Google rolled out Nano Banana 2 image generation directly inside Google Earth on the web today, letting anyone zoom into any location on the planet and prompt the AI to render what it looked like in the past, what a vacant lot could become, or what a building might look like before ground is broken. The feature went live globally this morning at no additional charge, using actual satellite, aerial, and 3D terrain data as a structural anchor for every image it produces — a technical distinction that sets it apart from general-purpose AI image tools that start only from a text prompt, according to Google Earth's official launch post.

Before trying to find it on your phone: the feature is available only on Google Earth's web version at earth.google.com. It is not present in the iOS or Android apps, which together account for more than 500 million downloads on the Play Store alone. And if you already pay for a Google AI or Google One subscription, that quota does not extend here — Google Earth operates on its own separate subscription and generation tier, independent of any standing Google account plan.

What Nano Banana 2 Is and Why It Powers This Feature

Nano Banana 2 is Google's mid-tier AI image model, technically known as Gemini 3.1 Flash Image. It sits between the lightweight Nano Banana 2 Lite — which generates standalone images in roughly four seconds at $0.034 per image via the developer API — and the higher-fidelity Nano Banana Pro, which is reserved for complex professional rendering. Google launched Nano Banana 2 in February 2026, and since then has expanded it into the Gemini app, AI Mode in Search, Google Ads, and NotebookLM. The Google Earth integration is the newest deployment, and the most architecturally distinct.

How the Geospatial Anchor Works: What It Actually Does

When a user in Midjourney or DALL-E types "show me Pompeii in 78 AD," the model generates whatever it has learned to associate with that description. The spatial result may be plausible, but the scene could be geographically inaccurate — wrong scale, wrong terrain, buildings in the wrong relative positions.

Google Earth's implementation is structurally different. The model receives the user's current viewport — the actual rendered satellite and 3D terrain data for that location — as input alongside the text prompt. The AI then generates an image that is constrained to that spatial footprint: the ruins of Pompeii appear in the correct positions, at the correct scale, within the correct topographical context, as confirmed by multiple reviewers who tested the feature at launch. That constraint is what no general-purpose image tool can replicate without access to the same proprietary two-decade archive of satellite, aerial, and photogrammetric imagery that Google Earth holds.

The model also retrieves factual knowledge about named locations. For a prompt requesting an infographic of the Statue of Liberty, Nano Banana 2 draws on Gemini's world-knowledge training to populate the graphic with historical data — construction date, materials, dimensions — rather than inventing them. Whether that retrieval is accurate is a different question, one worth treating with caution (more on that below).

What the Geospatial Anchor Does Not Guarantee

Spatial accuracy is not historical accuracy. When a user prompts the model to render "what Pompeii looked like in 78 AD," the terrain and building footprints are anchored to where things were — but the rooflines, wall colors, street textures, and human figures in the scene are AI imagination. No satellite camera existed 2,000 years ago, and no reliable color photographs of Roman streets exist to train a model toward historical accuracy. What the model produces is spatially plausible and visually coherent, but it is creative speculation presented in a photorealistic style that can read as documentary evidence, a concern flagged by reviewers who tested the feature at launch.

For educational use — which Google Earth product manager Bryan Horowitz specifically highlights in the official launch materials — this distinction matters. A teacher using an AI-generated "hyper-realistic view" of Pompeii in 78 AD to illustrate a history lesson should make explicit to students that the image is a spatial reconstruction, not an archaeological record. The visual authority of a photorealistic rendering, grounded in what looks like real satellite data, creates a credibility effect that goes beyond what the underlying evidence supports.

Google has applied both SynthID watermarking and C2PA Content Credentials to images produced by this feature. SynthID embeds an invisible pixel-level watermark that survives screenshots; C2PA attaches cryptographically signed metadata recording the image's origin and creation tool. Together, they allow any downstream viewer to confirm, using Google Search, Lens, or compatible tools, that an image was AI-generated — an important provenance signal as historically grounded visualizations spread across educational and news contexts. Both systems are part of Google's broader provenance infrastructure for Nano Banana 2.

Five Use Cases Google Has Documented

Google's launch materials describe five specific workflows for the feature:

Historical reconstruction. Zoom to a heritage site and prompt the model to render a period scene. The example from the official launch: transforming the Pompeii ruins into a "bustling, colorful street scene from the Roman Empire." The spatial anchor places the scene correctly; the historical content is AI-generated speculation.

Educational infographics. Generate illustrated explainers tied to specific locations. Google's example involves the Statue of Liberty, where the model combines a visual rendering with factual annotations retrieved from Gemini's knowledge base.

Real estate and urban planning. Architects and clients can prompt a vacant lot to be rendered as a proposed development — "a vibrant shopping and retail district with open spaces" — composited over the actual satellite view of the site.

Pre-construction visualization. Homeowners and developers can see a rendered structure placed on actual terrain before building begins — "a modern lakefront cabin built from sustainably sourced local materials" dropped into the real landscape of the target parcel.

