What it does

Wraps the Tiny YOLO model in a TensorFlow.js package you can npm install. Feed it a 416×416 image tensor, get back bounding boxes with class names and probabilities. There’s a live demo, and the whole thing runs client-side — no API calls, no GPU farm.

The interesting bit

The README is admirably blunt: full YOLO won’t run in tfjs, and “your user’s computers probably can’t handle it either.” So this is deliberately the small, slow version — ~800ms per frame on a 2014 MacBook Pro, which the authors cheerfully note gets you “1 FPS.” The honesty is almost refreshing.

Key highlights

• Ships as tfjs-yolo-tiny on npm/yarn with a one-line model downloader
• Returns clean box coordinates (top, left, bottom, right) plus className and classProb
• Configurable thresholds for class probability, IoU suppression, and max boxes
• Supports custom models: swap in your own anchors, dimensions, class count, and class names
• Peer-depends on tfjs; test coverage is explicitly “poor” per the authors

Caveats

• Locked to an old tfjs version (0.7.0 for tests); may need pinning to avoid breakage
• ~800ms/frame is real-time only if your definition of real-time is generous
• No full YOLO support, and the authors don’t pretend otherwise

Verdict

Worth a look if you need a quick browser prototype or want to avoid server costs for low-stakes detection. Skip it if you need speed, modern tfjs compatibility, or anything approaching production reliability.