AI Edit for QGIS: The Complete Guide

Suburban block with infill housing added (after) — Add infill housing to a suburban block, from one prompt. Drag the handle to compare.

Suburban block with infill housing added (before) — Add infill housing to a suburban block, from one prompt. Drag the handle to compare.

Turn the aerial imagery in your QGIS canvas into edits, clean maps, and vector data, all from plain-language prompts. Draw a box, describe the change, and the result drops back onto the map, perfectly aligned. This guide walks through the whole tool: your first edit, the prompt library, Markup, Vectorize, and reference images.

Examples

Drag the slider on any example: the original aerial image is on the left, the AI result on the right.

Agricultural land cover classification (after) — Classify agricultural land cover

Agricultural land cover classification (before) — Classify agricultural land cover

Solar farm added to a field (after) — Add a solar farm

Solar farm added to a field (before) — Add a solar farm

Coast under simulated sea level rise (after) — Simulate sea level rise

Coast under simulated sea level rise (before) — Simulate sea level rise

Redevelopment concept render (after) — Concept render for redevelopment

Redevelopment concept render (before) — Concept render for redevelopment

Install in QGIS

Install AI Edit from the QGIS Plugin Repository: open Plugins -> Manage and Install Plugins, search for AI Edit, and click Install. It runs on QGIS 3 and 4, on Windows, macOS, and Linux, and is powered by the Nano Banana 2 model.

Then open the panel and click Sign in / Sign up. Your browser opens, you create an account (email, Google, or Microsoft), and the plugin connects itself, no credit card, with free edits to start.

Your first edit

Launch and draw your zone

Click the green Launch AI Edit button, then hold the left mouse button and drag a box on the map over the area you want to change.

Click Launch AI Edit, then drag a box on the map to set your zone.

Describe the change

Type a prompt, or open the Library for a ready-made one. Keep the box roughly square and not too large.

A prompt ready to generate

Pick the output detail

Detail	Resolution	Plan
Standard	1K	Free
Detailed	2K	Pro
Maximum	4K	Pro

Generate

Click Generate. A few seconds later the result is added as a new georeferenced raster, exactly over your zone.

The result on the map

The panel keeps every version (Original, V1, V2, and so on). Pick one as the base, prompt again, and generate to refine. The Before / after button slides between input and output.

The result panel with its version strip

The prompt library

Click Library to browse ready-made templates, each with a before/after preview and an editable prompt. They are grouped by theme: Cartography, Urban, Segment, Land cover, Cleanup, Forestry, Agriculture, Climate, Energy, and more.

The prompt library

Markup

Markup lets you draw on your zone to tell the AI where to act. Your marks guide the edit and are removed from the result. Use it for a change in one specific place, not across the whole image.

Open Markup and pick a tool

With a zone drawn, click the pencil button, then choose a tool (pencil, arrow, circle) and a color.

The Markup panel

Draw on the map

Here, a purple outline traced around the greenhouse roofs we want to cover.

Purple markup traced around greenhouse roofs

Prompt and generate

Reference your mark by color, then generate. For example:

Add solar panels on the roofs inside the purple markup. Keep everything outside untouched.

Greenhouse roofs covered with solar panels inside the purple markup (after) — Solar panels added only on the marked roofs, the markup removed from the result. Drag the handle to compare.

Greenhouse roofs covered with solar panels inside the purple markup (before) — Solar panels added only on the marked roofs, the markup removed from the result. Drag the handle to compare.

Vectorize

Vectorize traces a flat-colored AI result (buildings, land-cover classes, parcels) into real vector polygons you can select, measure, style, and export. Use it after a Segment or Land-cover template, when you need GIS features, not just a picture.

Run a segmentation, then click Vectorize

On a result like Segment all buildings (buildings in red), click Vectorize this result.

Vectorize call-to-action on the result

Confirm the color and run

The panel pre-fills the color to extract. Click Vectorize, or use Pick on map to sample a different color.

The Vectorize panel

Refine live

Adjust tolerance, despeckle, simplify outlines, and round corners. Changes re-run instantly on the same layer.

