plus AIWhat AI can do inside QGIS, and what it cannot
One plugin turns your imagery into vectors. The other turns it into different imagery. Both run on any laptop with no GPU, and clicking one object at a time is free and unlimited.
QGIS 3.0 or later. Windows, macOS, Linux. No GPU, no Python setup, no card.
Every object, as a polygon
Click an object and get its outline. Or name what you want ("buildings", "solar panels", "trees") and get every match across a zone, as an editable layer in your project CRS.
A new raster, same coordinates
Select an area, describe the change in plain words, and get imagery back that no longer shows the clouds, the cars or the winter. It lands aligned to the same extent and CRS.
Why these exist
The slow parts of a GIS job
Neither plugin invents new work. Both take something you already do by hand and make it quick.
Someone has to draw the map
Every building, every field, every tree starts as a person clicking corner after corner. On a real site that runs to days, and the layer is stale before it ships.
The imagery is never the one you need
Cloud sits over the site. The flight was in January and the report is about summer. The car park is full and you wanted the surface. The tile is 50 cm and the print is A1.
And the fix used to need a GPU
Models this good want an NVIDIA card, CUDA and a Python environment. GIS work happens on a work laptop. So the model runs on our servers, and QGIS just asks it for the answer.
Which one
Find your job, get your plugin
The two do different work. Find the task you came here for.
- Get building footprints as polygons
- Extract a road network from aerial imagery
- Map individual tree crowns
- Count solar panels across a district
- Digitize water bodies or field boundaries
- Trace one awkward object without clicking every corner
Need both? The Pack covers the two plugins on one subscription. See pricing
Real runs
What comes back
Output from real runs in QGIS, one block per plugin.
Detection results as they land in the project. Every object gets its own colour, which is how the layer looks before you export it.
City blocks
Central Paris. One pass returned 635 polygons.
Tree crowns
Parkland around a lake. One pass returned 925 polygons.
Parked cars
Street parking across a district. One pass returned 575 polygons.
Imagery edits, one prompt each. Drag the handle to compare. Every result comes back on the extent and CRS of the tile that went in.
Raise the water
The quays and the square go under. Only the tree crowns stay above the line.
Panels on every roof
A residential street fitted out, to show a scheme before anyone builds it.
Sharpen for print
A soft tile resolved for A1. The model invents the fine detail, so read it as a picture, not as evidence.
Getting started
Running in about a minute
Install from the plugin manager
In QGIS, open Plugins, then Manage and Install Plugins. Search the plugin name and click Install. No admin rights needed.
Sign in once
Open the panel and sign in with your email. No card. The free tier starts immediately and your key works on every machine you use.
Load a raster and run
Any raster works: a GeoTIFF on disk, an XYZ basemap, a WMS layer. What is on screen is what the plugin reads.
Being straight with you
We are not your only option
Six plugins now run Segment Anything inside QGIS and several are free. Geo-SAM was rebuilt in June 2026 with one-click dependency install, though you prompt it with points and boxes rather than words. GeoOSAM does free text-prompt detection across the visible canvas. SamGIS runs SAM 2.1 fully offline, no account and no CUDA. If one of those fits your work, use it.
The catch on the newest model is hardware. SAM 3 wants an NVIDIA GPU with CUDA and has no CPU path, so Macs, AMD cards and integrated graphics cannot run it locally at all. That is the gap we fill: the model runs on our servers, so any laptop can send a tile and get polygons back.
Bunting Labs and Mapflow also do commercial feature extraction in QGIS. And if you have the hardware and don't mind the setup, the free local route costs nothing. We would rather you knew that than found out later.
Questions people ask before installing
- Yes, quite a few. TerraLab ships two: AI Segmentation, which detects objects in imagery and returns them as vector polygons, and AI Edit, which changes the imagery itself. Both install from the standard QGIS plugin manager. Geo-SAM, GeoOSAM, SamGIS, Bunting Labs and Mapflow are other options, and three of those are free.
- There is a free tier with no card. Manual segmentation, where you click an object and get its outline, is unlimited and runs locally. Automatic detection across a zone and imagery edits use a monthly allowance that resets. Paid plans start when that allowance runs short. For scale, a paid plan costs about twenty minutes of billable consulting time a month, or roughly one percent of a commercial GIS seat.
- No. The heavy models run on our servers, so a standard work laptop is enough. Manual segmentation runs locally on the CPU. This matters more than it sounds: the newest Segment Anything release needs an NVIDIA card with CUDA and has no CPU path, which rules out every Mac and every machine on integrated graphics.
- Yes. Anything QGIS can load as a raster works: satellite scenes, aerial orthophotos, drone flights, scanned maps, XYZ basemaps and WMS layers. Detection quality tracks ground resolution more than the source, so a sharp drone orthophoto usually gives the best results.
- GeoAI is the use of machine learning on geographic data: satellite and aerial imagery, point clouds, vector layers. In practice it means models that read a map the way an analyst would, finding objects and patterns. Running it inside QGIS is what turns it from a research topic into a working tool.
- Manual segmentation never leaves your machine. Automatic detection and imagery edits send the selected tile to our servers to run the model, and we archive the tile so we can investigate a failed job. We do not sell it or train public models on it. The details are in the privacy policy.