Google Earth AI Alternative: Generate on Your Own Imagery in QGIS

For one day, from 30 to 31 July 2026, Google Earth on the web could generate imagery. You pressed a button marked create image, typed a sentence into the box under the map, and Nano Banana drew a photorealistic version of that exact spot. Then Google took it away. My replacement is AI Edit, ours: it runs the same model inside , on the aerial photography you already have open, and gives the answer back as a map layer that knows where it belongs, not a screenshot.

What the button did

Image generation in Google Earth was a create image button in Earth on the web, live from 30 to 31 July 2026, that turned a typed sentence into a photorealistic view of the place on screen, drawn by Nano Banana from Google's own satellite, aerial and 3D imagery of that spot. The result stayed inside Earth as a picture, never as a file a map could place.

Because the model was handed Google's imagery of the place, the result looked like that place rather than a generic city.

Google Earth in a browser, showing an empty waterfront lot in Tokyo, with a create image label at the top left and a text box under the map holding a typed sentence about turning the lot into a shopping and retail district. — From Google's own launch video: a create image label top left, one text box under the map. That box was the entire input. Image: Google.

What came back had arrows to page between attempts, a Refine image button, and a Save to project button that kept it inside Earth. There was no way to get it out as a file a map could place.

The same Tokyo lot rendered as a lit waterfront district at dusk, with buttons reading Refine image and Saved to project. — The result on the same view. It was live globally for about twenty-four hours. Image: Google.

Google rolled it back on 31 July, after people began sharing generated scenes of real places as if they were real.

The two side by side

	AI Edit (ours)	Google Earth, create image
Status	Live	Removed on 31 July 2026
Where it runs	A plugin	A button in Earth on the web
What it draws on	The imagery in your own project	Google's imagery of that spot
What comes back	A map layer, placed on the ground	A picture inside Earth
Point of view	Straight down, 2D. 3D in progress	Any angle, photorealistic 3D
Setup	Plugin manager, one install	None

Which you want depends on one thing. If the picture has to go back on a map next to your own data, use AI Edit. If it is going on a slide and nothing after it cares where the place is, the Gemini app is quicker and cheaper. If you need a tilted, street-level view, neither one replaces Earth: AI Edit looks straight down, in 2D, and a 3D view is something we are building rather than something you can use today.

How I checked this

One of these two I could test and the other I could not, so it is worth saying which is which before the cards below.

I ran AI Edit myself, over a 339 m by 339 m square in Nantes, on public aerial photography from IGN. Earth's button no longer exists, so everything I say about it comes from Google's launch post, rollback statement and plan pages, all read on 31 August 2026.

Disclosure: I co-founded TerraLab, so AI Edit is ours. It gets the same card, the same lines and the same stated limits as anything else I would review.

1. AI Edit (ours)

A plugin that redraws the aerial photography on your screen from a written instruction. Outline an area, describe the change, and the answer comes back as a map layer sitting exactly on the photograph it was built from, so everything else in your project still lines up with it.

An aerial view redrawn as a flat site plan, with white building footprints outlined in black, mid grey roads and parking, and pale sage green vegetation. — The site plan template over the test square. The white footprints land on the real footprints, so you can trace or measure this against the survey underneath.

Best for: a town planner, landscape architect or flood engineer showing a jury what a real site could become.
Not for: counting or measuring.
Runs on: our servers. Your imagery leaves your machine on every run.
Install: the plugin manager, version 3.22 and up. No graphics card needed.
Price: a free tier with an email and no card, then a subscription.

Where it wins

The result lands on the exact rectangle you outlined, so two runs side by side meet cleanly and a long corridor can be done in pieces.
It works on your imagery, not Google's: your drone flight, your 2019 survey, your national mapping agency's feed.
It reads the site. In the park example below the car parks go, while the exhibition hall, both roundabouts and every mature tree stay.

Where it falls short

It works in 2D, and Earth worked in 3D. This is the real gap. If what you have to hand a client is a view from the pavement, this tool does not produce it today. We are working on a 3D output. It is not shipped, and I am not going to pretend otherwise.
Coverage is whatever you brought. Earth had the whole planet indexed.
Expect two or three attempts on an area before one is worth showing.

2. Google Earth, create image

The button described above. Global on day one, no waitlist, no enterprise tier, gone in twenty-four hours.

Best for: nothing today. Google paused it on 31 July 2026.
Not for: anything that had to end in a file you would open elsewhere.
Runs on: Google's servers.
Install: none, it was a browser button.
Price: never sold on its own. All three Earth plans lost it.

Where it won

Every place on the planet was one search away, with photorealistic 3D on top. No plugin gives you that.

Where it fell short

What you saved was a picture and nothing more, with no record of where on the ground it belonged.
It is gone, with no published return date.

