What AI video models get right and wrong in UGC ads

What is an AI video model good at in UGC ads?

More than I expected when we started:

• A person talking to a phone camera. Front-camera selfie framing, arm's-length distance, slight handheld drift. This is the core of UGC, and the model does it convincingly when the prompt describes it in physical terms.
• Real-looking places. A bathroom vanity, a kitchen counter, the driver seat of a parked car, a desk at home. Specific, lived-in places come out far better than generic ones.
• Voice and sound in one pass. The model generates the spoken line, lip movement and ambient sound together, so there is no separate voiceover or sync step. A can cracking, a pump bottle clicking or keyboard noise make clips feel real, and you get them by asking.
• Emotion that reads at a glance. Surprise, annoyance, laughter. A big readable expression in the first second is something it does well when told to.
• Stylised formats. Claymation, cut-paper animation, chalkboard drawing and podcast-studio scenes are all within reach, which is why we can offer 16 formats without a separate pipeline for each. See the claymation format for an example.

The common thread: it is good at physical, continuous, everyday moments, which is exactly what UGC ads are made of.

Where does an AI video model fail in UGC ads?

These are the failure modes that cost us real renders, and what we changed:

The text one matters most. Asking for captions inside the scene, the way you would brief a human editor, gives unusable results, and each attempt is a full paid render. Now the model never sees a request for words on screen. Our finishing pass adds the hook as a subtitle layer in the upper third, inside the Reels and TikTok safe zone, in the same encode that upscales the video.

How do you write a prompt for a UGC-style AI video ad?

Every Addy shot prompt is written by a language model following fixed rules, and the shape is always the same. Roughly 90 to 160 words, in this order:

1. Format style first. For a talking head that is a vertical 9:16 front-camera phone video, handheld with slight shake, real setting, window light, natural skin texture, slightly imperfect framing.
2. Who exactly. Age, look, clothing. Specific beats generic: a 29-year-old in a grey hoodie at her bathroom vanity, not "a woman".
3. Where and what light. One light source from the side.
4. The actions in order, with timing. Hook with the product in frame in the first 3 seconds, the problem, the product in use, the call to action.
5. The product by its physical look and real sound. Never its logo.
6. The exact spoken line in quotes, with how it is said. Deadpan, exasperated, laughing. A human emotion, never a brand adjective.
7. The ambient sound.

For the script itself we push hard toward how people actually talk: first person, contractions, one light filler word, one specific everyday detail, one mild honest criticism. A line with a small flaw in it sounds more like a real review than a perfect one. We also ban openers we see constantly, like "stop scrolling" and "this is your sign". If you want help with hooks specifically, our free hook generator is a quick way to get openers to start from.

How do you make a longer AI ad with the same person?

This was the hardest problem. A single generation tops out at 30 seconds, and if you prompt two clips separately with the same description, you get two different people. Close, but obviously not the same person, which breaks the illusion of one creator talking.

Our fix is to anchor every scene to one image. For 45- and 60-second ads, Addy first makes a starting portrait of the person (or uses the recurring creator's photo on plans that have one), then renders three scenes that all start from that same first frame, in the same place and outfit. The three clips are then stitched with ffmpeg. Because the clips share codec, size and frame rate, a straight stream copy usually works; when it does not, we re-encode.

The scenes follow a fixed arc: hook and problem, then discovery and demo, then result, objection and call to action. Each scene's script is written to fit its length, so nobody has to talk fast to finish a sentence.

It is not perfect. Small details like jewellery or hair can drift between scenes. For most ads we recommend 20 seconds in one continuous take, which avoids the problem entirely and is our default.

Which resolution should you render AI video ads at?

Our default is to render at 480p and upscale. The model's price doubles at each step, from $0.05 per second at 480p to $0.10 at 720p and $0.20 at 1080p, at list. Our Standard quality renders at 480p, then scales to 720x1280 with a Lanczos filter and a light sharpen. That halves the cost of a 20-second ad from about $2.00 to about $1.00.

What you lose is fine detail: hair, fabric and small product text are softer if you pause. What you keep is the thing that matters for UGC, which is that it looks like phone footage. We test hooks in Standard and re-render proven winners in HD. The full cost breakdown, including render time, is in what an AI video ad really costs.

What can AI video models still not do for ads?

Our honest list as of October 2026:

• Show your real product exactly. The model renders something that matches the description. For products where the exact label or shape is the selling point, you still need real footage or a product image as the first frame.
• Show your real app. We do not fake app screens, and we would not recommend anyone does: they come out wrong, and showing a UI that does not exist is a bad idea in an ad.
• Make claims safely on its own. We stop scripts from inventing results, ratings, customer counts or studies, and from making medical or guaranteed-result claims. A number appears only if it is on the brand's own site. That is a rule on our side, not something the model knows.
• Get every render right. Some clips come back with odd hands, a flubbed word or a product that looks off. Failed jobs are refunded; bad-but-finished ones are paid for. We do not have enough volume yet to publish an honest keep rate.
• Keep files forever. Our provider keeps outputs for about a week, so we copy every finished file to our own storage immediately. If you build your own pipeline, do the same.

If you want to see where the line currently sits, our examples page shows real outputs, and the AI video limitations post covers the platform-policy side.