Why Your AI Video Camera Won't Stay Still, and How to Rewrite the Prompt
Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05
In September 2026 we reviewed a 15-shot AI drama episode: the camera drifted in at least 12 shots, and 11 of those prompts said to keep it still.
So "static camera" in a prompt was doing almost nothing. And at 12 out of 15, this was not bad luck. Re-rolling fixes random glitches, like an extra hand that shows up once. It does not fix a mistake that happens nearly every time. That kind of mistake lives in the prompt.
What the camera actually did
| What happened | Count or size |
|---|---|
| Camera moved on its own | At least 12 of 15 shots (80%) |
| Of those, prompts that explicitly said the camera stays still | 11 |
| Asked for a slow push-in | Pushed in roughly twice as far as asked |
| Started at a knee-up medium shot | Ended on a face close-up; the face grew about 2.5x |
The last one hurts drama the most. Two people talk in a medium shot, and then the next shot is suddenly a giant face. The audience reads that as a story beat. It is just the model pushing in on its own.
Pretty prompts get misread
A related failure: prompts written like a novel. One of ours said, roughly, that the camera tilts up along the folds of her shirt. The video showed the character stuffing a sheet of paper into her shirt. The model did not read "along the folds" as a camera path. It read it as something to act out.
Here is an illustration. This is an example written for this article, not one of our original prompts.
Before:
Late afternoon, rain on the glass. The camera glides slowly up his jacket as light plays across his face; he stares at the phone in his hand, lost in thought, camera stays still.
It asks for a glide and for stillness. "Light plays" and "lost in thought" leave the model to invent. And the stillness request comes last, after a camera move has already been described.
After:
Locked camera on a tripod, no push-in, no pull-out, no pan; nothing new enters the frame. Medium shot, a man seated at a desk. 0-2 s: he looks down at the phone in his hand. 2-3 s: he sets the phone face-down on the desk. Final frame: phone face-down, his hand resting beside it. Avoid: double exposure, a second phone, waving, looking into the lens, an extra wristwatch.
Five things changed.
Five rewrite rules
1. Lock the camera in the first sentence. Spell it out: no push-in, no pull-out, no lateral move, nothing new enters. A stillness note at the end loses to whatever motion was described before it.
2. One action per shot, in time order. Writing "0-2 s ... 2-3.5 s ..." is about sequence, not exact seconds (more on that below). We reviewed a script line that packed in three or four actions; the clip showed only the first.
3. Name what you don't want. Our list was double exposure, a second sheet of paper, waving, looking into the lens, an extra wristwatch. Build yours from the defects you actually see. Whether to phrase it as a negative depends on the model; see the next section.
4. Cut still shots to about 3 seconds. This is the one people resist. The longer a shot runs, the more the model invents to fill it. Five seconds of someone standing and listening is five seconds for the camera to wander and the hands to start waving. If your tool cannot generate anything shorter than 4 or 5 seconds, generate the shortest length and trim in the edit.
5. Describe the final frame. The model needs to know where the action ends.
Three more traps we hit that throw prompts off:
- Use "the woman," "the man," or the character's name. On the demo drama we once kept character names out of prompts; with no subject left, we got sentences like "the figure lifts the figure."
- Do not swap nouns or add words with two meanings. A slight bow that got "salute" added came out as a raised-hand salute. A letter that got renamed "newspaper clipping" became the wrong prop.
- Do not move the camera in every shot. Alternate push, pull and locked shots; dialogue shots can be almost still.