What you can give it besides a sentence

That is the comparison. The rest of this page is about the one of the two that still works, and about what decides your results more than the wording of your prompt: what else you put in. Earth took a sentence and nothing else. Inside , two other kinds of input reach the model.

Any layer already open in your project

You can hand the model one of your own map layers as guidance: an elevation model, a cadastral plan, a flood map, a land-use map with its legend. It is cropped to the area you outlined and never drawn on your canvas. A prompt box in a browser cannot do this, having no project and no layers to reach into.

Below is an elevation model of Vaison-la-Romaine, a map where the colour of each pixel is the height of the ground at that point.

A digital terrain model of Vaison-la-Romaine, showing black and purple low ground along the river and orange and yellow high ground. — The reference: black and purple are the low ground along the river, orange and yellow the high ground.

The flood result with no reference, then the same result with the elevation model attached (after) — Same flood prompt, same town. Nothing attached on the left, the elevation model attached on the right, where the water fills only the low ground and the hilltop old town stays dry. Drag the handle to compare.

The flood result with no reference, then the same result with the elevation model attached (before) — Same flood prompt, same town. Nothing attached on the left, the elevation model attached on the right, where the water fills only the low ground and the hilltop old town stays dry. Drag the handle to compare.

Marks you draw on the map itself

The second input is simpler: stop describing a spot in words and point at it. Draw on the map with a pencil, a line, an arrow or a circle, in a colour you then name in the prompt. The marks travel with the request and are taken back out of the result.

A bare field with a pink outline drawn on it, then the same field with trees planted inside the outline (after) — Trees planted only inside the pink outline, and the outline itself gone from the result. Drag the handle to compare.

A bare field with a pink outline drawn on it, then the same field with trees planted inside the outline (before) — Trees planted only inside the pink outline, and the outline itself gone from the result. Drag the handle to compare.

Examples

Four runs on the same 339 m square over Nantes. One photograph, four different questions.

The same aerial view redrawn as a flat site plan with white footprints, grey roads and pale green vegetation (after) — A site plan in four flat classes

The same aerial view redrawn as a flat site plan with white footprints, grey roads and pale green vegetation (before) — A site plan in four flat classes

The same aerial view in winter with snow on roofs, lawns and canopies and wet cleared roads (after) — The same place under fresh snow, carriageways cleared

The same aerial view in winter with snow on roofs, lawns and canopies and wet cleared roads (before) — The same place under fresh snow, carriageways cleared

The same aerial view with apartment blocks and courtyards built on the former car parks (after) — Residential infill on the car parks only

The same aerial view with apartment blocks and courtyards built on the former car parks (before) — Residential infill on the car parks only

The same aerial view with the car parks replaced by lawn, trees, paths, a pond and a play area (after) — The car parks rebuilt as a neighbourhood park

The same aerial view with the car parks replaced by lawn, trees, paths, a pond and a play area (before) — The car parks rebuilt as a neighbourhood park

Two habits decide whether your own first result is usable. Name what must not move: "keep every building, road, roundabout and existing tree exactly where they are". And say the classes, not the mood: the site plan above came from "buildings as white footprints with a thin black outline, roads and parking as mid grey, vegetation as pale sage green".

Questions people ask

Why did Google remove AI image generation from Google Earth?

Google rolled it back on 31 July 2026, one day after launch: "We've seen geospatial professionals using this feature for a range of useful purposes, however we've also seen people sharing screenshots of generated imagery that appear to violate our policies. So we're rolling back this feature in Google Earth while we work on implementing stronger guardrails." The images carried a SynthID watermark and never appeared in Earth for anyone else, but a screenshot walked them past both controls.

Is it coming back?

Google called it a pause and has published no date. The problem it has to solve is the one the screenshots showed: whatever Earth generates has to stay identifiable once it leaves Earth.

Does the generated image line up with my map?

In , yes. AI Edit writes a GeoTIFF, an image file that carries its own coordinates, covering the exact rectangle you outlined. A picture made anywhere else has no coordinates at all.

Can I put a generated aerial image in a planning application?

Treat it as an illustration of intent, labelled as generated, never as a survey or a record of what exists. Our output is marked as AI-generated for that reason.

Sources

Google, Transform any place with Nano Banana in Google Earth, 30 July 2026. The two Earth screenshots are frames from the demo video in that post.
Google's rollback statement, 31 July 2026, quoted from the reproductions in NDTV Profit and unite.ai, the post itself being unreachable.
Google, Learn about Google Earth plans, read 31 August 2026.
Nantes figures generated in on 31 August 2026 over IGN BD ORTHO, Licence Ouverte. Flood and mark-up figures from the AI Edit guide.