Introducing how designer Justin Chavez combined Google's 3D map data and open-source AI to bring 1,300 miles of the San Francisco Bay Area to life as an intricate pixel art game map.
Imagine waking up in the morning, turning on your computer screen, and seeing the massive metropolis of San Francisco—which we see every day in news and movies—unfold like a charming pixel art screen from a classic 1990s role-playing game (RPG). Forests of buildings have transformed into intricate dot graphics, while winding hills, blue seas, and red salt pans are elaborately recreated as if in a fairytale village.
To put it in perspective, it delivers the visual shock of transporting an entire real-world city into a retro-styled toy block world. The ‘Isometric SF’ project recently unveiled by developer and designer Justin Chavez is a remarkable work that brings this exact imagination to reality [Source 1]. Going beyond mere pretty artwork, this project represents a prime example of how cutting-edge 3D map data and artificial intelligence (AI) technology can meet to create new artistic experiences. Let’s explore how this fascinating digital city was born and what technical secrets are hidden within it.
Why This Matters
There have been countless attempts to transfer the real world we live in into digital space, such as Naver Maps and Google Maps. However, their purpose mostly remained confined to practicality like “finding accurate directions” or “checking traffic info.” In contrast, Isometric SF completely breaks away from the mold of practicality, focusing instead on “artistic sensibility” and “the joy of exploration.”
As the entire vast San Francisco Bay Area spanning 1,300 square miles (over 3.36 million square kilometers, more than 5 times the area of Seoul) transforms into a classic game map [Source 1], we get to view a familiar city from an entirely new perspective. It is immensely fun to explore corner by corner how famous landmarks like Apple Park (Apple’s massive circular headquarters), the Palace of Fine Arts (a beautiful San Francisco structure reminiscent of ancient Roman architecture), and UC Berkeley (a world-renowned public university) are expressed in dot graphics [Source 1].
This technology offers profound inspiration for various fields such as the metaverse (spaces where virtual reality and reality interact), indie game development (independent games produced on a small scale without major funding), and virtual tourism. It wide opens the possibility of accomplishing large-scale handmade pixel art work—which in the past would have taken hundreds of graphic designers years of painstaking labor—with far less cost and time by combining AI and raw geographic data.
Easy to Understand: How 3D Maps Meet AI
How was this massive project possible? To easily metaphorize the core principle, it’s like equipping a “magic filter that applies the art style of a retro game master onto a physical 3D model captured with a cutting-edge camera.”
Simply put, this process consists largely of two stages.
Stage 1: Preparing Raw Materials (Google 3D Tiles)
Justin Chavez didn’t draw from scratch. He utilized immense data called Google Photorealistic 3D Tiles (data that precisely scans major cities worldwide down to the centimeter to construct a 3D virtual globe) [Source 1]. This data is the skeleton of a virtual globe that three-dimensionally scans major cities around the world with centimeter-level precision, preserving actual building heights, terrain curvatures, and road locations. San Francisco’s signature steep hills, rows of Victorian-style residential areas, and winding road networks are all captured here [Source 2, Source 4].
Stage 2: Teaching AI an Art Style (LoRA and Open-Source AI)
Simply bringing the skeleton doesn’t immediately turn it into pixel art. Artificial intelligence was mobilized to transform cold 3D shapes into warm dot graphics resembling 1990s arcade games.
Here, LoRA (Low-Rank Adaptation - a lightweight additional learning module that efficiently trains only specific art styles or traits without altering the massive AI model) technology plays a core role. During the initial prototyping phase, the developer used a tool called ‘Nano Banana 2’ to intensively teach pixel art styles to a Qwen-based image editing AI model [Source 1].
You can visualize it like this: showing 100 works by a famous pixel art master to an art major college student (the basic AI model) and giving them special tutoring (LoRA training) by saying, “When drawing city buildings, follow this art style.” Through this tutoring, the AI perfectly masters which color dots (pixels) to place on building windows, street trees, and ocean waves to bring out a retro feel.
Current Status: The Power of Open Source Breaking Through the Cost Wall
One of the most surprising aspects of this project is the issue of “cost” and “efficiency.”
If the high-performance premium model (Nano Banana 2) initially used to flesh out the idea had been pushed forward as-is, completing this entire map (1,300 square miles) would have cost an astronomical roughly $7,000 (about 9 to 10 million KRW) [Source 1]. That is a considerable amount for a personal project started out of interest.
However, the developer didn’t stop there. He newly trained ‘Klein 4B LoRA’ based on Qwen, a smaller, lighter open-source (software with publicly accessible source code allowing anyone to freely modify and use it) model [Source 1]. Without relying on big tech’s expensive paid servers or massive models, he customized a small, efficient open-source AI model to fit his needs.
The result was a resounding success. While almost perfectly replicating the original art style and precision, production costs were dramatically lowered to one-fourth of the original [Source 1]. This proves that an era has wide-open where individuals can combine AI and open-source tools to freely materialize their imagination even without massive capital.
Looking at the completed map, its meticulous care commands admiration. The salt pan areas in the South Bay shine in a vivid harmony of red, yellow, and white, beautifully surrounded by green wetlands and the teal waters of the Bay [Source 3]. It boasts the visual density of someone assembling all of San Francisco block by block with hundreds of thousands of Lego bricks.
What Lies Ahead?
The success of projects like ‘Isometric SF’ holds the potential to completely change how we approach digital maps in the future.
First, the popularization of personalized virtual worlds. Like the ‘Isometric NYC’ project covering New York [Source 1, Source 4] and this San Francisco project, countless cities around the world can be reborn in their own unique game graphic art styles. In the future, services where users can convert their neighborhood into a ‘Zelda’-style or ‘Pokemon’-style pixel art map with a single click to explore will likely become common.
Second, a paradigm shift in indie game development. By utilizing open-source AI and lightweight LoRA models, game developers can instantly prototype visual assets based on real-world geographic data without having to manually dot vast open-world backgrounds piece by piece.
Of course, limitations exist. AI-generated pixel art looks flawless at a glance, but microscopic details may exhibit illogical building structures, road connections, or awkward noise. Therefore, rather than jumping straight to fully automated commercial services, careful prompt engineering (the technique of inputting optimal instructions to AI) and style adjustments by artists or developers must accompany the process, just as in this case.
In the not-too-distant future, we may live in an era where we can ask AI to turn our desired travel destinations or hometown memories into retro game maps to keep as keepsakes, and virtually stroll through them.
MindTickleBytes’ AI Reporter Perspective
When massive tech giants are bogged down in a multi-hundred-billion-dollar competition for Large Language Models (LLMs, large-scale AIs capable of human-like conversation and reasoning), true creativity shines through individual experiments utilizing small, agile open-source AI. ‘Isometric SF’ proved that AI is not a threatening entity stealing human jobs, but rather the friendliest magic brush that can visualize anyone’s cherished childhood imagination into real-world data.
References
- Google Photorealistic 3D Tiles
- OpenStreetMap
- Apple 3D Map Data
- Nano Banana 2 LoRA
- Klein 4B LoRA
- GPT-5 Vision Model
- 130 square miles
- 1,300 square miles
- 13,000 square miles