The vendor docs disagree with each other
Generic prompt formulas skip this. On two points, negative phrasing and timestamps, official guides from different vendors point in opposite directions.
| Vendor guide (checked 2026-10-05) | Locked camera | Negative phrasing | Exact timestamps |
|---|---|---|---|
| BytePlus Seedance 2.0 prompt guide | "Fixed shot" is a standard term; one camera move per shot, since mixing push, pull, pan and move adds instability | Recommends constraint lines such as no subtitles, no logo, no watermark | Precise timing like "0-3 s" is unstable; do not over-constrain per-shot duration |
| Runway Gen-4 prompting guide | Use positive terms such as locked camera | Not supported; may give unpredictable or opposite results. Write "locked camera," not "no camera movement" | Treat each 5-10 s clip as one scene |
| Google Gemini Omni Flash guide | Not covered in the sections we read | No separate field; write it in the prompt | Timing can be written in natural language or as bracketed time ranges |
| Google Veo | Camera terms in the docs include dolly shot and tracking | The Vertex SDK exposes a negativePrompt field | Not covered in the sections we read |
What we take from this:
- On Runway, an "avoid" list can backfire. Turn every negative into a positive description: "camera locked on a tripod," "the man stays seated."
- Use timestamps for order, not precision. BytePlus says exact seconds are unstable; Google's Omni guide teaches time ranges. We keep "0-2 s, 2-3 s" because it forces one or two actions per shot. We do not expect the model to switch on the dot.
Limits: when these rules don't help
- The sample is small. The numbers come from one 15-shot episode with one set of settings. Try the rules; do not quote the percentages as a general rate.
- A rewrite lowers the odds. It does not guarantee anything. Some shots will still drift. That is when a re-roll makes sense.
- Not for shots that should move. Chases, reveals and transitions need camera motion. The rule is one move per shot, not zero moves everywhere.
- Your tool may not render 3 seconds. Many generators have a minimum length. On TWno1's generate page, clips run 4-30 seconds. For a 3-second still shot, render the minimum and trim.
Even after the rewrite, a few shots will keep drifting. In TWno1's Studio you can regenerate just those, and the shots that already hold still stay as they are. One catch: a drifting camera is usually not a failed generation. The clip rendered and is simply unusable, so the automatic refund for failures does not apply. Fix the prompt first, then regenerate the stragglers.
Related reading: Seedance Prompt Examples by Use: Vertical Drama, Hooks, Motion Reference, Transitions, Product Clips; Shot Sizes in AI Short Drama: Rules From a 244-Frame Storyboard With 4 Wide Shots; How Long Does a Mandarin Line Take? A Characters-to-Seconds Table for Short Drama.
FAQ
Why does the AI video camera move when the prompt says it should stay still?
Usually the stillness instruction sits at the end of a prompt that already describes motion, or the clip is long enough that the model adds movement to fill time. In a 15-shot AI drama episode we reviewed, the camera moved in at least 12 shots, 11 of which said it should not. Our rewrite puts a full lock sentence first, gives one action per shot, and trims still shots to about 3 seconds.
Can AI video prompts use negative phrasing like "no camera movement"?
It depends on the model. Runway's official Gen-4 guide says negative phrasing is not supported and may produce the opposite result, so write "locked camera" instead of "no camera movement." BytePlus's Seedance 2.0 guide, by contrast, recommends constraint lines such as no subtitles or no logo. Both checked 2026-10-05.
Should AI video prompts include timestamps like 0-2 s?
You can, but treat timestamps as an order of events, not exact cut points. BytePlus's Seedance 2.0 guide says precise timing such as "0-3 s" is unstable, while Google's Omni Flash guide allows time ranges. The real benefit of timestamps is that they force you to give each shot only one or two actions.
How long should a still AI video shot be?
About 3 seconds. The longer an AI video shot runs, the more likely the model is to add walking, waving or a camera push to fill the time. If your generator's minimum is 4 or 5 seconds, render the minimum and trim to about 3 seconds in the edit.
Will re-rolling fix an AI video camera that keeps moving?
Re-rolling fixes random errors, not systematic ones. When the camera moved in 12 of 15 shots in the episode we reviewed, the cause was how the prompts were written, so a new seed usually drifts as well. Rewrite the prompt first, then re-roll the few shots that still move.
References
- BytePlus ModelArk: Seedance 2.0 series prompt guide (checked 2026-10-05)
- Runway: Gen-4 Video Prompting Guide (checked 2026-10-05)
- Google Gemini API: Omni Flash (checked 2026-10-05)
- Google Gemini API: Veo video generation (checked 2026-10-05)