The Best AI Plugins for QGIS in 2026, Compared

Eight AI plugins you can install in QGIS today, grouped by what they do. Prices, licences and install counts were read from each plugin's own source on 18 August 2026.

The comparison

Install counts come from the official QGIS plugin index, the same field for every row. The AI Segmentation and AI Edit product pages quote a lower number, the sum of the per-version rows, which a reader can recount by hand.

Extract objects from imagery

1. Mapflow

Pick an area, pick a model, get vector layers back. Buildings with optional height, roads, agricultural fields, forest with optional crown, construction sites.

• Runs on Geoalert's servers. No graphics card, nothing to install beyond the plugin.
• Price $0.10 a credit. Buildings cost 10 credits per km², rounded up to a whole km². Paid from $50 for 500 credits, then $100 and $300 a month.
• Free tier 250 credits, capped at 25 km², download disabled.
• Licence the file says GPL v3, the README says GPL v2 or later. Read the file.

Docs · Video walkthrough of a building run

2. AI Segmentation

AI Segmentation runs its model on a server, so nothing installs beyond the plugin: no weights download, no CUDA, no Python environment. Plugin manager, QGIS 3.22 and up. Draw a zone, type building, get every one back as a polygon.

Filtering happens in the panel, not the attribute table. On one Paris run, dragging confidence and area took 635 polygons down to 82, and those 82 still held 84% of the built area.

Free accounts get 200 automatic detections a month with no card and zones up to 15 km², and the Semi-Auto mode on your own machine stays unlimited. Pro is €39 a month for 5,000 detections and no zone limit. Output is a styled GeoPackage with a label, a class, a confidence score and an area per feature.

Imagery is processed on a server in Europe, so it does not suit work that may not leave the machine.

Try it free in QGIS, no card needed

Run a model on your own machine

3. GeoAI

The deepest toolbox here, from Qiusheng Wu. SAM, DeepForest for tree crowns, OmniWaterMask, Mask R-CNN, and a trainer for your own model.

• Runs on your machine. CUDA recommended, Apple MPS supported, CPU fallback. SAM 3 needs an NVIDIA card.
• Price free, MIT.
• Version pinned to the QGIS 3.44 series, to avoid a regression in 3.42.2.
• Install budget an evening. I failed on both my Mac and my Windows machine, and my hardware was not enough for the runs I wanted. Reported failures are overwhelmingly Windows and the maintainer closes them quickly: #827, #519, #414, all closed.

Docs · Step-by-step setup video

4. Deepness

Hand it an ONNX file, point it at a raster, it runs the model tile by tile. Its zoo covers land cover, buildings, roads, solar panels, tree tops and cars, each only reliable on data resembling its training set.

• Runs on your machine. Works on CPU, a GPU needs your own CUDA setup.
• Price free, Apache 2.0. Published in SoftwareX in 2023 if you have to cite it.
• Install about twenty minutes for me, and worth it. The plugin pip-installs OpenCV and ONNX Runtime on first use, and that is what breaks: #197, #235 still open, #193.
• macOS the maintainers have no Mac to test on (#209) and the install page reads "MacOS - SOON".

Docs · Model loading and inference video

Classify pixels into land cover

5. SCP

Luca Congedo's plugin, in the repository since 2013 and installed 2.6 million times, more than everything else here put together. You draw training areas, pick minimum distance, maximum likelihood or spectral angle mapping, and get a classified raster. It downloads Landsat and Sentinel-2 too.

• Runs on your machine, CPU.
• Price free, GPL v3.
• Not the non-AI option any more version 9 added random forest, a neural-net classifier and pretrained models.
• Check first the 9.x line needs QGIS 4, so QGIS 3 gets 8.5.0 from November 2024. Apple Silicon has an open crash report.

Manual · Supervised classification walkthrough

Three that do a different job

6. Bunting Labs AI Vectorizer

Autocomplete for hand digitizing. Start editing a layer, place two vertices along a road or a river, and it traces ahead of your cursor. It never returns every match in an area.

• Runs on their servers. The imagery around your cursor is sent, the geometry comes back.
• Price no free tier, only a first day at $1. Personal $29 per seat per month, capped at 120 completions a day. Professional $99, uncapped, no data retention.
• Licence GPL 2.0.
• Note their listing advertises georeferencing too, but the pricing page meters AI Georeferencer separately, at 25 or 100 a month.

