Lightly
Zurich data curation, labeling and evaluation for computer vision, with Apache-2.0 LightlyStudio and LightlyTrain
Quick Overview
| Company | Lightly AG |
|---|---|
| Category | Data Labeling Platforms |
| Headquarters | Zurich, Switzerland |
| Founded | Not stated |
| EU Presence | Yes - Switzerland |
| Data Location | See website |
| Open Source | Apache-2.0 (LightlyStudio repository on GitHub) |
| GDPR Compliant | Yes |
| Self-Hosting | Yes - open source LightlyStudio |
| Main Features | Embedding-based sampling, similarity search, annotation and QA, evaluation with confusion matrix, SAM3 plugins, role-based access, COCO, Pascal VOC and KITTI export |
| Pricing | LightlyStudio free to try; commercial pricing on request via demo |
| Best For | Computer vision teams reducing labeling cost |
| Replaces | Weights & Biases Tables and Artifacts, Roboflow, V7 Darwin |
Detailed Review
Lightly addresses the data step: which of millions of images to label. Everything you index gets an embedding, and six sampling strategies, including diverse coverage, deduplication, outlier detection, similarity and class balancing, select the subset worth sending to annotators.
The imprint names Lightly AG, Bahnhofstrasse 86, Zurich, registered in the Commercial Register of the Canton of Zurich (CH-020.3.046.583-0).
Curate, label, evaluate
LightlyStudio combines curation, annotation and QA in one place and gives each labeled object its own embedding, so you can search objects as well as images. SAM3 and LightlyTrain plugins auto-label the easy cases.
An evaluation view compares predictions with ground truth and shows per-sample metrics and a confusion matrix, sortable by false positives, false negatives or mIoU.
Teams and formats
Roles are Viewer, Labeler, Editor and Admin. Data exports to formats including COCO, Pascal VOC and KITTI, and the stack integrates with PyTorch, ONNX and Ultralytics.
Plugins are Python functions with a form, which lets non-developers run custom steps from the UI.
Training and services
LightlyTrain pretrains vision models without labels, and the site also lists LightlyEdge and a data collection and annotation service delivered into LightlyStudio.
LightlyStudio's repository is Apache-2.0 licensed.
Limitations to Consider
Lightly is for vision data, so tabular, text or audio teams get less value. It does not track experiments, orchestrate pipelines or serve models, which means pairing it with ZenML, Valohai or DagsHub. Commercial prices and data location are not published on the pages read. Annotation and data services are an extra commercial layer.
Who Should Use Lightly
Swiss and European computer vision teams, including robotics and inspection, that want to cut labeling volume before training.
Pros and Cons
Pros
- Embedding-based sampling reduces labeling
- Annotation, QA and evaluation in one tool
- Apache-2.0 LightlyStudio
- Swiss company with published register number
- Common export formats
Cons
- Vision data only
- No experiment tracking or serving
- Commercial prices not published
- Data location not stated on pages read
Alternatives to Lightly
Other European data labeling platforms worth comparing before you decide:
Kili Technology
France · Expert-in-the-loop labeling and AI evaluation platform
Kognic
Sweden · Annotation platform and workforce for autonomous driving data
Prodigy
Germany · Scriptable self-hosted annotation tool from the spaCy makers
Supervisely
Estonia · Labeling, data management and model training in one platform
See all European data labeling platforms or compare them directly against Weights & Biases Tables and Artifacts on our alternatives page.
Frequently Asked Questions
The site says LightlyStudio is live and free to try, and its GitHub repository is Apache-2.0. Commercial plans are arranged through a demo.
It selects which images to label using embeddings and handles annotation and QA, which is data curation rather than run tracking.
The pages describe computer vision data, including images, video and sensor data for physical AI use cases.