Google Earth Pulls Back Nano Banana 2 AI Image Feature After Unexpected Abuse

Google recently integrated the Nano Banana 2 image‑generation model into the web version of Google Earth, allowing users to create AI‑generated visuals directly on the map by clicking a new “Create image” button.

Feature showcase

The model can produce hyper‑realistic scenes such as reconstructing ancient Pompeii, generating custom infographics of landmarks like the Statue of Liberty, visualizing professional real‑estate proposals, and even imagining futuristic cityscapes or post‑apocalyptic versions of real locations. Users simply describe the desired scene, and the system renders a photo‑level image within seconds.

Technical foundation – Geospatial Grounding

Unlike generic text‑to‑image tools, Nano Banana 2 receives a composite input consisting of the current satellite or aerial view, 3D elevation and terrain mesh, and the camera’s spatial parameters. This “geospatial grounding” enforces three strict constraints: the terrain cannot be altered, perspective and spatial relationships must match the real world, and subject consistency is maintained for up to five subjects and fourteen objects, preventing visual collapse when the view changes. Conditioned generation under these constraints typically requires about two minutes per image.

World knowledge retrieval

Powered by Gemini’s multimodal and search‑grounding capabilities, the model queries Google’s extensive knowledge base to supplement images with factual details such as construction year, materials, and historical context. While this enriches the output, the article notes that the retrieved information may not always be accurate.

High‑fidelity output

Nano Banana 2 supports 2K and even 4K resolution renders and can embed clear, legible text directly into the scene, enabling the creation of detailed infographics without post‑processing.

User workflow

The interface offers a before/after view, a “Refine image” option for iterative prompting, and a “Save to Project” feature that pins the result as a placemark, preserving camera angle and allowing others to explore the generated view.

Limitations and withdrawal

The feature is unavailable in Street View, limiting its utility for developers and architects who need precise spatial geometry. Media outlets described the uncurated results as “AI slop,” and after users quickly abused the tool to produce extreme visualizations (e.g., zombie‑infested Philadelphia), Google added “enhanced safeguards” and removed the feature within a day.

Industry context

In the crowded AI image‑generation market—where OpenAI’s GPT‑image‑2 leads the Arena leaderboard, with Adobe Firefly and Midjourney also competing—Google’s differentiator is its proprietary geospatial dataset accumulated over two decades, covering satellite imagery, aerial photos, and 3D city models for dozens of countries. This creates a new category of spatially grounded AI visualization that emphasizes geographic fidelity over artistic style.

Overall, the Nano Banana 2 experiment demonstrates both the potential of tightly integrating AI with real‑world geodata and the challenges of ensuring reliability, usability, and responsible deployment.