A Rust-backed data curation tool that runs on your laptop

What it does

LightlyStudio wraps dataset exploration, filtering, and annotation review into a local web UI. You point it at image folders, video directories, or existing COCO/YOLO datasets; it indexes them into an embedded DuckDB database and starts a GUI server on localhost. The pitch is keeping everything on-premise—no cloud upload required, though it can read from S3, GCS, or Azure if asked.

The interesting bit

The team wrote performance-critical parts in Rust, which is unusual for a Python data tool. They claim you can work with COCO and ImageNet on an M1 MacBook Pro with 16 GB RAM. The database is just a local .db file, so sessions survive restarts and you can append new data incrementally without re-indexing everything.

Key highlights

• Supports images, video, and multi-modal data (captions, image+text pairs)
• Imports COCO, YOLO, and ImageNet; exports back to standard formats
• Cloud storage via fsspec (S3, GCS, Azure) with optional dependency install
• Embeddable in Jupyter and Google Colab via iframe/kernel port tricks
• Python 3.9–3.14, cross-platform (Windows, Linux, macOS)

Caveats

• Several features are marked “support in progress” (🛠️): classification visualization, 3D point clouds, text handling, and keypoints are entirely unsupported (❌)
• Cloud-hosted annotations only work for COCO object detection and segmentation masks; other importers expect local files
• The README is heavy on quickstart snippets and light on architecture or embedding methodology

Verdict

Worth a look if you need a self-hosted, browser-based dataset browser for computer-vision work and prefer not to pay for cloud labeling platforms. Skip it if you need keypoint annotation, 3D point-cloud support, or a mature, fully-baked feature set today.