AI UGC Ads: Script, Selfie Shot, Lip Sync, Captions

Weak prompt, stronger prompt

Here's the prompt most people start with:

a woman talking about my skincare product, UGC ad

You'll get a polished, studio-looking stock clip. Nothing about it says "a real person filmed this on her phone."

Here's the stronger version:

Selfie-style vertical photo of a woman in her late 20s sitting in a parked car, daylight from the side window, phone held at arm's length, slight wide-angle distortion, holding a small white moisturiser tube near her face, relaxed half-smile, no makeup polish, natural skin texture

The difference is that the second prompt describes how the shot was taken, not just who's in it. That's most of what makes ai ugc ads feel believable.

What UGC style actually means

UGC (user-generated content) ads imitate a customer or creator talking to their phone. They usually share four traits: a handheld selfie angle, an ordinary location, a casual spoken script, and burned-in captions. Take away any one and the clip starts to look like a commercial.

So the job breaks into four steps: script, selfie image, talking video, captions. You can do each by hand or let a tool chain them.

Step 1: Write a script that sounds spoken

Scripts written like ad copy sound fake the moment a voice reads them. Write for the ear.

• Hook in the first line. A problem or a reaction, not your brand name. "I kept breaking out every winter, so I tried this."
• One idea per sentence. If you can't say it in one breath, split it.
• One product claim, one benefit. Don't invent results, numbers or "clinically proven" lines you can't back up.
• Keep it short. A 15 to 25 second clip needs only a few short sentences. Read it aloud with a timer.
• End with one action. "Link's below" is enough.

If you're stuck, Mediafy Chat can help you brainstorm hooks and tighten a draft. Ask for three versions in different tones, then pick the one you'd actually say.

A fast script template

1. Hook: the problem or the surprise.
2. Proof moment: what you noticed after using it.
3. Detail: one specific thing (texture, size, smell, price tier).
4. Call to action: one short line.

Step 2: Build the selfie shot

The still image sets the whole ad. The video model will animate whatever you give it, so fix problems here.

Prompt the phone, not a camera:

• Arm's-length framing, face slightly off-centre, a little wide-angle distortion.
• Real places: kitchen counter, bathroom mirror, parked car, a messy desk.
• Soft window light rather than "cinematic lighting."
• Skin and hair left imperfect: "natural skin texture, flyaway hairs."
• The product held visibly at chest or chin height, label facing the lens.

Two mistakes show up constantly. First, over-polished faces: words like "stunning" or "perfect" push the image toward a beauty-ad look. Second, garbled product labels. Image models still struggle to render exact packaging text, so the safer path is to use a clean reference of your real product where the tool allows it, or to composite the real label afterwards. Our guide to getting readable text inside AI images covers what works and what doesn't.

Shoot vertical. Aspect Ratio for Social Media explains why 9:16 is the default for this format.

If you want the same "person" across several ads, build a consistent character rather than re-rolling a new face every time. See Keep a Consistent AI Character Across Images and Video.

Step 3: Make her talk

You have two routes.

Route A: image to video with a lip-synced voice. You animate the selfie and drive the mouth with recorded or synthetic speech. You get control over the exact words, which matters when the script has to be legal and accurate. Start frames with a clear, front-facing, unobstructed mouth animate best. Our post on picking a start frame that animates well goes through the checks.

Route B: a video model that generates speech and sound itself. Some models create audio along with the picture. Google's Vertex AI video prompt guide recommends stating audio explicitly and putting audio details in their own sentences. It describes three audio types: sound effects, ambient noise and dialogue. Its dialogue example uses a speaker tag and a colon, as in "the man in the red hat says: Where is the rabbit?", and it suggests specifying tone or accent. That guide is written for Google's Veo models, so treat the syntax as a starting point for other models, which may behave differently. More detail in Veo 3.1 Native Audio: How to Prompt Sound and Speech.

A prompt in that style might read:

Handheld selfie video, the woman in the car looks into the lens and says: I kept breaking out every winter, so I tried this. Quiet car interior ambience, faint traffic outside. Slight natural head movement, she lifts the tube toward the camera.

Keep the spoken line short. Long lines are where mouth movement and timing tend to drift, and results vary run to run, so plan to generate a few takes and keep the best one. Models such as Kling and Veo variants are on offer in Mediafy Studio's Video mode, and How to Choose an AI Video Model by Job helps you pick by task. For natural movement prompts, see Kling AI Video: How to Prompt for Realistic Motion.

Where lip sync still breaks

• Hands crossing the mouth or the product covering the lips.
• Fast head turns and profile views.
• Products held close to the face, which can warp.
• Hands themselves: extra or merged fingers still happen.
• Multiple speakers in one frame.

The workaround is boring but effective: use one speaker, face the lens, keep the hands low, and cut around any bad second.

Step 4: Add captions

Most people watch ads with the sound off, so captions do real work. Some practical rules:

• Two lines at most, a few words at a time.
• Keep them clear of the bottom area where platform buttons and descriptions sit.
• Check every word against your script. Auto-generated subtitles misspell product names.
• Use one font and one high-contrast colour. Highlighting a key word is fine, a rainbow isn't.

Do the chain in one go

Doing these four steps in separate tools means lots of downloading and re-uploading. In Mediafy Studio, the UGC Factory takes a character and a product, writes a short script, makes a selfie-style image, lip-syncs it into a talking video and adds subtitles. If you'd rather control each step, a Flow can chain image, video and voice-over on one canvas (Flows in Mediafy Studio). Either way, the credit cost shows on the Generate button before you run, and failed generations are refunded automatically.

Be honest in the ad

This part isn't optional. A UGC ad that implies a real customer used your product, when none did, is a problem.

The US FTC's final rule on fake reviews and testimonials covers reviews and testimonials that misrepresent who wrote them, including AI-generated fake reviews attributed to people who don't exist. The FTC can seek civil penalties against knowing violators. In plain terms: an AI character shouldn't pretend to be a real customer sharing a real experience. Present the character as a spokesperson or actor, and only say what's true about the product. If your business operates elsewhere, check local rules, and consider legal advice for health or beauty claims.

Platforms have their own rules. TikTok's announcement on labelling AI-generated content says creators must label AI content with realistic images, audio or video, and lets them use its AI-generated label or another disclosure such as a sticker or caption. That post is from 2023, so check the current ad policy for each platform before you publish.

A quick checklist before you post

1. Script read aloud: under 25 seconds, one claim, one call to action.
2. Selfie image: phone-style framing, readable product, natural skin.
3. Video: mouth visible, hands low, best take chosen from several.
4. Captions: checked word by word.
5. Disclosure: AI label or caption added, no fake customer story.

Want to see how the full chain works? You can sign in with Google and try the UGC Factory yourself. If you want a structured path through AI video production, the AI Creator Lab course is there too.

Sources

• Federal Trade Commission, "Federal Trade Commission Announces Final Rule Banning Fake Reviews and Testimonials": https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials
• Google Cloud, Vertex AI documentation, video generation prompt guide: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide
• TikTok Newsroom, "New labels for disclosing AI-generated content" (September 2023): https://newsroom.tiktok.com/new-labels-for-disclosing-ai-generated-content?lang=en