Models and files

Everything this site can download into your browser to do its work, with the exact size, licence and fingerprint. None of it loads until you press the button that needs it, and every one of these files travels one way: to your device. The picture never travels to them.

The describer: not released

There is no captioning model on this site yet, so nothing here writes a sentence about what a photograph shows. The plan is an MIT-licensed model that runs in the browser, published here with its size, licence, parts and known failures before its first byte can be downloaded. Until then the subject of a photograph is the one thing you type, and the checker, the measurements, the detectors and the text reader below work without it.

Text reader

Engine: Tesseract 5 LSTM (tesseract.js 7.0.0), compiled to WebAssembly, in a background worker
Licence: Apache-2.0 (engine and language data)
Comes from: This site, under /ocr/
Runtime files: 3.9 MB (62,961 + 112,122 + 3,899,472 bytes; the core build is chosen by your browser’s SIMD support)
English data: eng, tessdata_best (integer), 2,952,873 bytes, SHA-256 45b4cb346724ac1774f1c36f42f182b887bcdb28ebe63e6fff90ac41f3fcff91
Hindi data: hin, tessdata_best (integer), 1,389,692 bytes, SHA-256 f3b6a0d320df38d886178cdd727b90dbf9df3db053adb32bd9cf73f0463cda07
Adds: The words of printed English and Hindi text in a picture, with a box and a confidence for every line and word
Fails on: Handwriting (not supported), decorative or very small type, text at a steep angle, low light and heavy compression. Mixed Hindi and English lines need both languages loaded. Measured accuracy figures for this site’s test pictures are not published yet, so none is quoted.

Detectors

Three small detectors that place things in a picture. They count and locate; none of them identifies anyone. Their model files come from our model file server and are checked against the fingerprint below before they run. The runtime that executes them, ONNX Runtime Web (13.3 MB, WebAssembly, single-threaded), comes from this site.

Face finder

Model: YuNet 2026may
Licence: MIT
Size: 229,738 bytes (224 KB)
SHA-256: ebafce4e3c118d6554634be5c27ab333b4c047a9a8c3faf1d7cf93101c22f0f0
Adds: where faces are and how many — never who
Fails on: Misses faces turned away, small faces under about 10 pixels at the size it reads, and heavily occluded faces. Painted or printed faces can be counted. It never says who anyone is.

Person finder

Model: PP-HumanSeg 2023mar
Licence: Apache-2.0
Size: 6,163,938 bytes (5.9 MB)
SHA-256: 552d8a984054e59b5d773d24b9b12022b22046ceb2bbc4c9aaeaceb36a9ddf24
Adds: how much of the frame is a person, and where
Fails on: Trained mostly on upright people in portrait framing: people lying down, very small in the frame, or in dense crowds are under-counted, and mannequins or statues can be marked as people. Touching people merge into one area.

Depth finder

Model: MiDaS v2.1 small
Licence: MIT
Size: 66,764,249 bytes (63.7 MB)
SHA-256: 2d8c6cb8f415229daf1eb041024208e2608c9f98e17c81cc7c6ecb449c56fd58
Adds: which parts of the frame are near and which are far
Fails on: Gives relative near/far, not distances. Mirrors, windows, sky and flat graphics confuse it; on a chart or a screenshot its answer means nothing.

What the browser keeps

Detector models are kept in the browser’s Cache Storage so they do not download twice, and the studio’s detector panel deletes them. The text reader’s files sit in the ordinary HTTP cache. No picture, no text and no result is stored.