Refine controls

From aerial image to red mask to vector footprints — 174 building footprints extracted from one suburban block.

Each feature carries a clean attribute set (feature_id, class_name, class_color, and a geodesic area_m2) and is saved to a GeoPackage in your output folder, ready for analysis.

The attribute table

Reference images and layers

References give the AI extra context that is cropped to your zone but never shown on the canvas. Use them to control style, match a color legend, or feed real data. The free tier allows one reference; Pro allows several.

A reference image

Attach an image from disk: a style sample, a target look, or a color legend. A common use is a land-use map with your exact legend, paired with a prompt like:

Transform into a land use map in the style of the reference.

The output then matches your palette every time.

A land-use reference attached in the panel

Reference legend, aerial image input, and the land use map result at 4K — Left: the reference legend. Middle: the aerial image input. Right: the result at 4K (Pro), reproducing the legend's residential, commercial, industrial, agriculture, forest, and water colors.

A QGIS layer as reference

This is the part worth getting right. The AI works from two things at once: the aerial image visible on your canvas (the input it edits) and any reference layers you hand it on the side. The catch is that a reference must stay out of the visible view, otherwise it is rendered into the input image instead of being sent as context.

Here, the setup is a Google Satellite basemap on top (the only thing the canvas shows) with an OSM Standard layer kept underneath it. The OSM layer is never drawn into the input; it is dragged onto the prompt as a reference image instead.

Layers panel: Google Satellite on top, OSM Standard underneath — The aerial basemap sits on top and is the only thing visible on the canvas. The OSM layer stays underneath, out of the input, ready to be sent as a reference.

Drag the OSM layer onto the prompt box (or use Ref image). It is cropped to your exact zone and sent as context, perfectly aligned with the aerial input but never shown on the map. Now the model sees the aerial image and knows precisely where every building footprint is.

The OSM layer attached as a reference image, with the segmentation prompt — The OSM layer dropped onto the prompt as a reference. The prompt asks for a flat 2-color building map and tells the model to use the reference to segment more accurately.

Aerial image input, OSM footprints reference, and the building-map result — The aerial image is the input; the OSM layer is sent as context; the result is a clean building map aligned to real data.

The reason this is powerful: every reference is co-located, cropped to the same zone and aligned to the same ground as your input. So you can stack several and the model reads them as one synchronized scene: a DEM or 3D terrain for elevation, several aerial or satellite captures of the same place, contour, water, or land-use layers. Each extra layer is one more aligned source of context, and the more the model has to work with, the more accurate and better-aligned its result.

Free vs Pro

The full prompt library, Markup, Vectorize, and references all work on the free tier. Pro raises the limits: higher resolution, several reference images per edit, and a larger monthly allowance.

Resolution is the one that changes the output itself. A higher setting raises the pixel count on both sides, the input image the model reads and the result it returns, so it resolves detail that simply is not there at 1K and comes back sharper, cleaner, and better aligned.

1KStandard

~1024 px · ~1 MP

Draft and explore

Free

2KDetailed

~2048 px · ~4 MP

Sharper, cleaner edges

Pro

4KMaximum

~4096 px · ~16 MP

Finest detail, best aligned

Pro

Each square stands for a block of pixels. A higher tier packs in more of them on both sides: a larger input image the model reads, and a larger result it generates, so finer detail survives.

The same shadow-removal edit on a Barcelona block generated at 1K, 2K and 4K, zoomed to the same spot, showing the 1K result soft and the 4K crisp — A real run: the same edit (remove shadows) on the same Barcelona block at each resolution, zoomed to the same spot. 1K is soft; 4K resolves every rooftop, car and tree.

See plans and pricing

Tips

Set your basemap and zoom before drawing: the visible pixels are the input.
Keep zones square-ish and modest in size for the sharpest results.
For flat-color maps you plan to Vectorize, ask for solid fills and clean edges.
Keep reference layers hidden.

Need help? Use the ? menu in the panel for the tutorial, the terms, and to report a problem.