How to make a natural AI UGC video: reference, voice, and review

Natural is a production constraint, not a style word

A generated face can be convincing for one still frame and fail as soon as the person moves or speaks. The light stops matching the room. A hand changes shape. A voice sounds like a polished commercial even though the shot is framed as a casual phone video. Adding the word ‘natural’ to a prompt does little unless the brief says what the camera, movement, and sound should actually do.

Start by deciding which part of the ad must be credible. If the product claim depends on real use, record the product action and preserve that footage. Use synthetic performance to help with structure or presentation, not to manufacture a customer experience. A fictional presenter should never say that they personally bought or used an item when that would make viewers think a real customer did.

This guide combines lessons from our research toolkit with the current VibesUGC product flow. They are different systems. The audio findings below came from a separate Gemini Omni experiment, while VibesUGC Copy Video follows a motion reference and currently produces silent video.

Pick the reference before writing the line

Look for camera height, distance, light direction, background, and the amount of movement you need. A steady phone at chest height in a normal room gives a generator fewer reasons to invent a dramatic camera move. Watch the first second at full size. If the opening is black, blurred, or halfway through a transition, trim it or choose another clip.

VibesUGC takes a 3–30 second reference clip and a character portrait in Copy Video. The browser captures a scene frame near the beginning, and the motion job follows the reference timing. That makes the opening important: the current flow cannot automatically inspect several candidate frames and pick the cleanest one. Bring footage you own or have permission to use and a portrait you have rights to. A public video on Pinterest or TikTok is useful for research, but public availability alone is not a license for a commercial motion reference.

1. Play the candidate clip at normal speed and write down its camera distance and movement.
2. Pause near the start. Check that the face, clothing, and setting are visible and consistent.
3. Remove opening frames that are black, blurred, covered by text, or in a cut.
4. Keep the source file and permission record with the creative brief.

Write for the visible action

An eight-second clip has room for one useful idea. Try a line that points to a real, visible action: ‘The clasp opens with one hand,’ while the camera shows the clasp. Do not ask the performer to explain every feature, deliver a testimonial, and finish with a sales pitch in the same short take. Read the line aloud at the target pace before rendering or recording.

Separate speech from movement direction. If a client supplies exact dialogue, keep the words exact and put directions such as ‘small hand movement’ outside the quote. Asking for a head tilt, smile, eyebrow raise, and hand gesture on each phrase can produce a rehearsed ad or a continuity error. Give the performer a simple social situation and a few boundaries instead.

What our audio test showed, and what it did not

In a separate Gemini Omni research run, we made two short talking clips with the same reference image, dialogue, model, duration, and resolution. The reviewer found the first voice artificial. The preferred version asked for speech that sounded casually recorded through a phone microphone, ruled out narration and voice acting, and gave less choreography to the face and hands. The dialogue transcribed correctly in both clips.

That is useful feedback, but not a controlled proof. We changed several prompt sections at once and generation itself varies. The practical lesson is to test ordinary conversational pacing and restrained motion with your own line, then listen to the result. A transcription check confirms words, not naturalness or lip sync.

Do not confuse that Gemini Omni experiment with a VibesUGC feature. VibesUGC Copy Video currently renders video with audio disabled. Plan the voice separately, record it with a performer or create it in an audio workflow you are allowed to use, and add it in editing. Keep the spoken claim tied to footage that actually supports it.

Review the export as an ad, not a demo

A completed MP4 is not necessarily a usable ad. Play it once with sound at normal speed, then pause at the opening, middle, product close-up, and final frame. Compare hands, clothing, packaging, labels, and background against the source. Look for text that appears and disappears, a gesture that does not finish, or a product detail that changes between frames.

Review captions only after the clean clip passes. Baked-in captions can cover the product or contain spelling errors; add them in the edit where you can control safe areas and timing. Put a synthetic-presenter disclosure where your campaign and jurisdiction require one. Our disclosure guide covers that separate decision.

• Reject altered product labels, missing parts, or an interface state the app cannot actually reach.
• Listen for repeated words, rushed endings, and a voice that sounds detached from the room.
• Check that the synthetic actor does not claim a real personal experience.
• Save the reference, brief, prompt or settings, output, and reviewer notes together before making a variation.

For the wider research-to-test sequence, read the reference-first playbook. If you are comparing generation tools, use the cost worksheet to count approved outputs instead of just completed renders.