// HOW-TO · AVATAR VIDEO

How to create an AI video clone of yourself (2026)

How to create an AI video clone of yourself: record training footage, clone your voice, pass the consent check, test the likeness, and publish on a cadence.

Last verified · 2026-11-07 · by Moe Ameen

An AI video clone is a digital replica of you — your face, your voice, your lip movement — that you drive with a typed script to produce talking-head video you never actually filmed. Done right, it lets one person hold a daily posting cadence off a single afternoon of recording. Done carelessly, it produces flat, slightly-uncanny videos that spend your credibility and, if you skip the consent step or clone someone who isn't you, break platform terms and likeness law.

This is the honest, do-able sequence: decide what the clone is for, pick the fidelity tier that matches, record clean training inputs, clone your voice, pass the platform's consent verification, test the likeness before you trust it, write to the clone's strengths, then caption, disclose, and publish on a schedule. The conceptual deep-dive — how cloning works and where it breaks — is in [AI video cloning](/guides/ai-video-cloning); for the audio half specifically see [voice cloning for video content](/how-to/voice-cloning-for-video-content).

The steps

Decide what the clone is actually for. Write down in one sentence the recurring job the clone will do — daily tips to camera, product explainers, multilingual announcements, a newsletter delivered as video. A clone pays off when you need the same presenter saying new things often; it is the wrong tool for one-off performance, anything emotional, or content that needs you to physically be somewhere. Naming the job now stops you from cloning yourself and then discovering half your content still needs a real camera.
Choose your fidelity tier: footage-trained vs one-photo. There are two paths at two quality levels. Footage-trained (HeyGen custom avatars, Synthesia personal avatars, Argil) learns from a few minutes of you talking to camera and holds up on longer, varied clips — choose it if the clone will carry a recurring channel. One-photo image-to-video animates a single still plus a voice sample; it is faster to stand up but flatter and weaker on long clips. For anything you will post repeatedly, pay the footage cost up front.
Record clean training footage (or pick your best photo). For the footage path, record yourself talking naturally to the lens for the length the tool asks — even, bright light, a quiet room, a plain background, eyes on camera, minimal big gestures. The clone can only be as good as this clip, so treat it like the one shoot that replaces a hundred. For the one-photo path, pick a sharp, well-lit, front-facing photo with a neutral expression and nothing crossing your face.
Clone your voice and lock the settings. A video clone needs a matching [voice clone](/glossary/deepfake-voice) or it will speak in a generic reader's voice. Record or upload a clean voice sample per your tool's spec (ElevenLabs is the common quality benchmark), then lock the voice and its stability and clarity settings once and build a short pronunciation glossary for names and jargon you repeat. Locking it means every future render sounds like the same person instead of drifting between videos.
Pass the consent verification — it is mandatory and it protects you. Reputable platforms will not train a clone without a recorded consent clip: you film yourself stating you authorize the use of your likeness, submitted alongside the training footage and biometrically matched to it. The face in the consent video must be the face being cloned, which is exactly why you cannot clone a colleague, a client, or a public figure from footage you merely have. Record the consent clip in the same lighting and framing as your training footage so the match passes cleanly.
Generate a test script and inspect the likeness. Before you trust the clone, render a short test — a few varied sentences, not one calm line. Watch the lip-sync on consonant-heavy words, listen for the voice holding tone across the clip, and check that pauses and emphasis land where a person would put them. If the mouth drifts or the delivery goes robotic, re-record cleaner footage or more voice samples; a weak clone is cheaper to fix now than to discover at video thirty.
Write to the clone's strengths. Script short, direct, to-camera lines — explainers, tips, announcements — and keep each video tight. Avoid asking the clone for big emotion, shouting, laughing, handling props, or gesturing off-screen; those are exactly where a near-real face reads as uncanny. Put the human, unscripted moments (a live reaction, an on-location shot) on a real camera and let the clone carry the repeatable, informational half of your calendar.
Caption, reframe, disclose, and publish on a cadence. Burn in word-synced captions (most short-form is watched on mute), reframe to 9:16 for TikTok, Reels and Shorts, and apply each platform's AI-content label — a realistic clone of a real person is precisely the case disclosure rules exist for. Then schedule rather than post by hand: a clone's whole advantage is cadence, and cadence is a queue, not a daily scramble. Repurpose each script into a carousel, a post, and a newsletter so one render session feeds a week.

