Why AI faces drift between clips, the regeneration counts that prove it, and the reference-anchor workflow that locks one actor across a whole batch.
HighQualityUGC Team
Editorial
We run UGC ad tests daily and publish what holds up: real credit costs, real hook rates, no vendor fluff.
AI actors drift because video models have no memory. Feed the same photo and the same prompt twice and you get two slightly different people: different jawline, warmer skin, a new haircut. This guide shows you why it happens, how bad it gets, and the reference-anchor workflow that keeps one actor identical across an entire batch of ads.
Get this right and you can run a testimonial, an unboxing, and a demo with the same believable face. Get it wrong and every clip looks like a different creator, which kills the trust UGC is supposed to buy you.
An AI actor is a generated or cloned person your video model renders from a prompt. It drifts because each generation is an independent event: the model has no persistent understanding that this is Character A, so it re-approximates the face from your description every single time.
Inside one clip, consistency holds. The model renders every frame in a single pass, so the face, lighting, and proportions stay stable from second 1 to second 15. The break happens on the next clip. That generation starts from fresh noise with no memory of the last one, and the person shifts.
That is the whole problem in one sentence. Same description, different bone structure, because nothing carried the identity forward.
The drift is not subtle, and it is measurable in one number: how many times you have to regenerate a scene before the face matches. Independent testing on a 60-second, 8-scene project put real numbers on four tools.
| Tool | Regenerations per scene | What broke |
|---|---|---|
| Runway | About 18 | Skin shifted warmer outdoors, face structure moved |
| Kling | ~80% consistent | Clothing changed shade nearly every scene |
| Seedance | 22+ on single scenes | A different person in 3 of 8 scenes |
| Pika | High on angle changes | Character shifted on medium-to-close cuts |
The average across all four tools was 15 to 20 regenerations per scene. In separate testing, up to 40% of scenes showed visible identity drift on the first pass.
That matters for one reason: every regeneration is a video render, and video renders cost money. Drift is not just an aesthetic problem, it is a line item. If you are budgeting a test, the true cost of a UGC video climbs fast when a third of your clips need 20 retries to hold one face.
A reference anchor is a single approved image of your actor that every downstream generation is forced to match. You lock the face once, then pass that exact image into every clip as the seed, instead of hoping the model re-draws the same person from words.
This is the technique that actually works in 2026. Generate and approve one static hero frame, then use that frame as the foundation for every image-to-video render. The geometry, skin tone, and hair transfer forward because the model is copying a picture, not interpreting a prompt.
HighQualityUGC builds this in as a gray-background contact sheet. Every actor, generated or cloned, gets one, and no ad render happens without it attached. The contact sheet is the memory the models do not have.
Lock one hero frame
Generate or upload a single clean shot of the actor on a neutral background. Approve it before anything else. This is your identity source of truth.
Pass it into every clip
Attach the anchor as the reference image on every scene generation, not just the first. Each clip copies the same face instead of inventing one.
Preview before you render
Check the actor on cheap storyboard images first. Catch a drift for the cost of an image, not a full video render.
Render only approved frames
Once the anchored storyboards look right, render the video. The face is already locked, so you are not paying for retries.
There are three ways to hold an identity, and they are not equal on cost or speed. Prompt-only is what most people try first, and it is why they end up regenerating 20 times.
| Method | Setup cost | Consistency | Best for |
|---|---|---|---|
| Prompt only | None | Low, heavy drift | Nothing that needs the same face twice |
| Reference image / contact sheet | One approved frame | High | UGC ad batches, fast turnaround |
| LoRA fine-tune | 20 to 30 images, hours of training | Highest | Long-form film work, big budgets |
LoRA fine-tuning gives the most consistent result, but it needs 20 to 30 training images per character and hours of compute before you render a single clip. For ad testing that is overkill. A single approved reference frame gets you most of the consistency at a fraction of the setup, which is why it is the right default for UGC volume.
Even with an anchor, two details drift first: skin tone and clothing shade. They move because lighting changes between scenes push the model warm or cool, and wardrobe is often under-specified in the prompt.
Two fixes handle most of it. First, desaturate the skin slightly in your anchor image, which discourages the model from swinging warm outdoors and cool indoors. Second, describe wardrobe explicitly and identically in every scene prompt, down to color and material, so the model has nothing to reinvent.
A consistent face with a different voice in every clip breaks the illusion just as fast. Voice drift is the identity problem most tools ignore, because they only anchor the visuals.
The fix mirrors the visual one: pin a base audio sample and pass it as reference audio on every render. Some models, including Seedance 2.0, support reference audio natively, so the actor sounds like the same person across every clip and every batch. Anchor both the face and the voice, or the ad still reads as stitched-together strangers.
Be honest with your workflow about the one thing anchoring does not fully solve: two characters in the same shot. Multi-character scenes still produce identity blurring on every major tool as of mid-2026, where the models bleed features between the two people.
The practical workaround is to generate characters separately and composite them, or to keep your UGC ads single-actor, which is how most high-performing UGC looks anyway. One believable person talking to camera outperforms a crowded scene, and it sidesteps the hardest unsolved problem in the category.
A locked actor is not only about polish. It changes what you can test. When the same face carries across every clip, the actor stops being a variable and your creative levers get clean.
You can run the same actor across ten hook variations and know the face is not what moved the numbers. That isolation is what makes hook rate testing actually mean something. If you want to see the full storyboard-first flow that anchors the actor before any render, the how it works walkthrough lays it out step by step.
AI actors drift because the models forget, not because the tech is broken. Lock one reference frame, pass it into every clip, anchor the voice the same way, and preview on cheap images before you spend on video. Do that and one believable actor carries your whole batch, instead of a different stranger showing up in every scene.