AI avatar video generators: script or performance

Two quite different products sell themselves under the same phrase. One assembles a speaker out of typed text. The other takes a performance that a person actually gave and moves it onto someone else's face. They cost different things, they fail in different places, and the difference decides which one belongs in a paid ad.

What sits behind the phrase

An avatar video generator returns a video of a person talking, built from something you supply. What you supply is the whole argument.

In the script route you supply words on a page. A voice model reads them and a face is animated to fit the audio. In the performance route you supply a take: someone on a phone, saying the words out loud, and the system moves that delivery onto a different face.

Both hand back a video of a person who is not you. Only one of them contains a performance that a human gave.

The script route, step by step

You type or paste the script. You pick a stock avatar, or build one from a photo or a short enrolment recording. You pick a voice, from a library or cloned from a sample of your own. The system synthesises the audio and animates the mouth, and usually the head, to match it.

The appeal is real. No camera, no room, no retakes, and no negotiation with anyone about appearing on screen. Twenty language versions of one script are twenty generations, and changing a price in the copy is an edit rather than a reshoot.

The bill arrives at delivery. A voice model reads punctuation, not intention. It puts the stress where the grammar suggests rather than where the meaning is, so an ironic line lands flat, and a question at the end of a list is read as another item on the list. Faces animated from that audio inherit the problem: the mouth is busy, the rest of the face is calm, and the eyes point at the lens without ever quite looking at anything.

The performance route, step by step

You record the take. Your voice, your pauses, the half second where you glance down at the product and back up, the breath before the hook. The system reads that recording as a performance and drives another face with it. The timing comes from the recording rather than from a model's guess about how the sentence should sound.

The bill arrives earlier here. Somebody has to stand in a room and deliver the script convincingly, and a weak take produces a weak ad on a nicer face. Twenty language versions are twenty recordings, not twenty pastes, unless the speaker happens to be fluent in twenty languages.

The lip sync question

Lip sync is the thing people check first and the least interesting difference between the two routes.

Script-driven tools are usually good at it, because the mouth is being animated from the audio waveform that the same system just produced. The two can hardly disagree. Performance-driven tools take their mouth shapes from a real mouth, so the sync is as good as the source recording and degrades with bad audio, heavy occlusion or a face half out of frame.

What separates them is everything above the mouth. Audio carries phonemes cleanly, and it carries almost nothing about where the speaker was looking, when they tilted their head, or which word they decided to lean on. A performance carries all of that because it recorded it happening.

Gesture, gaze and the parts nobody writes down

Read a script aloud and count what you do that is not in the text. You mark the list with your hand. You look away on the concession and back on the claim. You speed up through the caveat because you know it is the boring part.

None of that is in the words, so a script-driven system either omits it, which reads as stiff, or samples it from a bank of idle motions, which reads as busy and disconnected. In a feed the viewer does not diagnose any of this. They scroll.

Which one belongs in a creator-style ad

Creator-style ads are built around a demonstration: the product enters the frame, one claim is shown rather than stated, and a person tells you what to do next. The anatomy of a UGC ad is that structure, and it is mostly physical.

A script-driven avatar cannot hold your product, because it was never in the room with it. That makes it a reasonable fit for an announcement, a policy update or a spoken explainer, and a poor fit for an ad whose job is to show something working. The guide to AI actors covers the casting side of the same decision, and the guide to UGC covers where the format came from.

What you have to disclose either way

Two obligations attach, and neither cares which technique produced the face.

The FTC's Endorsement Guides define an endorser to include parties that merely appear to be people, which reaches virtual and fabricated endorsers. A material connection between that endorser and the advertiser still has to be disclosed clearly. The Commission's rule on consumer reviews and testimonials, effective 21 October 2024, separately prohibits testimonials from someone who does not exist or never used the product, which is exactly what a generated presenter reciting a first-person customer story is.

In the EU, Article 50 of the AI Act applies from 2 August 2026 and obliges deployers who generate or manipulate video constituting a deepfake to disclose that it is artificially generated. Plan the label into the creative rather than bolting it on afterwards.

How Sosie fits

Sosie is the performance route. You record up to forty seconds on a phone, saying your own script, in whatever room you are in. You choose a presenter from the catalogue, describe one for Sosie to generate, or upload an image of someone who gave you written permission; public figures are blocked. Your voice, your timing, your expressions and your gestures are kept and delivered by that presenter, who is placed in your own room by default, with its walls and its light, or on their own backdrop instead.

What comes back is a vertical 1080p MP4 with no watermark, cleared for commercial use including paid placements, in about five minutes for a twenty-second video. One credit covers up to twenty seconds and two cover up to forty. Credits come as monthly subscriptions or as one-time packs that never expire, from €39 for five credits, and no plan exists without payment. A render that fails is refunded.

A quick way to tell them apart

Write one line that only works with a shrug in it, and put it in the middle of the script. Generate the video both ways. The script-driven version will say the words. The performance-driven version will shrug, because somebody did.