The Isometric SF map renders 1,300 square miles of the Bay Area in a retro pixel style, highlighting major landmarks.
*A developer turned 1,300 square miles of Google 3D tiles into a pixel‑art map of the San Francisco Bay Area. The project showcases how open‑source AI models can cut production costs to a fraction of commercial alternatives.*
A hacker‑news post this week unveiled a pixel‑art rendition of the entire San Francisco Bay Area. The developer transformed 1,300 square miles of Google Photorealistic 3D Tiles into a single, scrollable map that feels like a retro video game. Landmarks such as Apple Park, the Palace of Fine Arts, and UC Berkeley sit side by side in a cohesive, low‑resolution world. The project doubles as a technical case study, showing how open‑source AI models can be fine‑tuned to produce high‑quality visual assets at a fraction of the cost of commercial services. The creator’s blog details the pipeline, model choices, and cost calculations, offering a blueprint for others who want to blend geographic data with artistic rendering.
The creator scraped Google Photorealistic 3D Tiles covering 1,300 square miles of the Bay Area. Those tiles were fed into Nano Banana 2, which trained a Qwen‑Image‑Edit 2511 LoRA to prototype the pixel‑art style. For the final version, the developer switched to Qwen to train a Klein 4B LoRA, reproducing the same aesthetic at one‑quarter the cost. The workflow is documented in a technical blog post, which details the data pipeline, model fine‑tuning, and rendering steps. The result is a cohesive, navigable map that preserves geographic relationships while rendering everything in a retro pixel palette.
The pixel map spots several iconic Bay Area sites. Users can locate Apple Park, the corporate campus of the tech giant, rendered in blocky green. The Palace of Fine Arts appears as a stylized dome surrounded by water. The University of California, Berkeley campus is marked with its distinctive spires. Each landmark is positioned accurately relative to the surrounding streets and terrain, allowing explorers to recognize familiar silhouettes despite the low‑resolution aesthetic. The map’s interactivity invites users to wander and discover these points of reference, turning a technical demo into a playful geographic tour.
Two open‑source models formed the backbone of the project. First, Nano Banana 2 trained a Qwen‑Image‑Edit 2511 LoRA, establishing the visual language for the prototype. The second stage employed Qwen to train a Klein 4B LoRA, which delivered the final style at a quarter of the original cost. Both LoRAs are lightweight fine‑tuned adapters that modify the base models without requiring full retraining. The developer highlighted the cost efficiency of this approach, noting that the Klein 4B LoRA achieved comparable quality while dramatically reducing compute expense.
If the map had been produced entirely with Nano Banana 2, the estimated expense would have been around 7,000 USD. By transitioning to the Klein 4B LoRA trained on Qwen, the developer reduced that figure to roughly one‑quarter, saving several thousand dollars. The cost differential underscores the financial advantage of leveraging smaller, open‑source models for large‑scale visual projects. It also demonstrates that high‑quality, region‑wide pixel art can be generated without the budget of commercial AI services, opening the door for similar community‑driven mapping efforts.
The map covers 1,300 square miles of the San Francisco Bay Area.
Nano Banana 2 trained a Qwen‑Image‑Edit 2511 LoRA for the prototype, and Qwen trained a Klein 4B LoRA for the final map.
Using Nano Banana 2 alone would have cost around 7,000 USD.
The Isometric SF map proves that open‑source AI can deliver large‑scale visualizations without breaking the bank. By swapping a costly model for a lean Klein 4B LoRA, the developer saved thousands while preserving artistic intent. The result is both a functional map and a proof‑of‑concept for cost‑effective AI‑driven cartography. As more creators adopt similar pipelines, we may see a surge in community‑generated, high‑resolution maps that challenge the dominance of proprietary tools.
Sources: Hacker News, https://isometricsf.com/, https://www.justinchavez.com/blog/isometric-sf-how/