Lightly
Data curation and active learning platform that picks what is worth labeling before you label it.
About
Lightly is an ETH Zurich spin-off focused on the step most labelers skip: deciding what is worth labeling. Its platform uses self-supervised embeddings to select the most valuable samples, find near-duplicates and edge cases, and run active learning loops over image and video datasets. The flagship open-source library has over 3,800 GitHub stars, and the company is ISO 27001 certified and GDPR compliant.
The newer LightlyStudio (Apache 2.0) unifies curation, annotation and embedding search with a Rust backend built to handle millions of images, while LightlyTrain (AGPL-3.0, commercial licenses available) handles self-supervised pretraining and distillation from foundation models like DINOv2 and DINOv3 into production architectures such as YOLO. Caveat: the hosted cloud version of Studio is still rolling out, so some teams will self-host.
What it can do
- Embedding-based sample selection and deduplication
- Near-duplicate detection and edge-case discovery
- Active learning loops for labeling prioritization
- LightlyStudio: curation plus native annotation
- LightlyTrain: self-supervised pretraining and distillation
Setup
pip install the Lightly library, or run LightlyStudio locally and point it at your image or video dataset.
Prerequisites
Python environment; dataset on local disk or cloud storage (S3, GCS, Azure).
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Strengths: ETH Zurich spin-off, lightly repo 3821 GitHub stars (verified via API, pushed 2026-10-15), lightly-studio 896 stars, ISO 27001, GDPR; self-supervised curation + active learning + pretraining/distillation in one stack; Docked for: (1) hosted cloud Studio still rolling out; (2) AGPL on Train may deter some commercial users. Manual review: caveat on cloud rollout noted in entry. License: Apache-2.0 (LightlyStudio) / AGPL-3.0 with commercial licensing (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
- 9.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Strengths: ETH Zurich spin-off, lightly repo 3821 GitHub stars (verified via API, pushed 2026-10-15), lightly-studio 896 stars, ISO 27001, GDPR
- · self-supervised curation + active learning + pretraining/distillation in one stack
- · Docked for: (1) hosted cloud Studio still rolling out; (2) AGPL on Train may deter some commercial users
- · manual review: caveat on cloud rollout noted in entry
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.