Satellite truth meets generative “what if” — inside the map, not a separate art app.
On July 30, 2026, Google Earth launched AI image generation with Nano Banana on Google Earth web: pick any place using satellite, aerial, or 3D views, tap create image, and prompt a custom scene. Product manager Bryan Horowitz’s official post pitches history class, real-estate pitches, backyard dreams, and sci-fi campus makeovers — all grounded in Earth’s real imagery rather than floating in a blank latent space.
Same day as Gemini Robotics 2, Google is shipping physical-world AI in two directions: robots that act, and maps that reimagine.
TL;DR
| Question | Answer |
|---|---|
| What? | Create image in Google Earth web via Nano Banana 2 |
| When? | July 30, 2026 — available globally (web) |
| Input | Real Earth satellite / aerial / 3D context + text prompt |
| How | Zoom → create image → type what you want to see |
| Why it matters | Generations stay place-anchored |
| Hero uses | History, infographics, real estate, personal builds, fun makeovers |
| Try | earth.google.com |
| Docs | Transform any place with Nano Banana |
What people are asking
“Is this just Imagen with a map wallpaper?”
No. The differentiator Google sells is conditioning on Earth’s geospatial imagery — the empty Tokyo lot, Pompeii ruins, or lakeside slope you are looking at. Nano Banana 2 still does the generative lift (Nano Banana 2 family), but the scene is supposed to respect the real footprint you zoomed to. That is closer to “reskin this parcel” than “draw a city from vibes.”
“Web only?”
Launch announcement targets Google Earth on web. Do not assume iOS/Android parity on day one — check the in-app Earth clients before teaching a classroom on phones.
“Can I use this for real permits / legal filings?”
No. These are concept visuals. Zoning boards and lenders still need stamped drawings. Use Earth generations for client storytelling and teaching, then hand off to CAD / survey data. Same ethics as any generative real-estate marketing: disclose AI.
“Will people confuse AI scenes with real places?”
Yes — that is the risk. Google’s image stack widely uses SynthID invisible watermarks; always treat outputs as synthetic, keep prompts/records, and prefer visible “AI-generated” labels when publishing. See also LinkedIn Content Credentials / C2PA.
Five official try-it paths
Google’s blog lists five patterns — steal these prompts as templates.
1. Bring history to life
Example: Pompeii ruins → “Render a hyper-realistic view of what these ruins looked like in 78 A.D.”
Classroom value: students see space, not only a textbook plate. Accuracy will still hallucinate; teachers should compare against archaeological sources.
2. Learn while you explore
Example: Statue of Liberty → “Create an easy to understand infographic … with key historical facts.”
Behind the scenes: Gemini retrieves facts, Nano Banana layouts the graphic. Fact-check every claim — retrieval + generation can still invent.
3. Professional real estate / urban concepts
Example: empty Tokyo lot → vibrant shopping district with open space.
Pitch decks get a site-specific mood board in minutes. Still not a construction document.
4. Visualize before breaking ground
Example: lakefront lot → modern cabin from sustainable materials.
Homeowners and designers get a photorealistic nest in the actual landscape.
5. Playful makeovers
Example: Google Mountain View campus → sci-fi utopia with biodomes and flying pods.
Marketing and entertainment; the feature that will flood social feeds.
How to run a clean Earth generation session
1. Open Google Earth web (signed in if your org requires it)
2. Navigate/search to the exact parcel or landmark
3. Set a useful camera (tilt/altitude matter for 3D context)
4. Tap Create image
5. Prompt with: place intent + style + constraints (“keep street grid”, “no new roads”)
6. Iterate with narrower follow-ups
7. Export/share with AI disclosure
8. Never present as documentary photography
Prompt patterns that work better
| Goal | Prompt skeleton |
|---|---|
| History | Hyper-realistic reconstruction of [site] in [year], keep terrain footprint, daylight |
| Real estate | Reimagine this lot as [program], respect parcel boundary, contemporary materials, golden hour |
| Education | Simple labeled infographic of [landmark], 5 key facts, clean icons, no fake citations |
| Personal | Add a [building type] using local materials, match surrounding tree line and slope |
| Speculative | Futuristic redesign, clearly fantastical, neon accents — for concept art only |
Why geospatial grounding matters for AI products
Most image models start from noise + text. Earth starts from a CRS-backed place humans already trust. That changes product design:
| Free-floating image gen | Earth + Nano Banana |
|---|---|
| “A park in Tokyo” | This parcel’s geometry + surroundings |
| Easy to invent roads | Harder to ignore existing blocks (still not GIS-accurate) |
| Great for art | Better for site conversations |
| Weak for urban planning meetings | Stronger mood for those meetings |
For builders, the lesson maps to any domain with a trusted spatial base layer — maps, BIM, medical imaging, factory digital twins. Generative overlays beat generative voids when stakeholders share a coordinate system. Adjacent explainx.ai threads: Pascal Editor / 3D web buildings, world models, AlphaEarth lineage in DeepMind’s planetary mapping work.
