On-device inference test

Runs the vehicle inspection models entirely in your browser, checks their output against a reference, and measures how fast your device is.

Pick a test

2 min Full test — run this one Both models across 8 inspection photos, all three runtimes. Ends with a Copy report button — paste that back to the team. 3 min Accuracy benchmark (optional) 200 validation images, computes detection accuracy on your GPU and compares it to the desktop reference. Extra 16 MB.

Before you start

  1. Open in Chrome (Android) or Safari (iPhone) — not inside another app.
  2. Stay on Wi-Fi. First run downloads ~40 MB; after that it's cached.
  3. Keep the screen on and don't switch apps while it runs.
  4. When asked, enter your name and your phone's model — e.g. Rahul and Pixel 8.

What it actually does

Two models run one after the other on each photo:

StageModelOutput used
1. Vehicle detectionYOLO11n, 80 COCO classes, 640×640 Boxes for car and truck
2. Parts segmentationFine-tuned YOLO, 21 car parts, 640×640 Which parts are visible
Processing details

The 8 photos

A full walkaround of one vehicle — four faces and four corners, so every panel is visible at least once.

Runtimes being compared

#RuntimeUsed when
1LiteRT + WebGPUPreferred — smallest and usually fastest
2TensorFlow.js + WebGPUIf LiteRT fails the correctness check (e.g. iPhones)
3TensorFlow.js + WebGLIf the device has no WebGPU at all

The test runs all three regardless, so we can compare. In production only the first one that passes would be used.