Everything about Neural models

A face detector of 190 KB returns a grid of anchors, a box and five landmark points, and everything after that is arithmetic: head height, eye line, the angle to level by, millimetres to pixels. An application photo, explicitly not an official passport photo, made without uploading your face anywhere.

A scratch is thin and disagrees with its surroundings in almost every direction at once. A real edge disagrees in one or two. That single difference finds the damage with plain arithmetic, and an inpainting model fills what you agree to. All of it in your browser.

An encoder gives the hedge behind your subject as many bits as the subject. A saliency model knows better, but a browser will not let you set quality per region, so the encoder is not steered: its input is prepared instead. On the sample that was 41 per cent off the file at an unchanged quality setting.

A tool that guesses which text is sensitive will miss the one that mattered. This one covers every text field it finds and lets you click back what can stay. It also refuses to default to a blur, because blurring and pixelation can be undone, and there is published work showing exactly how.

JPEG damage is not random. It is made in blocks of eight by eight pixels, in a known order, which is why a model trained on compressed images beats any amount of sharpening. It runs in your browser, repairs at four times the size and hands the result back clean.

A cutout blur draws a line and blurs one side of it, which is a thing no lens has ever done. A depth model gives every pixel a distance instead, so sharpness can fall away gradually the way it does optically. The map is a downloadable asset in its own right, and none of it leaves your browser.

A clone tool copies pixels from elsewhere. An inpainting model predicts what should have been behind the thing you removed, which is a different question with a much better answer on grass, hedges and brickwork. It runs in your browser, and the rest of the photograph comes back byte for byte unchanged.

A background remover that works by colour asks every pixel a yes or no question, and the honest answer for a pixel that a strand of hair passes through is neither. Matting gives that pixel a fraction instead. Two small networks, both running in your browser, and the grey alpha view that tells you in a second whether the cutout will work.

A brightness slider adds the same amount everywhere, which is why a dark photo goes pale instead of bright. A small trained network predicts a different correction for every pixel, plus one gamma and one colour matrix for the whole frame. It runs in your browser tab and your picture never leaves it.