Speculative and creative reimagining. The feature also supports pure imagination, including transforming existing campuses into fictional environments. Google's example involves prompting its Mountain View headquarters to become a "futuristic sci-fi utopia."

Engadget reviewer Lawrence Bonk, who tested the feature at launch, noted that the historical reconstruction example "isn't exactly mind-blowing" and judged the output "not all that hyper-realistic." He also flagged the real estate visualization use case as potentially underwhelming for serious commercial purposes, writing that a credible investor might expect a commissioned architectural render rather than an AI image from a mapping tool.

Generation Speed and Platform Limits

There is a meaningful performance gap between Nano Banana 2 in standalone use and Nano Banana 2 running inside Google Earth. The standalone model, available in the Gemini app and AI Studio, generates images in a few seconds. The Google Earth implementation takes up to two minutes per image.

That gap reflects the compute overhead of the geospatial conditioning pipeline — the model is not just generating an image from a text description, but integrating satellite and 3D terrain data as a structural input, retrieving knowledge about the specific named location, and compositing the result against the viewport. The tradeoff is that the output is spatially credible in a way that a two-second standalone generation cannot be.

After generating an image, users can click "Refine Image" to iterate on it and optionally save it to a Google Earth project. One current workflow limitation: the viewport cannot be adjusted after a prompt is started. To change the view angle or zoom level, the user must discard the prompt, reposition the map, and begin again.

Who Cannot Use This Feature Today

The web-only launch excludes a large share of the people who use Google Earth. The app has more than 500 million downloads on the Play Store, plus a substantial iOS user base. For users who interact with Google Earth primarily on mobile — which, based on Play Store download figures, is a significant portion of the total base — this feature does not exist at launch.

Google has given no timeline for a mobile release. Forbes contributor Paul Monckton described the desktop web limitation as "a major oversight" given the scale of the mobile user base.

For users who already pay for Google AI subscriptions, the feature carries a secondary surprise: their existing generation quota does not apply here. Google Earth operates under its own subscription tiers, independent of any Google One or Google AI plan, which means users who expect their AI subscription to cover Earth image generation will encounter separate limits they did not anticipate.

How This Feature Fits Into the Generative AI Image Market

AI image generators have become a crowded, fast-moving category. OpenAI's GPT-image-2 leads the public Arena image-generation leaderboard; Microsoft's MAI-Image-2.5 is fourth; Nano Banana 2 Lite sits fifth. Adobe Firefly, Midjourney, and ByteDance's Seedance — which has drawn criticism from Hollywood studios including Disney and Paramount over intellectual property concerns — all compete for the same users, as detailed in TechTimes' coverage of Google's July AI tool launches.

Google's strategy with the Earth integration is differentiation through a data asset no competitor holds. Twenty years of satellite and aerial imagery, combined with 3D terrain modeling covering hundreds of cities across more than 40 countries, constitutes a proprietary spatial dataset that cannot be replicated through a text prompt. Tying Nano Banana 2 to that dataset creates a specific use case — place-specific, spatially credible AI visualization — that sits outside the competitive landscape occupied by general-purpose generators.

The feature is available now at earth.google.com. No additional subscription is required, though daily generation limits apply and the feature operates under Google Earth's own separate subscription tiers rather than any standing Google account plan.

Frequently Asked Questions

How do I use AI image generation in Google Earth?

Open Google Earth at earth.google.com in a web browser. Navigate to any location, zoom in to your desired view, and tap the "create image" button. Type a text prompt describing what you want to see — a historical scene, a proposed building, or a creative reimagining — and the model will generate an image anchored to the satellite and terrain data of the location you have selected. Each generation can take up to two minutes. After generating, you can refine the image or save it to a project.

Why isn't AI image generation available in the Google Earth mobile app?

As of the July 30, 2026 launch, the feature is available only on Google Earth's web version at earth.google.com. It is not present in the iOS or Android apps, which together have more than 500 million Play Store downloads. Google has not announced a timeline for mobile availability.

Does my Google AI or Google One subscription cover Google Earth AI image generation?

No. Google Earth operates under its own separate subscription and generation tier, independent of any Google One or Google AI plan. Users who subscribe to Google AI may be surprised to find that their standard quota does not apply inside Google Earth. Daily generation limits apply to the Earth feature under Earth's own tier structure.

How historically accurate are AI-generated reconstructions in Google Earth?

The geospatial anchor makes reconstructions spatially accurate — the terrain, building footprints, and scale of the location are drawn from real satellite and 3D data. But the historical content — what buildings looked like, street textures, colors, and human figures — is entirely AI-generated and not based on documentary evidence. For ancient sites like Pompeii, no camera data from the relevant period exists, so the AI speculates visually within the spatial constraints of the real-world location. Outputs should be treated as creative visualizations, not historical records, and images carry SynthID watermarks and C2PA Content Credentials to identify them as AI-generated.