What a Deepfake Is, in Plain Terms

The word joins «deep learning» and «fake». Behind it sits one concrete task: to carry one person's face onto video of another so that the angle, the expression and the lighting all agree.

How this differs from compositing

Classic compositing cuts and pastes pixels. It works as long as the subject stays still: turn the head and the paste comes apart. A deepfake does not paste an image; it rebuilds the face from scratch in every frame, fitting it to the current turn, expression and light.

Hence its defining property: the result survives movement. And hence its central difficulty — the model must understand a face as a three-dimensional object rather than a flat picture.

What it needs as input

A reference: a set of frames of the face being transferred. The more angles and expressions it contains, the less the model has to invent. A single frontal portrait works too, but quality on turns will be lower.

Second, the target video. It supplies the movement, the light and everything outside the face: hair, figure, background.

Why the technology arrived when it did

All the components existed for years but converged recently: affordable graphics cards, open libraries and, above all, vast quantities of labelled facial images. The first public implementations appeared in 2017, and since then the barrier to entry has fallen from research-laboratory level to home-computer level.

Where it is used lawfully

• Film and television: de-ageing actors, doubles for difficult scenes, completing interrupted shoots
• Advertising: localising a spot for different markets without a reshoot
• Education: animating historical portraits in museums
• Dubbing: matching lip movement to a translated soundtrack

All these cases share one thing — the consent of the person whose face is used. Without it the application stops being a technology and becomes an offence.

This material is intended for adults over 18 and is informational in nature.