How to tell an AI video model what must not change: preserve clauses for image-to-video
A face that melts, a label that turns to nonsense, a logo that drifts. Write a short keep-list into your image-to-video prompt, and stop adding motion words.
You animate a product shot and the label turns into soup by the third second. You animate a portrait and by the end it is a slightly different person. The usual reaction is to rewrite the motion: slower, smoother, more cinematic. That rarely helps, because the motion was never the problem. Nothing in the prompt said what had to stay.
This post is about that missing half: the preserve clause, a short sentence that names what the model must keep from the start frame. How to pick motion words is covered in the image-to-video motion prompts post, and the whole workflow from still to clip in how to turn a still image into a video. Here we only deal with the keep-list.
Why more motion words make it worse
Every verb in a video prompt asks the model to redraw something: an orbit redraws the sides of an object it has never seen, a gust redraws hair and fabric, a light change redraws every surface. Each redraw is a chance for detail to drift. Stack six motion phrases and you have asked for six kinds of change and given no reason to protect anything.
A preserve clause works the other way: it names the fixed points, so the change goes somewhere else. One motion and a clear keep-list holds together better than five motions and none.
What belongs on a keep-list
- Identity: the face, the hairstyle, the clothing. Same face, same outfit throughout.
- Printed things: a label, a logo on a mug, a sign. Name the object that carries the print and ask for it to stay sharp.
- Shape: a bottle that must not stretch, a building with straight walls.
- Composition: the subject's place in the frame, the horizon line. A static camera is the strongest composition keeper there is.
- Background and look: the room or sky, the colour grade, the lighting direction, the style of the still.
Keep the list to two or three items, the things you would reject the clip for. A list of ten gets treated like any long prompt: a few words count and the rest are dropped at random.
The shape of a good prompt: move one thing, keep the rest
Write the motion first, in one sentence, then the preserve clause in a second sentence that starts with a clear verb: keep, stays, unchanged. Concrete nouns beat general ones. "Keep everything the same" is too vague to act on; "the label and the bottle shape stay unchanged" is something a model can check against the start frame.
Slow push-in toward the bottle, condensation drops slide down the glass. The label stays sharp and facing the camera, the bottle shape and the background unchanged.Static camera. She blinks, smiles slightly and turns her head a little toward the window. Same face, same hairstyle and same green sweater throughout, the room behind her stays still.Static camera. Steam rises from the cup and drifts to the left. The logo on the mug stays readable and in place, the table and the window light unchanged.Choose motions that do not touch what you keep
A preserve clause is a request, and a request loses when the motion contradicts it. "Orbit around the bottle" and "the label stays facing the camera" cannot both be true. "She turns to look over her shoulder" and "same face throughout" are at odds, because the model has to invent the side of the head it never saw. So plan the motion around the keep-list: if the label matters, move the camera in, not around; if the face matters, move the hair, the light or the background, and keep the head roughly where it is.
Where the keep-list starts: the frame itself
The model can only keep what it can see. A label that is small, angled or slightly blurred in the start frame will not get sharper in motion. Make the still right first: the important print large, sharp and facing the camera. To fix a still before animating, use an edit prompt with its own keep clause, as in the post on writing image edit prompts.
Then animate. In the studio, the Animate action on any image in your feed or library opens the video composer with that image attached as the start frame, or you can attach your own picture. Some models also show a Negative prompt field under Advanced. It is the right place for what should not appear (flicker, warping, extra fingers), as the negative prompts post explains; what must stay goes in the main prompt, in positive words.
Test cheaply, then commit
Drift shows up early, so a short draft tells you most of what you need. Render a 5-second test on Motion Lite, from €2.00 a clip, at a low resolution, and watch the things on your keep-list frame by frame. If they hold, move to the model and length you want for the final. Models differ in how faithfully they hold a start frame; the models page lists them, and the post on choosing an AI video model says which suits what. The price of each run shows on the Generate button before you press it.
What a preserve clause cannot do
- It is not a lock. It raises the odds that a detail survives; it does not guarantee it. Small text on a moving object is the weakest point of every video model today, and a long line of small print will often blur.
- Long clips drift more. Clips run up to 30 seconds on some models, but every extra second is more frames to redraw. For a label or a face that must be exact, a short clip is safer.
- It cannot recover what is hidden. If the motion reveals the back of an object or the side of a face, the model invents it.
- It will not repair the frame. Warped letters in the still stay warped in the clip, often more so.
If a clip still drifts, cut a motion before you add a keep item. Fewer changes, named anchors, a short clip: that is the whole method.
Questions
What is a preserve clause in a video prompt?
A short sentence after the motion that names what must stay as it is in the start frame, for example: the label stays sharp and facing the camera, the background unchanged.
Should I put what must stay in the Negative prompt field?
No. The negative field is for things that should not appear, such as flicker or warping. What must stay belongs in the main prompt, in positive words like keep, same and unchanged.
Why does the face change in my image-to-video clip?
Usually because the motion asks the model to redraw it: a head turn, an orbit or a long clip. Use small motions, a static camera, a short length, and add same face throughout to the prompt.
Can I keep a product label perfectly readable?
Often, not always. Start from a large, sharp label facing the camera, move the camera in rather than around, keep the clip short, and check a cheap draft before the final render.
