Lightly

Lightly is a computer vision data platform for ML engineers and organizations scaling image and video models. It combines two products: LightlyStudio, a data hub for curating, annotating, evaluating and visualizing datasets, and LightlyTrain, which trains production models from that data. In LightlyStudio, everything indexed is embedded automatically, so teams can filter by metadata, search by text or image, and run sampling strategies such as diverse coverage, deduplication, outlier detection, similarity and class balancing to choose which samples are worth labeling. Annotations can be created, edited and reviewed in the same place, with object level search and auto labeling plugins built on SAM3 and LightlyTrain models. Evaluation runs against ground truth and produces per sample metrics and a confusion matrix for detection, classification and segmentation. Plugins are Python functions with a form, so teams can add their own. Roles for viewers, labelers, editors and admins control access, and data exports to COCO, YOLO, Pascal VOC and other formats through the interface or Python. LightlyTrain releases include real time detectors and depth estimation inference. The company was founded in 2019 out of ETH Zurich research. The site offers a free start and demo booking. Paid plan prices are not published.