How to make AI talking head ads with lip sync

What is an AI talking head ad?

An AI talking head ad is a short video in which one AI-generated person speaks straight to the camera and recommends, explains or reviews a product. The face, the voice and the mouth movement are all generated, so the lip sync is what makes or breaks it.

It is the simplest AI ad to describe and one of the hardest to get right, because viewers watch the mouth. Public examples, seated and in the car, are on AI talking head ads, and the expert-presenter version is on AI expert ads. Those come from posts on X; they were not made on or for Anglr.

Talking head is also the core of the selfie UGC format. This guide covers the speaking part: casting, script, dialogue formatting and lip sync checks. The AI UGC ads guide covers the rest of a UGC ad.

Tools creators use for it

The table counts how many of the reviewed tutorials named each tool for a speaking clip. A tutorial that used two tools counts once for each.

Based on 10 public tutorials posted between 27 Jan 2026 and 8 Oct 2026, reviewed by Anglr.

Single tutorials also used Arcads for a prompt-built speaker with a voice, ElevenLabs for a separate voiceover, Google Flow for lip sync on animated stills, Nano Banana for the start frame and Claude for the script and prompts. We also track HeyGen, an avatar tool none of these tutorials used. Two tutorials did not name their tools. Few public posts teach this format step by step, so the counts are small.

How do you make an AI talking head ad?

You make it in six steps: choose the route, cast the speaker, lock the script, write the lines in the model's format, generate short clips, then check and cut. The order is our synthesis of the tutorials.

1. Choose how the voice gets made

There are three routes, and the choice decides everything after it.

Speech generated inside the video model. The model makes the voice and the mouth together from the lines in your prompt. This is the most common route in the tutorials: Mike Futia keeps speech, lip sync and ambience inside Seedance with no separate voiceover step, @AvelyrahnAI gets narration from one Kling prompt, @davidfigeira generates the whole talking clip from one Seedance prompt, and Chibueze Kingsley gets lip sync from Google Flow.

An avatar tool. You give it a face or a description plus a script, and it returns a speaking video. @1littlecoder uploads a product photo and a script and gets a vertical talking head back. @David_TornAI defines face, voice and personality in one prompt and renders in Arcads. Fast, but you get less control over setting and camera.

A separate voiceover. Will Connell makes the voiceover in ElevenLabs and the motion in Kling. This gives the most control over the voice, but the mouth has to be matched to audio that already exists.

2. Cast one speaker as a realistic still

Create the speaker as an image before any video. @hamzaoudghiri2 writes age, skin texture, outfit, setting and a front-camera, unretouched look into the prompt. @sixugc starts from an avatar reference and makes a selfie-style start frame from it. Mike Futia writes three casting prompts and picks the person before spending a video credit. @David_TornAI fixes the voice and personality at the same time as the face.

Pick a face that suits the claim. An expert-style presenter suits explainers; a casual, slightly messy look suits reviews. Keep the speaker fictional, and if they play a professional, avoid anything that implies a real credential the brand cannot back up. To reuse the speaker, @hamzaoudghiri2 saves the face as a Soul ID in Higgsfield.

3. Lock the script before you generate

Write every line the speaker says, word for word, before you make any video. Chibueze Kingsley locks the script first because changing dialogue later means generating again. @1littlecoder feeds the finished script straight into the tool. Mike Futia's skill builds the full script, question by question, before it writes a single video prompt.

Write for the mouth as well as the ear. Short sentences, one idea each, and words people actually say. Mike Futia has his speaker hesitate and concede a point instead of listing benefits, which reads as more human. Our UGC script template and ad hooks guide cover the hook and structure.

4. Write the lines in the model's own dialogue format

Each video model reads dialogue in its own syntax, and the vendors document it. Using the wrong one is a common reason a line is skipped, said by the wrong person or turned into on-screen text.

Our prompt pages for Seedance, Kling, Veo and Higgsfield show each format in full.

5. Generate short clips and keep the voice the same

Generate the ad as short speaking clips, a few lines each, rather than one long take. Mike Futia splits his ad into four segments, gives each three beats, spells out the voices in a sound block, and feeds the last frame of each clip in as the reference for the next. Higgsfield's help center describes the same last-frame chaining.

Voice drift between clips is the other half of consistency. Describe the voice the same way in every prompt, or bind it once where the model allows it. Google's Veo guide adds a practical warning: voice does not extend well from a clip with no speech in its final second, so end each clip with the speaker still talking if you plan to extend it.

6. Check the mouth, then cut

Watch every clip at full size before it goes into the edit. Look for mouths that keep moving after the line ends, teeth that change shape, lips that do not close on sounds like "b", "m" and "p", and eyes that never blink. Regenerate only the clips that fail.

Then cut the clips together, add captions, and add product cutaways over any weak moment. @davidfigeira makes many versions from one script by changing the face, tone and setting; once one version works, that is the cheapest way to test.

Prompt template for a talking head clip

This template is ours. It follows Google's Veo 3.1 order (camera, subject, action, context, style), with speech in quotation marks after who says it and labelled sound lines. Duration and aspect ratio are settings, not prompt text. Build your own in the AI ad prompt builder.

Speaker still (image model):

A vertical phone front-camera photo of a fictional [age] [gender] [role, e.g. home cook] sitting at [kitchen table or car seat], facing the lens at eye level. [Two or three stable features: hair, top, glasses]. Soft daylight from the side, natural skin texture, relaxed mouth, unretouched look. [Product] on the table beside them, label facing the camera.

Speaking clip (Veo 3.1, image-to-video with the still as reference):

Static phone camera at eye level, slight handheld movement. The woman from the reference image looks into the lens, picks up [product] and holds it beside her face. She says, "[line one, under ten words]." She pauses, then says, "[line two, under ten words]."
Style: casual phone video, natural light, no music.
SFX: the soft tap of the bottle on the table.
Ambient noise: quiet kitchen room tone.

Kling version of the same lines:

The woman from the image looks into the lens and lifts [product] beside her face.
Woman (warm, American accent): [line one]
Woman (warm, American accent): [line two]
Quiet room tone, no music.

Common problems and fixes

What creators report

These are the authors' own claims, not results we checked.

• @1littlecoder reports getting a vertical talking head video in a couple of minutes from a photo and a script.
• @davidfigeira reports generating a full talking clip in under two minutes from one prompt.
• Mike Futia reports holding character, voice and product steady across four 15-second segments with his chaining method.

Talking head ads are mostly script and casting. If yours already perform, the creators page explains how Anglr pays for them.