Isometric NYC

3 months ago

The author embarked on a project to create an isometric pixel-art map of New York City using generative AI models.
AI coding agents (like Claude Code, Gemini CLI, and Cursor) were extensively used, reducing the need for manual coding.
Initial attempts with CityGML data were replaced by Google Maps 3D tiles API for better consistency.
Nano Banana Pro was initially used for image generation but proved inconsistent and expensive, leading to fine-tuning a Qwen/Image-Edit model.
An 'infill' strategy was developed to ensure seamless tile generation by staggering adjacent tiles.
Micro-tools were created to handle various tasks, such as visualizing tiles, classifying water, and generating training data.
Edge cases, particularly with water and trees, required manual intervention despite AI assistance.
Scaling up involved using Lambda AI for faster, cheaper model inference, enabling parallel generation of tiles.
Automation challenges highlighted the limitations of current AI in understanding complex algorithms and image editing.
The final application used OpenSeaDragon to display the generated tiles, though performance issues arose.
Key takeaways include the transformative potential of AI in reducing drudgery and unlocking new creative possibilities, but current image models still lack reliability and editing capabilities.

Hasty Briefsbeta