Why Faces Drift in Generated Video and How to Reduce It
The mechanics of error accumulation between frames, the factors that amplify it, and concrete measures — from source selection to how you phrase the motion.
"The face drifts" is the commonest complaint about video generation. It shows up in different ways: features shift slightly between frames, eyes slide, the jawline breathes, and by the end of the clip the person has become someone else. All these symptoms share one cause.
Where drift comes from
A video model builds frame after frame, relying on the previous state. Each frame is the result of a computation, and every computation deviates slightly from the ideal. Individually the deviation is invisible, but the next frame is built from the already-shifted state, and the errors compound.
It is the same principle as copying a copy: one pass is barely visible, the twentieth clearly is. That is why drift always grows towards the end of a clip rather than being spread evenly.
What amplifies drift
- Duration. A direct relationship: more frames, more accumulation.
- Speed of motion. Fast movement demands larger changes between neighbouring frames, and therefore larger deviations.
- Complexity of motion. Several simultaneous actions force the model to divide its attention, and the face gets less of it.
- Head turns. They require showing what the source never had: the model invents the hidden half, and invents it slightly differently each frame.
- A weak source. A small or blurred face gives an imprecise anchor, and drift starts immediately.
- A busy background. It consumes capacity that would otherwise go to the face.
What reduces drift
- A shorter clip. The most effective measure: five seconds instead of ten halves the accumulation.
- Slower motion. The words "slowly", "smoothly", "barely" directly reduce the size of change between frames.
- One action instead of several.
- A large, sharp source with the face in focus.
- A simple background that does not compete for the model's attention.
- Avoiding large turns: a tilt rather than a rotation.
Drift versus flicker
Two similar defects are worth distinguishing. Drift is a slow, one-directional change: the face gradually becomes someone else. Flicker is fast oscillation: a detail jitters back and forth between frames.
Drift is treated with duration and tempo. Flicker is usually tied to fine detail in the source — fabric patterns, grain, a busy background — and is treated by removing or blurring those.
Why you cannot simply "lock the face"
The obvious wish is to ask the model not to change the face at all. But a video with no facial change is a static picture: expression is that change. The task is not to keep the face unchanged but to change it only where intended.
Hence a practical consequence: the more precisely you state what should move, the less the model alters everything else. A vague request gives a vague result, literally.
When drift does not matter
For atmospheric clips where the face is not the focus — a wide shot, a silhouette, environmental motion — mild drift goes unnoticed. Strict requirements apply only where the viewer studies the face: close-ups, portrait video, avatars.
How to check the result
Do not judge the clip at playback speed — motion masks drift. Compare the first and last frame side by side: if it is one person, the clip passes. If they are different people, shorten the duration and slow the motion.
Frequently asked
Can I ask the model to keep the face unchanged?
A video with no facial change is a static image — expression is the change. What works is stating precisely what should move; then everything else changes less.
How is drift different from flicker?
Drift is a slow one-directional change towards the end of a clip, treated by duration and tempo. Flicker is fast oscillation of fine detail, treated by simplifying the background and source.
What reduces drift the most?
Short duration and slow motion. But before either — a good source: an imprecise anchor starts the drift from the first frame.
- #артефакты
- #видео
- #качество