AI video cloning is the step past avatars: instead of picking a stock presenter, you build a digital replica of a specific real person — their face, their voice, their lip movement — and then drive it with a script to produce video they never filmed. In 2026 the technique went mainstream the moment a named operator put it to work in public: HeyGen and Ryan Serhant launched @HeySerhant, an avatar of his face and voice set to post daily for a year, with every post labeled AI. The demo is seductive and the capability is real, but the gap between "I made a clone" and "my clone runs a channel" is where almost everyone stalls. This guide is the honest version: what a clone actually is and how it differs from a stock avatar, the two very different ways one gets trained, what sits inside it beyond a face, why reputable tools gate the whole thing behind a recorded consent check, the jobs a clone is genuinely good at, the ones where it still visibly breaks, and the part the launch videos skip entirely — that a clone is a render, not an operation, and turning it into a recognizable feed across platforms is a separate problem with a separate answer.
AI video cloning is the production of new video featuring a digital replica of a specific, real person. You train the clone once on that person, then type or paste a script and the system renders them saying it — in their face, their voice, with their mouth shaped to the words — in footage they never actually filmed. The output is a talking-head presenter who is unmistakably a particular individual, generating content on demand.
The distinction that matters is clone versus stock avatar. A stock avatar video uses an invented presenter — a synthetic person who looks like nobody in particular, picked from a library. A clone is modeled on an actual human being, which changes everything downstream: the fidelity bar is higher because viewers may know the real person, the identity is an asset someone owns, and the consent question becomes central rather than optional. Cloning is the technique behind a creator turning themselves into an always-on channel, a company turning its CEO into a presenter who speaks forty languages, and — on the adversarial side — a bad actor turning a stranger's footage into words they never said.
There are two training paths and they sit at genuinely different fidelity tiers, so choosing the wrong one is the first mistake people make. The higher-fidelity path trains the clone from a short piece of consented footage — the person talking calmly to a camera for a few minutes under clean light and sound. Because the model learns from real motion and real speech, a footage-trained clone holds up on longer clips, varied sentences, and natural head movement. This is the path HeyGen's custom studio avatars, Synthesia's personal avatars, and Argil use.
The faster path builds a talking-head from a single photo plus a voice sample, animating the still image to match generated or cloned audio. HeyGen's image-to-video model and a wave of one-photo tools work this way. It is remarkable that it works at all, and it is the right choice for a quick stand-up or a presenter you will only use in short bursts — but it is flatter, it tends to degrade on longer clips, and it cannot reproduce how the person actually carries themselves, because it never saw them move. The practical rule: if the clone is going to carry a recurring channel, pay the footage cost up front. The one-photo version is covered in depth in AI avatar videos from selfies.
A convincing video clone is not one model but a stack of them working together, and the quality of the weakest layer is the quality the viewer perceives. The appearance layer reproduces the face and, on footage-trained clones, the head and upper body. The voice layer is a separate voice clone — a text-to-speech model tuned on samples of the person's speech so new narration sounds like them, not a generic reader. The lip-sync layer is the join between the two: it reshapes the mouth frame by frame so the visible articulation matches the audio, and it is usually the first thing that gives a weak clone away.
The frontier layer, and the one still least solved, is mannerism and motion — the specific way a person gestures, pauses, shifts their weight, and lets emotion cross their face. Footage-trained clones capture a shadow of this; photo-trained ones have almost none of it. This is why a clone can be flawless on a ten-second calm delivery and fall apart the moment the script asks for a laugh, a shout, or a gesture toward something off-screen. The face is largely solved. The behavior behind the face is where cloning still announces itself.
Because a clone depicts a real person, the mainstream platforms refuse to train one without that person's provable agreement — and this is not a checkbox. HeyGen, Synthesia and their peers require a recorded consent clip: the person films themselves stating that they authorize the use of their likeness, and that clip is submitted at the same time as the training footage and biometrically matched against it. The person depicted has to complete the verification themselves; you cannot upload footage of a colleague or a public figure and clone them, because the face in the consent video has to be the face in the training video.
Treat that gate as a feature, not an obstacle. It is the single mechanic that separates a legitimate owned clone from a deepfake, and it is why the consent question in avatar production is upstream of the disclosure question — they are not the same thing, as AI avatar likeness consent lays out. Consent is permission to build the clone at all; disclosure is telling the audience the finished video is AI. You need both, and in 2026 the legal exposure for skipping either is rising, not falling — see unauthorized AI likeness ads for how the creator-rights fight is unfolding.
Play to the strengths and a clone is a serious production asset. The first is volume from one shoot: film the training footage once and you can produce dozens of videos a week without filming again, which is the entire reason a solo creator or a small team can suddenly hold a daily cadence. The second is localization — a cloned presenter can deliver the same script in many languages with the voice preserved, which is why global teams reach for it for internal comms and product explainers. The third is consistency: the presenter is identical in every video, so a channel built on a clone has a stable, recognizable identity by construction.
The format fits a specific content shape: direct, informational, to-camera talking-head — explainers, tips, announcements, product walkthroughs, newsletters-as-video. It is a poor fit for anything that depends on physical performance, spontaneity, genuine emotion, or being somewhere real. Know which half of your content is clone-shaped and which half still needs a camera, and the tool earns its place instead of making everything you post feel slightly synthetic.