Classroom and client workflows that do not embarrass you
Teachers
- Open the real site in Earth first (ruins as they are).
- Generate the historical reconstruction.
- Split-screen compare and ask: What did the model invent?
- Assign a short source check against a museum or paper.
Architects / brokers
- Capture the true parcel with labels (roads, neighbors).
- Generate 2–3 program options with explicit constraints.
- Export concepts into the deck with “AI concept — not a survey” on every slide.
- Move winners into CAD; never reverse the order.
Product builders copying the pattern
If you own a digital twin, BIM viewer, or factory layout tool, Earth is a template: trusted geometry in → generative overlay out → human gate. Pair with provenance standards (C2PA / Content Credentials) so downstream social platforms can show AI labels when users upload.
Same-day Google context
July 30 also brought Gemini Robotics 2 — whole-body VLAs and embodied agents. Earth Nano Banana is the perception/imagination surface; Robotics is the action surface. Both sell “AI that understands the physical world,” one through pixels on a globe, one through torque on Apollo and Spot.
Consumer Gemini users already saw Nano Banana / Omni media features (free video promo, NotebookLM Shorts). Earth is the geo-native distribution channel for the same image stack. Expect prompt tourism (“turn my house into…”) to dominate week one; the durable use is site-specific storytelling for people who already argued about a plot of land.
Builder / educator checklist
□ Web Earth account works in your country/network
□ Lesson plan includes “this is AI reconstruction” slide
□ Real-estate decks label AI concept art vs survey
□ Keep original Earth URL + prompt for provenance
□ Fact-check any Gemini-sourced “historical facts” panel
□ Don’t upscale and strip watermarks for stock misuse
□ Compare 2–3 camera angles before locking a pitch image
□ For serious planning, export ideas → CAD/GIS, don’t stop in Earth
Honest limitations
- Not survey-grade — expect melted cars, invented façades, wrong eras.
- Web-first launch; mobile parity unknown at announce time.
- Infographics can hallucinate facts even when Gemini “retrieves.”
- Real-estate misuse risk (misleading buyers) is on the human publisher.
- Cultural/heritage sites need sensitivity — speculative reconstructions can erase living communities’ narratives.
- Rate limits / quotas not detailed in the blog — expect consumer fair-use caps.
- SynthID / disclosure UX may vary; verify in-product.
- 3D tilt and altitude change what the model “sees”; regenerate from two camera angles before you trust a client-facing still.
- Do not upload Earth generations to stock libraries as documentary photos of a named place.
- Enterprise Workspace policies may disable generative features — check admin settings if Create image is missing.
Closing
Nano Banana in Google Earth turns the planet into a promptable mood board anchored to real coordinates. Use it to teach, pitch, and play — then keep the line bright between imagined and measured. For the robotics half of Google’s physical-AI day, read Gemini Robotics 2. If a generation looks too perfect for the messy street you know, assume hallucination until a survey says otherwise. Save the Earth URL with your prompt so reviewers can reopen the same camera view later.
Follow @explainx_ai for geospatial AI follow-ups.
Related on explainx.ai
- Gemini Robotics 2 — whole-body physical AI (same day)
- Gemini free 10 Omni videos promo
- NotebookLM Short Video — Nano Banana 2 Lite
- LinkedIn Content Credentials / C2PA for AI images
- Gemini Omni Flash video
- Google Photos Video Remix
- Pascal Editor — open 3D buildings
- What are world models?
- NVIDIA Cosmos 3 physical AI
Sources
- Google — Transform any place with Nano Banana in Google Earth
- Google Earth on X
- Google Earth web
- Nano Banana 2
- Short link from Earth post: goo.gle/44SUsOP
Feature availability and behavior as announced July 30, 2026 for Google Earth web. Quotas, mobile support, and watermark UX can change — verify in-product before classroom or client use.