Google Earth Adds Nano Banana 2 to Reimagine Real Places

Google Earth on the web now generates AI images from real satellite, aerial and 3D map data using Gemini’s Nano Banana 2 model. The feature rolled out globally on July 30, 2026, so any user can zoom to a place, tap Create image, and describe a change or explainer.

What used to be a pure window on the existing world now doubles as a grounded sketchpad. The images stay tied to actual geography instead of floating free in a blank prompt box.

That single shift changes the feel of every result. A prompt no longer invents a place from nothing. It reworks a place that already exists on the map, with real terrain, real building footprints and real surrounding context already locked in.

The Feature Lands on Web Earth

Product manager Bryan Horowitz introduced the update on the Google blog. Nano Banana 2 pulls Google Earth’s current view as the base, then builds the requested scene on top of that real foundation.

Examples in the launch materials range from a hyper-realistic Pompeii street in 78 A.D. to a Tokyo empty lot turned into a retail district. Users can also request simple historical infographics for landmarks such as the Statue of Liberty.

Those demos show the breadth Google wants users to try first. One path leans educational and factual. Another path leans commercial and speculative. A third path simply plays with the map for fun.

The tool is live for everyone on the web version. Mobile apps do not have it yet. Android Authority and other outlets first flagged the consumer pitch earlier the same day.

Web-first availability means anyone with a browser can test it immediately. It also means the full Create image workflow stays tied to the desktop-style interface where projects, placemarks and the toolbar already live.

How Generation Works

The steps stay short. Open a project or start fresh in Google Earth on the web. Frame the location you care about.

• Click Create image in the toolbar; the current map view becomes the base input.
• Type a prompt that can be practical or wild: more trees and bike lanes, a solar farm on a hillside, or futuristic towers downtown.
• Wait a few moments for Nano Banana 2 to finish.
• Use forward and back controls to compare before and after, then hit Refine image with a new prompt if needed.
• Save the result to the project as a placemark attachment if you want to keep it.

The refine loop matters as much as the first generation. Users can nudge a scene toward denser greenery, clearer labels or a different architectural mood without leaving the same map frame.

The experimental documentation for image creation stresses that generation uses overhead imagery only. Street-level or oblique angles are not the base.

That overhead constraint keeps every output aligned with the classic Earth viewing angle. It also means the model never starts from a ground-level photo a user might have hoped to alter.

Five Grounded Ways People Will Use It

Google’s own post walks through concrete starting points rather than abstract promises. Teachers can turn ruins into living streets. Planners can show clients a finished look. Homeowners can drop a cabin onto their actual lakeshore lot.

The same post lists the five official ways to try the tool with direct Earth links to the demo locations. Because the base is real terrain and buildings, the output feels more plausible than a pure text-to-image result.

Plausibility here comes from continuity. Roads still meet at the same junctions. Shorelines keep their real curves. Neighboring structures remain visible around the altered patch. The generated content has to fit inside a world the user already recognizes.

That fit is useful for early conversation. A planner can show scale and massing against true surroundings. A teacher can keep students oriented to the same map they already studied. A homeowner can judge whether a cabin idea crowds the lot or leaves room to breathe.

Limits Google Built In

The feature carries clear guardrails from day one. It is labeled experimental and pre-GA.

What we know

• Available only on Google Earth web, not the mobile apps at launch.
• Generated images save only inside Earth projects; direct download or external sharing is not supported.
• Every image carries a visible watermark plus invisible SynthID marking.
• Base input is overhead satellite, aerial and 3D data only.

What remains open

• Exact rate limits or quality differences by account type.
• Whether export or API access will arrive later.
• How aggressively safety filters block sensitive or deceptive prompts.

These constraints keep the tool inside Google’s ecosystem for now. That choice reduces casual deepfake distribution while still letting users iterate inside their own maps.

The project-only save path also shapes workflow. Results live next to the placemarks and layers a user already maintains. They do not become loose files that drift into chat threads or social feeds without extra effort.

Who Gains and Who Watches Closely

Urban planners and real-estate teams gain a fast visual aid that starts from true site conditions. A barren lot becomes a proposed plaza in minutes instead of days of custom rendering. Educators gain a way to make distant history feel immediate without leaving the map.

Architects on X already noted the value for early concept boards. Casual users simply get a more playful Earth: holiday lights on a square, extra greenery on a street, or pure sci-fi fun.

At the same time, verification specialists flagged a second-order problem. Realistic synthetic overhead views can complicate open-source intelligence work that once trusted Google Earth imagery as a relatively stable record. One widely shared demonstration planted a nuclear plant on an Iranian site and refugee camps near a border with a single sentence each. The images look plausible at a glance.

