Lab

The same kind of model I put inside Seelai, my thesis app, to tell partially sighted users what is in front of them. Here it runs on your camera or a photo, with no server involved.

yolov8n · 80 classes640 × 640

See what YOLOv8 sees

Point your camera at something, or drop in a photo. Every box is found right here, on your device.

Use camera Upload a photo

or tryBarkadaGraduation day

Camera Upload

Confidence0.40

inference

—

speed

—

objects

—

runs on

—

Runs entirely on your device. Your camera and photos never leave the browser.

1. step 1

   Fit the frame

   Each frame is scaled to fit 640×640 and padded with grey, the way the model was trained.

2. step 2

   Run YOLOv8n

   ONNX Runtime runs the model in a background thread, on your GPU (WebGPU) when it can, otherwise WebAssembly. It scores 8,400 candidate boxes across 80 classes.

3. step 3

   Keep the best

   Boxes below the confidence are dropped, and overlaps of the same class are merged (non-max suppression).

Model: YOLOv8n by Ultralytics (AGPL-3.0), trained on COCO. Runtime: ONNX Runtime Web 1.22. The model downloads once (12.8 MB) when you first use it.