Models
Every model this site can download, with the figures that let you check it.
The ledger
Each model below is fetched from our model file server only when you press its button, after the button has printed its size. Its SHA-256 fingerprint is checked before a byte of it is cached or run, and it then runs on your own processor. Nothing about your picture is in the request.
| Model | What it does | File and size | Licence and source | SHA-256 | Run time |
|---|---|---|---|---|---|
| Face finder | Proposes a region over each face it can see. | face_detection_yunet_2026may.onnx224 KB · 229,738 bytes · fp32 | MIT · model repository | ebafce4e3c118d6554634be5c27ab333b4c047a9a8c3faf1d7cf93101c22f0f0 | about 16 ms (median, one thread, WASM) |
| Person finder | Draws the person, so the background around them can be blurred. | human_segmentation_pphumanseg_2023mar.onnx5.9 MB · 6,163,938 bytes · fp32 | Apache-2.0 · model repository | 552d8a984054e59b5d773d24b9b12022b22046ceb2bbc4c9aaeaceb36a9ddf24 | about 90 ms (median, one thread, WASM) |
| Depth finder | Estimates near and far, for a blur that grows with distance. | midas-small.onnx63.7 MB · 66,764,249 bytes · fp32 · above the 25 MiB per-file guideline, so it is offered under Advanced only | MIT · model repository | 2d8c6cb8f415229daf1eb041024208e2608c9f98e17c81cc7c6ecb449c56fd58 | about 500 ms (median, one thread, WASM) |
The runtime
The models run in ONNX Runtime Web 1.29 (MIT), served from this domain: a 13.3 MB WebAssembly file (13,961,845 bytes) and a 24 KB loader. It is fetched only with the first model you ask for. The WebGPU build was measured at 27.8 MB, above the 25 MiB per-file limit, so the site runs on WebAssembly with one thread.
What runs without a model
- The text-line finder is arithmetic on the picture's own pixels: it finds strokes that sit in a row. Nothing is downloaded and no word is read.
- The blur checker reads flat fills, mosaics and blurs by measuring brightness steps. No model.
- Every blur, mosaic and solid fill is plain code in this site's own bundle.
Not here, and why
- Number-plate finder: no plate detector has yet passed this site's licence and measurement check, so plates are covered by hand with a plate-shaped drag.
- Reading text (OCR) to highlight numbers: it would need a new model on the model file server; until then nothing on this site recognises a letter.
- A general object finder for pets and products, and a smaller depth model: none is published on the model file server yet. The background blur uses the person finder or the oval you place.
A figure is added to this page only after the file has been downloaded, fingerprinted and timed.