That tension sits beside the creative upside. The same grounding that makes mock-ups useful also makes fabricated scenes harder to dismiss instantly.

Speed is the shared thread across both camps. Concept boards appear faster. So do deceptive overlays. The difference lies in intent and in whether the viewer knows the image began as a generated layer rather than a captured one.

The Trust Layer That Comes With Every Image

Google pairs the feature with its existing provenance stack. Nano Banana 2 images embed SynthID, the invisible watermark the company has pushed across its generative tools. Visible markers also appear so a casual viewer can tell the scene is generated.

The company has expanded verification tools elsewhere, including checks inside the Gemini app. That work arrives as rising deepfake fraud pressures on identity tools push platforms toward stronger signals of authenticity.

Still, no watermark is perfect once an image leaves its original context. Because Earth currently blocks easy external export, the immediate circulation risk stays lower than a free-floating generator. The longer-term question is what happens if or when those images travel more freely.

Today we’re bringing AI image generation with Nano Banana to Google Earth, letting you virtually reimagine anywhere in the real world.

That line from the official @googleearth account, which drew more than 1,800 likes and 180,000 views within hours, captures the marketing pitch. Replies and quote posts mixed delight with the “what could go wrong” jokes that now travel with every powerful image model.

Inside the product the dual markers do the quiet work. Visible labels help ordinary viewers. SynthID aims to survive common edits so downstream tools can still flag the content as generated.

What the Model Itself Brings

Nano Banana 2 is Google’s Gemini 3.1 Flash Image model. It arrived earlier in 2026 as the fast sibling that keeps much of the reasoning and world knowledge from the Pro tier. The Nano Banana 2 model capabilities and rollout post highlights better subject consistency, legible text, real-time search grounding, and production resolutions up to 4K.

In Earth the model’s world knowledge helps when a user asks for historical facts or plausible architectural styles. It is not a CAD package or a historical archive. It produces a concept image, not a certified reconstruction or construction document.

Users who already know Gemini’s image tools will find the Earth version familiar, only now the camera starts locked to a real place on the planet.

Subject consistency and legible text matter for the infographic style prompts. Real-time search grounding helps when a landmark request needs current or commonly known details. The 4K ceiling gives the output enough resolution to inspect on a large monitor while still living inside a map project.

Grounding Changes How Prompts Behave

The geographic anchor is the main difference from a blank Gemini canvas. Everything else about Nano Banana 2 stays shared across products, yet the starting frame forces the model to respect real layout.

Compare the two starting points side by side:

Because the terrain and nearby structures remain visible, wild prompts still have to negotiate with reality. Futuristic towers must rise from the same downtown footprint the user framed. A solar farm must sit on the same hillside the map already shows.

That negotiation is what makes the output feel more plausible. It is also what makes a deceptive prompt harder to spot at a glance when someone later sees only the finished frame.

Project Walls Slow Outside Circulation

The decision to block direct download and external sharing is more than a missing button. It keeps generated scenes inside the same workspace where the original map view lives.

Several practical effects follow from that wall:

• Images stay next to the placemark and project that produced them.
• Casual copy-out into social feeds requires extra steps the product does not encourage.
• Watermarks remain attached while the file stays in its native home.
• Collaboration happens through Earth projects rather than loose image files.

Verification specialists still worry about the day export arrives or about screenshots that strip context. For now the circulation path is longer and more deliberate than a free-floating generator offers.

Creative users feel the same wall as friction. A polished concept cannot yet drop straight into a slide deck or client email without leaving the Earth environment. The tradeoff is intentional: lower casual deepfake spread in exchange for a tighter creative loop on the map.

Frequently Asked Questions

What AI model powers the new Google Earth image feature?

Nano Banana 2, also called Gemini 3.1 Flash Image, supplies the generation. It combines Flash-speed iteration with advanced world knowledge and improved instruction following that earlier Nano Banana versions lacked.

Is the feature available on Google Earth mobile apps?

No. At launch it runs only inside Google Earth on the web. Mobile apps have not received the Create image button or the underlying pipeline.

Can I download or share the AI images outside Google Earth?

Direct download and external sharing are not supported. Generated images can be saved only as attachments inside a Google Earth project, and they carry visible plus invisible watermarks.

Does every generated image get a watermark?

Yes. Google applies both a visible watermark and its invisible SynthID digital watermark so the content can be identified as AI-generated even after common edits.

How is this different from generating images in the Gemini app?

In Earth the starting canvas is the live satellite, aerial or 3D view of a real location rather than a blank or uploaded photo. That geographic anchor is the main difference; the underlying Nano Banana 2 model is shared across products.

Anyone curious can open Google Earth on the web, pick a familiar corner of the map, and test a prompt. The results stay inside the project for now, watermarked and tethered to the real place that inspired them.