator becomes love."
What this means for designers, illustrators, and art directors
- Use AI for scaffolding; keep final passes human. Let models block in forms, patterns, and lighting. You handle composition, storytelling, and polish.
- Prefer small, fine-tuned models for style consistency. A tight dataset (even ~40 examples) can beat a giant model for coherent output.
- Define your "zoom level of truth." If the work ships at city-view scale, don't obsess over pixel-level artifacts. If it's for close inspection, budget time for cleanup.
- Build a repeatable pipeline: source tiles → model pass → human review → batch fixes → QA for consistency. Treat prompts, datasets, and review checklists like a style bible.
- Cost/time matters. Track inference speed, GPU time, and unit costs per tile. Small models keep experiments affordable and fast.
- Own the voice. As content gets cheap, taste, intent, and a recognizable style become your edge.
Tools mentioned
- Claude Code, Gemini CLI, Cursor
- Qwen/Image-Edit (fine-tuned)
- 3D CityGML data
- Google Maps 3D Tiles API
Try this playbook on your next visual project
- Pick a target scale and style reference before you touch a model.
- Assemble 20-50 clean before/after pairs and fine-tune a smaller model.
- Automate the boring parts; keep the final pass for your eye and hand.
- Ship, learn, refine the dataset, repeat.
If you want a curated starting point for creative tooling, this list of AI tools for visual work is useful: Generative Art Tools.
Why this matters
AI is making production cheaper. That doesn't make your work less valuable-it raises the bar. The thing that separates forgettable content from work people talk about isn't the tool, it's the creator's taste, decisions, and care.
Use AI to clear the slog. Keep the love for the craft to make it worth seeing up close.
Your membership also unlocks:
