Labelme
labelme.io
- Free trial
- Windows
- macOS
- Linux
What Labelme is
Labelme is a self-hosted, AI-assisted annotation tool for creating training datasets on Windows, macOS, and Linux. It supports polygons, rectangles, oriented rectangles, circles, lines, points, masks, and rich attributes, with PNG, JPEG, BMP, and TIFF image formats. Local models include SAM, SAM2, SAM3, EfficientSAM, and YOLO-World. SAM3 accepts box and text prompts; SAM2 and EfficientSAM support point prompts, while YOLO-World returns bounding boxes. After one-time model downloads, annotation and AI assistance can work offline. The maker says images, annotations, and exports stay on the user’s machine without cloud processing or telemetry. Pro adds dataset conversion, batch processing, review, resizing, label renaming, statistics, exports, and command-line automation. Starter costs $49, paid once, Pro $79, paid once, and Pro (Lifetime) $249, paid once. A 3-day trial excludes the Pro toolkit.
Who it suits
Labelme suits researchers, engineers, and teams creating image datasets locally. It is useful when offline or air-gapped annotation and local model processing are priorities.
Good to know, and what to watch
- Supports many annotation shapes and attributes
- Runs several AI models locally
- Annotation can work offline after model download
- Keeps images and exports on the machine
- Exports YOLO, Pascal VOC, masks, and visualizations
- No text annotation support
- No API or SDK access
- Pro toolkit requires Python
- Intel Mac support ends at v7.0.4
Plans
Every plan and priceStarter
Pro
Pro (Lifetime)
Free plan · 3-day trial
The verdict
Labelme provides local, model-assisted annotation with broad shape support and offline operation. Its Pro toolkit expands dataset preparation, but text annotation and API access are unavailable and some platform limits apply.
What the maker publishes
- AI models
- It supports SAM, SAM2, SAM3, EfficientSAM, and YOLO-World, which run locally rather than through a hosted API. source
- AI prompts
- SAM3 accepts box and text prompts, while SAM2 or EfficientSAM can be used for point prompts; YOLO-World returns bounding boxes. source
- Annotation types
- It supports polygons, rectangles, oriented rectangles, circles, lines, points, masks, and rich attributes. source
- Dataset tools
- The Pro toolkit includes batch processing, visual review, image resizing with annotations, label renaming, dataset statistics, export, and command-line automation. source
- Founded
- 2016 source
- Integrations and exports
- Pro automates conversion to YOLO, YOLO-OBB, and Pascal VOC, can import YOLO and YOLO-OBB, and produces masks, visualizations, and dataset statistics. source
- Maker and history
- The maker identifies Kentaro Wada as founder and developer, and says the open-source project began in 2016. source
- Notable limits
- The three-day trial excludes the Pro dataset toolkit, Starter cannot be upgraded to Pro at a discount, and Pro toolkit tools require Python. source
- Offline use
- After each model's one-time download, annotation and AI assistance work offline; the maker says air-gapped use is possible by moving models from a connected machine or requesting its offline pack. source
- Open source and licensing
- The annotation application is GPL-3.0, while the Pro dataset toolkit is closed source; the AI models have their own licenses. source
- Operating system limits
- The desktop app supports Apple Silicon macOS, 64-bit Windows, and 64-bit Linux; Intel Macs are supported only through v7.0.4. source
- Privacy
- The maker says images, annotations, and exports remain on the user's machine, with no cloud processing or telemetry; model weights download on first use. source
- Purpose
- Labelme is an AI-powered image annotation tool for creating training datasets, aimed at researchers, engineers, and teams. source
- Support
- Pro and Pro Lifetime include priority email support with a stated 48-hour response, and team-license inquiries receive a response within one business day. source
- Team licensing
- Team and enterprise licenses are arranged by email with Pro per seat, one invoice, and volume pricing; each seat is for one named person. source
Checked against the maker's pages on 4 Oct 2026.