The honest failure modes are worth naming so you design around them. Emotional range is narrow: a clone delivering calm information is convincing, the same clone asked to be genuinely excited or moved usually is not. Movement is limited: most clones are stationary from roughly the chest up, so anything needing the body, a change of location, or interaction with an object falls outside the format. And the uncanny valley cuts the wrong way for clones specifically — because the face is a real, recognizable person rather than a cartoon, viewers hold it to a higher bar and small lip-sync or timing errors register as unsettling rather than charming.
There is also a trust cost that has nothing to do with render quality. As clones proliferate, audiences are getting better at spotting them and more skeptical when they do, and an undisclosed clone that gets caught spends credibility it does not get back. The move that survives this is disclosure by default — label the content as AI, lean into the clone as an intentional format rather than hiding it, and keep the genuinely human moments genuinely human. The manipulation risk at the extreme end is covered in the AI influencer manipulation trend.
The subject went mainstream in 2026 because a named operator demonstrated it in public rather than in a demo reel. HeyGen and real-estate entrepreneur Ryan Serhant launched @HeySerhant, an avatar built from his face and voice and committed to posting daily for a year, with his team supplying the ideas and scripts, HeyGen rendering the videos, and every post labeled as AI. Read past the celebrity and the mechanic is the point: the clone is not the achievement — the daily pipeline behind it is. The full breakdown is in the @HeySerhant launch.
This is the gap almost everyone falls into. Standing up a clone is now the easy part — an afternoon of footage, a consent clip, a few test renders. The hard part is everything the launch video skips: writing a worthwhile script every single day, keeping the clone's voice and claims on-brand across hundreds of posts, cutting each render to vertical with captions, publishing it to every platform on time, and doing all of that without the output collapsing into recognizable sameness that the platforms now penalize. A clone is a render. A channel is an operation. They are different problems, and owning the first does not solve the second.
This is where Kompozy fits, and the honest framing matters: Kompozy does not compete with a frontier avatar lab on raw clone fidelity — it renders avatar video through HeyGen under the hood, so the face and lip-sync are HeyGen's. What Kompozy is built for is the half the launch videos skip: turning a clone into a recognizable channel. Its AI Influencer persona pool is the productized, consent-gated clone — one primary identity with a pool of variations — and the whole engine is wired to keep that clone on-model while it posts at cadence across platforms.
Concretely, the clone's output lives in named formats: Persona Shorts (talking-head with auto-captions and optional B-roll), Persona HeyGen (longer multi-scene avatar video), and Persona Frames (the clone composited into a brand-exact template). A written Persona Brief governs what the clone says — voice, audience, banned words — so video three hundred still sounds like video one instead of drifting into generic AI narration. Autopilot and a per-post review gate run the Serhant loop without the celebrity deal: you supply the ideas, approve each render, and the engine fans it across eight social platforms plus blog and email with per-platform captions and length limits handled. And because one script becomes a carousel, a quote graphic, a text post, a blog article, and a newsletter in the same pass, a single clone session becomes a week of posts rather than one clip.
Be exact on the boundary. Kompozy does not create your clone's consent record or train the likeness — that is HeyGen's gated flow, and the consent clip is yours to record. It does not write your angle or decide what is worth saying. What it removes is the operational wall between owning a clone and running a channel with it: the scripting-to-brand, the format fan-out, the consistency, and the cross-platform distribution that otherwise turn a promising clone into three posts and then silence. The adjacent reads are AI avatars in video for the full avatar taxonomy and voice cloning AI for video content for the audio half of the same stack.
AI video cloning is building a digital replica of a specific real person — their appearance, voice, and lip movement — and driving it with a typed or pasted script to produce new talking-head video that person never filmed. It differs from a stock AI avatar, which is an invented presenter who resembles nobody. A clone is modeled on an actual individual, which is exactly why reputable platforms require that individual's recorded consent before they will train it.
Two ways, at two fidelity tiers. The higher-fidelity path trains the clone from a short piece of consented footage of the person talking to camera — a few minutes is typical — which captures how they actually move and speak. The faster, lower-fidelity path builds a talking-head from a single photo plus a voice sample, animating the still to match the audio. Footage-trained clones hold up far better on longer clips and real motion; photo-trained ones are quicker to stand up but flatter.
Yes, and the mainstream tools enforce it at creation. HeyGen, Synthesia and their peers require a recorded consent clip, submitted alongside the training footage and biometrically matched to it, so the person depicted has to complete the verification themselves. You cannot clone someone else from footage you merely possess. Using a person's likeness without permission is both a terms-of-service violation and, in a growing number of jurisdictions, a right-of-publicity or deepfake-law problem.
It is a stationary presenter, strongest when speaking calmly to the lens. It struggles with wide emotional range, big physical movement, handling props, and genuine spontaneity, and small errors in a near-real face read as uncanny in a way a cartoon avatar never does. It also cannot invent what to say — a clone with no script, no angle, and no brand voice behind it just produces fluent, forgettable video faster.
The technology overlaps, but the word that matters is consent. A clone you built of yourself, disclosed as AI, is a legitimate production tool. A deepfake is the same class of synthesis applied to someone without their permission, usually to deceive. The line is not the model — it is whether the depicted person agreed and whether the audience is told. See [deepfake](/glossary/deepfake) for the adversarial side of the same technique.
AI video cloning creates a digital replica of a specific real person — their face, voice, and lip movement — trained from a short consented video or from a single photo plus a voice sample, then driven by a script to produce new talking-head video the person never filmed. It is strongest for always-on posting, localization, and scaling one presenter; it is weakest at open emotion and real movement, and reputable platforms gate it behind a recorded consent check matched to the training footage.
Get started → · ← All guides · Compare Kompozy vs other tools