You upload two photos of a couple in ordinary clothes, ask for a zombie scene, and get back a clip where the jacket is a different jacket. Nothing was asked to change it. So why did it change?
The model doesn't see a jacket
It sees pixels with statistical structure. From millions of examples it has learned that regions with this texture, this fold pattern, and this edge behaviour tend to be called "jacket" — but it has no concept of your jacket as an object with an identity to preserve.
So when it has to render frame 40, and the folds on the left shoulder are ambiguous in shadow, it makes a decision. That decision is locally plausible and globally inconsistent with the jacket it drew at frame 10.
Why the drift compounds
Each frame is generated conditioned on the previous ones. That means an error at frame 10 becomes part of the evidence for frame 11, which becomes evidence for frame 12. Small reinterpretations accumulate the way rounding errors do in a long calculation.
This is why clothing drift is worst at the end of a clip and almost invisible at the start, and why ten seconds is a very different proposition from thirty.
The four things that actually trigger it
1. Low contrast between the garment and its surroundings. A black jumper in a dim room is a soft edge the model has to guess at. Guessing produces change.
2. Pattern and detail density. Stripes, plaid, logos, fine knit. Detail the model can't hold at 720p gets simplified into something nearby.
3. Occlusion and motion. The moment a body turns and a sleeve crosses the torso, the model has to invent what's behind. What it invents rarely matches.
4. Loose fabric. A flowing coat or an oversized shirt gives the model more degrees of freedom than a fitted one, and freedom is where drift lives.
What the model is optimising for instead
Here's the part that makes the behaviour feel less like a bug: the model is trying to produce a convincing video, not a faithful document. Motion coherence, plausible lighting, and frame-to-frame smoothness all score well. Preserving a specific garment's pattern does not directly score at all, unless the model was trained to care.
So it will happily redraw a lapel to make a shoulder rotation look smoother.
What actually helps
Simplify the wardrobe in the source photo. Solid colours, fitted shapes, no logos. This isn't about taste — it's about reducing the number of decisions the model has to make.
Improve the light. Even, front-ish light on both people means fewer ambiguous regions, which means fewer guesses.
Choose scenes that don't require full-body movement. A ten-second scene where two people sit, lean, and look at each other asks the model to preserve far less than one where they walk across a room.
Accept the trade. Some drift is the price of motion. A completely stable jacket in a completely static shot is just a photo.
What doesn't help
- Repeating "same jacket" in the prompt. A text constraint is a weak signal compared to the visual evidence in the input frame, and it doesn't survive contact with a turning body.
- Higher resolution alone. It sharpens the drift rather than preventing it.
- Asking for a longer clip and cutting the good part. By the time it looks wrong, it's wrong for the remainder.
Is this a problem for a couple video?
Mostly no, and it's worth being precise about why. The audience is one person who knows you, watching for the faces and the joke. A jumper that gained a collar is far less noticeable than a face that lost its owner.
Where it does matter: a couple video built around a specific costume — matching Halloween outfits, a wedding outfit, a particular band t-shirt that's the whole point. If the outfit is the message, generation will betray you, and practical makeup plus a camera is the honest answer.
The structural fix
A locked scene. If the camera doesn't move, the bodies don't turn, and the light is even, this may reduce clothing drift, but it does not eliminate it or guarantee wardrobe continuity.
That's why a ten-second, fixed-camera, two-shot scene is the format this product uses. Not because it's modest, but because it's the configuration where the failure mode is smallest.
Try it at the generator; price at pricing.
Frequently asked questions
Why did the AI change my clothes?
The model infers a garment from pixels rather than preserving an object. Ambiguous folds, patterns, low light, and body rotation all force it to guess, and guesses differ between frames.
Can I stop it changing?
Reduce the causes: solid colours, even light, fitted clothing, limited movement. You can't eliminate it entirely, because some inference is always happening.
Does it affect faces too?
Faces drift less because they get more attention in training and more conditioning from the reference. Hands and clothing are the reliable weak points.
Does a longer clip make it worse?
Yes. Errors become evidence for later frames, so drift accumulates over duration.