Lightly
Data curation for computer vision: embedding-based selection, active learning, and self-supervised pretraining.
About
Lightly is a data-centric AI company spun out of ETH Zurich that helps vision teams train on the right data instead of all the data. Its core idea is embedding-based curation: self-supervised models index image datasets as vectors so teams can find near-duplicates, surface outliers, cluster visually similar samples, and actively select the most informative examples for labeling. That directly cuts annotation spend on large automotive, robotics, and manufacturing datasets.
The lineup now spans LightlyStudio, an open-source curation and labeling platform released in March 2026 with a Python-first SDK and DuckDB/Rust backend, and LightlyTrain, a PyTorch pretraining framework for YOLO, ViT, and detection models offered under AGPL-3.0 with commercial licenses. The open-source lightly library holds about 3,800 GitHub stars with commits days old, and the company reports ISO 27001 certification and GDPR compliance. It is less a full annotation workforce platform than a smart filter in front of one.
What it can do
- Embedding-based dataset curation and selection
- Near-duplicate detection and outlier surfacing
- Active learning for labeling prioritization
- LightlyTrain self-supervised pretraining (AGPL-3.0)
- LightlyStudio open-source curation platform
Setup
- Install:
pip install lightly(or the LightlyStudio package). - Point it at your image dataset in S3, GCS, or local disk.
- Run embedding-based selection to pick the most informative samples.
- Export the curated subset to your labeling tool, or pretrain with LightlyTrain.
Prerequisites
Python 3.8+; image data in cloud storage or on disk.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: 3,821 GitHub stars (GitHub API), MIT, pushed 2026-10-15; ETH Zurich spin-off documented; LightlyStudio launched Mar 2026 per company blog. Strengths: research-backed curation, strong OSS footprint, ISO 27001/GDPR posture. Docked for: (1) LightlyTrain under AGPL-3.0 limits commercial embedding; (2) curation-first, not a full labeling workforce. License: MIT (library) / AGPL-3.0 (LightlyTrain).
★ 3,821 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · 3,821 GitHub stars, MIT, pushed 2026-10-15 (GitHub API, lightly-ai/lightly)
- · ETH Zurich spin-off; ISO 27001 certified, GDPR compliant (vendor-reported)
- · LightlyStudio open-source curation platform launched Mar 2026 (company blog)
- · LightlyTrain under AGPL-3.0 with commercial licensing
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.