Common gotchas

Treating one photo as equal to footage. A single-image clone is fine for a quick stand-up but flattens on longer clips — if the clone will run a channel, train it from footage.
Skipping or faking the consent step. You cannot clone someone who is not you, and an undisclosed clone of a real person is both a terms-of-service violation and a likeness-law risk in a growing number of places.
Scripting for emotion the clone can't deliver. Calm, informational delivery is convincing; genuine excitement, laughter, and big movement are where a near-real face breaks and reads as uncanny.
Letting the voice drift. If you don't lock the voice model and its settings, narration wanders between videos and the channel stops sounding like one person.
Thinking the clone writes your content. A clone renders a script faster — it does not supply the angle, the accuracy, or the reason anyone should watch. Garbage in, fluent garbage out.
Never disclosing. Audiences are getting better at spotting clones; an undisclosed one that gets caught spends trust you don't get back. Label by default and lean into it as an intentional format.

Legal note

You may only clone your own likeness, or a person who completes the platform's consent verification themselves — cloning someone else from footage you merely possess violates the major platforms' terms and, increasingly, right-of-publicity and deepfake statutes. Separately, many platforms require you to label realistic AI-generated video; disclosure (telling viewers it's AI) and consent (permission to build the clone) are two different obligations and you need both. This is general information, not legal advice — check the rules in your jurisdiction and each platform's current policy.

Where Kompozy fits

Standing up the clone is the afternoon's work; the part that quietly eats the next year is steps seven and eight, run once per video, forever — write a worthwhile script in your voice, render it, caption it, reframe it, disclose it, and get it onto every platform on time without the feed collapsing into sameness. That recurring loop, not the render, is where a clone channel dies. [Kompozy](/) is built to run exactly that loop.

Here's the honest fit first: Kompozy renders avatar video through HeyGen under the hood, so the clone's face and lip-sync are HeyGen's — Kompozy is not a separate fidelity lab. What it owns is everything around the render. Its AI Influencer persona pool is your productized clone, and a written [Persona Brief](/glossary/persona-brief) governs what it says — voice, audience, banned words — so video three hundred still sounds like you instead of drifting into generic AI narration. Step seven's "write to the clone's strengths" becomes a brief it enforces on every script; steps eight's caption, 9:16 reframe, and per-platform limits happen inside the render. The output lands as [Persona Shorts](/glossary/persona-shorts) (talking-head with auto-captions), Persona HeyGen (longer multi-scene), or Persona Frames (your clone in a brand-exact template), and the same script fans into a carousel, a quote graphic, a text post, a blog article, and a newsletter — so one clone session becomes a week of posts, not one clip. [Autopilot](/glossary/autopilot) and a per-post review gate run the daily cadence behind a human approval, which is also where you keep the AI disclosure honest.

The boundary, plainly: Kompozy does not create your consent record or train your likeness — that is HeyGen's gated flow and the consent clip is yours to record (steps three through five stay yours), and it does not supply your angle or guarantee the clone is on-camera-ready. What it removes is the operational wall between owning a clone and running a channel with it. Starter ($199/mo for 5,500 credits) fits a solo creator posting a few clone videos a week; Pro ($499/mo for 18,000 credits) sustains a daily, multi-format, multi-platform cadence; Enterprise is available through Contact Sales for teams and agencies running clones across many accounts.

Frequently asked questions

How much footage do you need to clone yourself on video?

It depends on the path. A footage-trained clone typically wants a few minutes of you talking clearly to camera, and more clean footage generally yields a better result. A one-photo image-to-video clone needs just a single sharp, front-facing image plus a voice sample, but it trades fidelity for speed and is weaker on longer clips. For a recurring channel, record the footage.

Do you need consent to make an AI clone of yourself?

For yourself, you provide the consent — the mainstream platforms require you to record a consent clip and biometrically match it to your training footage before they will build the clone. You cannot, however, clone anyone else that way: the person depicted has to complete the verification themselves, so you can't clone a client, a colleague, or a public figure from footage you have.

Why does my AI video clone look slightly off?

Usually one of three things: training inputs that were low-light, blurry, or too short; a script asking for emotion or movement outside the clone's calm, stationary comfort zone; or a lip-sync layer struggling on consonant-heavy speech. Re-record cleaner footage, write shorter and more direct lines, and keep the delivery informational rather than performative.

Is cloning yourself on video legal?

Cloning your own likeness and disclosing the result as AI is a legitimate production tool. The problems start when you clone someone else without their verified consent, or when you pass off a realistic clone of a real person as unedited footage without the platform's AI label. Keep it to your own face, disclose it, and you stay on the right side of both terms of service and likeness law.