Docs · Their own demo of the autocomplete

7. QGIS MCP

Nicolas Karasiak's plugin opens a socket inside QGIS, and a companion server exposes 118 operations to any Model Context Protocol client: Claude Code, Claude Desktop, Codex, Cursor, VS Code. Layers, features, editing, styling, processing, rendering, layouts, atlas, SQL.

• Runs on your machine. The model runs wherever your client points.
• Price free. Plugin GPL 2.0, server MIT. You pay for your own LLM.
• Needs QGIS 3.28 or later, plus uv.
• Ships no detection model, so it is not an alternative to the others. It is the thing that could call them.

Setup and agent integration guide

8. AI Edit

AI Edit changes the pixels rather than reading them: clouds off a scene, a summer field turned winter, an orthophoto rendered as a clean land-cover map, then vectorized into a layer.

• Runs on a server. No graphics card.
• Price €29 a month, three free generations.

OriginalClouds removed

Drag to compare

Installing them

The plugins that run locally are free. The price is your afternoon. QGIS ships its own Python on Windows, dependencies land in the other one, and a numpy 1 against numpy 2 mismatch takes the whole plugin down. I failed to install GeoAI, Deepness and GeoOSAM on both my Mac and my Windows machine on the first attempt. Deepness eventually worked and was worth it.

• Install one at a time and restart QGIS between each. A failed dependency install often leaves the environment worse than it found it.
• Read the requirements page before the plugin manager. GeoOSAM states in bold that both of its install methods still need a manual pip install of nine packages.

Also worth knowing. Smaller or newer, all alive as of August 2026, and all free:

• SamGeo, Qiusheng Wu again, SAM segmentation from text, point or box prompts.
• Geo SAM, SAM 1 through 3 for landforms, one object at a time.
• GeoSeg Studio, a full semantic segmentation pipeline for raster.
• Aerial LiDAR Classifier, deep learning semantic segmentation for aerial LiDAR.
• AgenticGIS, an in-QGIS chat assistant that runs QGIS operations through an LLM.

Questions

What is the best free AI plugin for QGIS? For extracting objects from imagery on your own machine, GeoAI. It is MIT licensed, it covers SAM plus tree crowns, water masking and Mask R-CNN, and it has more installs than any other free option here except SCP. Budget an evening for the install. For land cover from your own training areas, SCP, which has been doing it since 2013.

Is there an AI plugin for QGIS that works without a GPU? Yes, four of them. Mapflow and our AI Segmentation run the model on a server, so any laptop works. Bunting Labs also runs remotely. QGIS MCP needs no model at all. Everything that runs locally wants a graphics card once you move past small areas, and SAM 3 specifically needs an NVIDIA card with CUDA.

What is a cheaper alternative to Mapflow? It depends on the area. Mapflow charges 10 credits per square kilometre for buildings, and a credit is $0.10, so a hundred square kilometres of buildings is about $100. Our AI Segmentation is a flat €39 a month for 5,000 detections with no area cap, and free up to 200 detections a month. If you want no bill at all, GeoAI and Deepness are free and run on your machine.

Does QGIS have Segment Anything? Not in core QGIS. Several plugins bring it: GeoAI and SamGeo from Qiusheng Wu, Geo SAM, and AI Segmentation, which runs SAM 3 on a server. The SAM guide compares them properly.

Which of these can I use on client work without a licence problem? GeoAI is MIT and Deepness is Apache 2.0, both permissive. SCP is GPL v3. Bunting Labs is GPL 2.0 and QGIS MCP is GPL 2.0 with an MIT server. Mapflow ships a GPL v3 licence file while its README says GPL v2 or later, so read the file if it matters to you. None of that covers the imagery you feed them, which stays your responsibility.

Can I run any of this on a batch of 400 tiles? Deepness tiles a raster natively. Mapflow processes an area server-side. For a real pipeline you want Python rather than a plugin, and segment-geospatial is the package to start from.

Pick by task

The SAM guide goes deeper on which plugins run which version of Segment Anything, and the building footprints guide walks one job end to end.