What gives AI UGC away

For my UGC Campaign I made three creator ads for a real skincare product, Dr.Jart+ Cicapair, a green cream that calms the look of redness. Everything in them is generated: the women, their rooms, their voices. It took more than 20 video generations to get three ads I was happy to show.

Inside a single video, the faces held up well. Smaller things gave the ads away, the kind a viewer might not be able to name but still notices. These are the ones I caught, and what each one changed about how I work.

The label is only as sharp as its pixels

In a product ad, people look at the product. At 720p, a jar held at arm’s length has its brand name about 12 pixels tall, and the small print is 4 or 5. At that size the letters wobble from frame to frame, and wobbly letters give a video away.

Asking for a sharper label doesn’t fix it. In one video edit I asked for a crystal-clear label and got the same label back, if anything a touch softer. The label needs more pixels: a higher resolution, and a script where she holds the jar close to the camera for a moment, label facing it. Words can’t add detail the frame has no room for.

In still images I don’t let the model write the label on its own. Left alone, it invents letters that look like the brand from a distance and turn into nonsense up close.

Cream doesn’t disappear on its own

In one take, she dabbed cream on her chin, left it there while she worked on her cheeks, and about twelve seconds later the dab faded by itself as she turned her head. Real cream only goes away under fingers.

So every bit of product now gets blended on camera, one area at a time, and nothing sits on her face “for later”. The amount matters as much. “A generous dollop” gave her a white, chalky face. A pea-sized amount, dabbed and blended, looked like skincare.

The model dresses your character in somebody’s brand

I described one creator’s sweater as “a cream knit sweater”. The model gave her a cream cable-knit with a small navy polo-player logo on the chest. A cream cable-knit reads as one famous brand’s sweater, so the model finished the thought for me.

I tried to remove it in post twice, by covering the logo with knit from elsewhere on the sweater. Both times the patch flickered. I saw it before the measurements confirmed it: the spot where the logo had been changed more from frame to frame than the plain knit around it. In the end a video edit model removed it. That took three tools and close to $30, and the picture came back slightly softer than the original.

Preventing it costs nothing. Every garment is now written as plain, with no logo, embroidery or text, and I zoom in on the chest and sleeves of the first generation before anything else gets built on top of it.

Brand names come out wrong

“Dr. Jart Cicapair” came out as “Dr. Jant’s Cicapack”. A pronunciation note next to the line did nothing. Writing the name the way it’s said, inside the line she speaks, fixed “Doctor Jart” right away. The product name took more work.

When a take is good and one word is wrong, I don’t regenerate the take. I re-voice that one word in her own voice and fit it to her lips. Regenerating means rolling the dice again on everything that was already right. Every talking video also gets transcribed and checked against the script before anyone sees it. The transcript flags the suspect words; my ear makes the call.

The model fills every gap you leave

Anything the prompt doesn’t pin down is a risk. When one creator stood up to leave, the camera saw a corner of her room that nobody had described, and that’s where a magazine with garbled text appeared. When a script had her glance at her phone screen, a second phone on a stand showed up in the shot. Describing the camera itself, instead of writing “a phone”, made the extra phone go away.

Faces drift when you ask for more detail than you gave

Faces slipped a few times, all of them in the campaign stills. A brand-new group photo made everyone younger, with the wrong hair, even with their character sheets attached. In one close-up I measured why: the model drew a creator’s face more than three times bigger than in her character sheet, the set of photos of her from every angle. It had to invent the missing detail, and what it invented was a generic face: younger, with a fringe she doesn’t have.

Order fixed it: saying who each person is at the very top of the prompt, before the product and the scene, kept all three women right on the first try.

Stills have one more giveaway: mirrors. If there’s a mirror in the room, it has to reflect something, and a missing reflection gives the image away. I keep mirrors in and say exactly what they show.

The checks

Every failure in this project turned into a check I now run on the first generation, before anything gets built on top of it: the label at full size next to the real packshot, the chest and sleeves, the transcript against the script, and every mirror.