Face distortion in AI video isn't random. It has a small number of specific causes, and knowing them lets you predict which photos will fail before you spend anything.
Cause 1: the source photo doesn't contain enough face
The model needs pixels to reconstruct from. A 200×200 crop from a group photo, a face turned three-quarters away, a face half in shadow — each of these gives the model a reconstruction problem it can't solve, so it interpolates, and interpolation looks wrong.
The file minimum here is 256×256 pixels for the whole image, not per face. It is a file requirement, not a likeness guarantee. A crisp 400×400 portrait front-on will beat a soft 900×900 one where the face is angled.
Cause 2: the teeth and the mouth
Mouths are the hardest region in generative video and the reason so many clips look subtly wrong without anyone being able to say why.
Teeth are repetitive, high-contrast, and structurally constrained — they have to be a specific arrangement, and any departure reads as a dental anomaly. Tongues are soft tissue with no fixed shape. And the interior of a mouth is dark, which is exactly where a model has least information.
If your source photo has a big open smile, expect trouble. A closed or barely-open mouth is much safer.
Cause 3: asymmetry being "corrected"
Human faces are asymmetric. Generative models have a prior toward symmetry, because averaged training data is symmetric-ish. So the model quietly evens out a crooked smile or a lopsided eyebrow, and the result is a face that resembles you and isn't you.
This is the most unsettling failure and the hardest to describe, which is why people say "it looks like a wax version of me."
Cause 4: eyes
Specifically, the direction of gaze and the shape of the catchlight. A model that gets the iris right but the highlight wrong produces a face that reads as lifeless. Blinks compound it: a blink is a fast, high-contrast event and models often render it as a lid sliding rather than a lid closing.
Cause 5: occlusion and contact
A hand near a cheek, hair across an eye, two faces touching. Where two surfaces meet, the model has to decide which is in front, and it doesn't always decide the same way twice.
This is why "kiss" scenes fail disproportionately and why two people leaning close is harder than two people standing apart.
Cause 6: profile and extreme angles
Most training data is roughly frontal. Profile views get less support, and a face that turns past about 45 degrees often loses structural accuracy entirely.
If your photo is a side-on shot because it's flattering, it's a bad input for generation.
What to do about each
| Cause | Fix |
|---|---|
| Low resolution | Use the original file, not a screenshot of it |
| Teeth | Choose a photo with a closed mouth |
| Asymmetry | Accept mild change, or use a shorter, stiller scene |
| Eyes | Front-on photo with even light and visible catchlights |
| Occlusion | Remove hands and hair from the face in the source |
| Angles | Front-on or slight three-quarter only |
Why "just retry" doesn't fix distortion
For a paid task marked failed, the private task page offers up to two free retries. Start a retry there rather than buying another order. If retries are exhausted, status is unclear or a completed video has a serious problem, contact support@zombie-couple-video-ai.com. Support reviews the case under the refund policy; an automatic retry or refund is not promised.
- Failure handling. Up to two free retries on a failed paid task’s page; unresolved cases and serious result problems are reviewed by support under the refund policy.
- Failure handling. Up to two free retries on a failed paid task’s page; unresolved cases and serious result problems are reviewed by support under the refund policy.
Knowing which one you're looking at saves money and arguments.
The uncomfortable truth about likeness
Even a perfect render is an interpretation. A model that has seen your face once, at one angle, in one light, is reconstructing a person from a single sample. It will get the silhouette and the broad features and lose the fine grain that makes a face recognisable to someone who loves it.
That's why the ceiling for AI likeness is "clearly resembling" rather than "indistinguishable." Products that promise more are overselling, and you should read their examples with that in mind.
Systematic approach
Pick the photo for the model, not for the frame. That means: front-on, evenly lit, one face, eyes open, mouth closed, no obstructions, full resolution. It might be a worse photograph and it will be a much better input.
Try it at the generator; price at pricing.
Frequently asked questions
Why do AI faces look distorted?
Mostly insufficient input: low resolution, an angled face, an open mouth, or occlusion. A smaller part is the model's prior toward symmetry.
What photo works best?
One face, front-on or slight three-quarter, evenly lit, eyes open, mouth closed, at least 256×256 and preferably much larger.
Can distortion be fixed after the render?
Not really. The clip is the output; there's no repair pass. The fix is a better input photo and a new render.
For a paid task marked failed, the private task page offers up to two free retries. Start a retry there rather than buying another order. If retries are exhausted, status is unclear or a completed video has a serious problem, contact support@zombie-couple-video-ai.com. Support reviews the case under the refund policy; an automatic retry or refund is not promised.
For a paid task marked failed, the private task page offers up to two free retries. Start a retry there rather than buying another order. If retries are exhausted, status is unclear or a completed video has a serious problem, contact support@zombie-couple-video-ai.com. Support reviews the case under the refund policy; an automatic retry or refund is not promised.