lightly-ai/lightly-studio

What it solves

LightlyStudio is a local-first data curation tool designed to help developers manage, annotate, and evaluate computer vision datasets. It addresses the problem of handling massive amounts of unlabeled or poorly labeled image and video data, reducing the cost of manual labeling by identifying the most diverse and representative samples for annotation.

How it works

The tool runs as a local server that opens in a web browser. Users index their image or video folders (including those in cloud storage like S3, GCS, or Azure) into a local database. It uses embeddings to visualize the dataset as a plot where similar images are grouped together, allowing users to lasso-select clusters of data to filter the grid. It also supports a Python API for scriptable indexing, querying, and exporting data in formats like COCO and YOLO.

Who it’s for

It is primarily for computer vision engineers and ML practitioners who need to curate high-quality training data for models like YOLO, and those who want to keep their data local for privacy and security.

Highlights

• Local-first architecture: Data and images never leave the local machine.
• Embedding-based curation: Visualizes data relationships to find duplicates or typical samples.
• Scalability: Handles over 2 million images on a single MacBook M1.
• Integrated annotation: Includes an annotation editor and supports plugins for auto-labeling (e.g., SAM).
• Model evaluation: Compare model predictions against ground truth to find failure patterns via confusion matrices.
• Multi-format support: Works with image folders, videos, COCO, and YOLO datasets.