Most ai video prompt examples on the internet are written for text-to-video, where the prompt has to invent everything. Yours is different: you already have two faces, and the prompt only has to decide what happens around them. That single difference is why copying a generic prompt list produces clips that ignore your photo entirely.
The template below has five slots. Fill them in order and the output becomes predictable.
What a prompt can and cannot control
A prompt steers the scene, the motion and the mood. It does not steer identity — that comes from the encoded photo, and no wording recovers detail the encoder discarded.
There is a mechanical reason for this. Image-to-video systems generate frames by denoising, with your photo acting as a guide; the strength of that guidance is a separate dial from the text, as the Stable Video Diffusion paper sets out when describing temporal layers bolted onto a still-image model. Diffusion guidance itself was formalised as a way to trade off diversity against fidelity:
"Classifier guidance is a recently introduced method to trade off mode coverage and sample fidelity in conditional diffusion models post training."
That trade-off is the whole game (arXiv:2207.12598). A prompt that demands dramatic motion effectively lowers fidelity to your photo, and the face starts to slip.
The five-slot template
Fill each slot with one short clause. Total length: 15-30 words. Longer prompts do not perform better; they just dilute.
| Slot | Question it answers | Example |
|---|---|---|
| 1. Subject | Who is in frame | "a couple in their thirties" |
| 2. Scene | Where they are | "inside a decaying chapel" |
| 3. Motion | What the camera and they do | "slow push in, she turns slightly" |
| 4. Light | Where the light comes from | "single candle from below" |
| 5. Mood | How it should feel | "quiet, unsettling, not gory" |
Slot 3 is the one people overload. One camera move plus one small body movement is the practical maximum before identity drifts.
Eight ai video prompt examples you can copy
Each of these is written to sit under the 30-word ceiling and to leave the faces alone.
- The chapel — "a couple standing in a ruined chapel, slow push in, dust in a single shaft of light, solemn and quiet"
- The porch — "two people on a wooden porch at night, static camera, one flickering bulb above them, uneasy but warm"
- The last train — "a couple seated on an empty platform bench, slow lateral drift, cold fluorescent light, waiting"
- The field — "two figures in tall grass at dusk, slight handheld sway, low sun behind them, nostalgic"
- The corridor — "a couple walking toward camera in a long corridor, steady forward move, hard overhead light, tense"
- The window — "a couple beside a rain-streaked window, almost still, grey daylight from the left, reflective"
- The photograph — "two people holding an old photo between them, tiny push in, warm lamp light, intimate"
- The fog — "a couple standing apart in fog, very slow zoom, no visible light source, distant and cold"
Note what none of them contain: camera specifications, film stock names, resolution instructions, or emotional adverbs stacked three deep. Those are text-to-video habits and they add noise here.
Three prompt habits that hurt
Naming a lens or film stock. The model has no reliable mapping from "35mm anamorphic" to a visible result, and the tokens compete with your scene description.
Stacking mood words. "Eerie, haunting, melancholic, bittersweet, cinematic" gets averaged into something generic. Pick one.
Describing your own faces. "A handsome man with a beard and a woman with long hair" is telling the model to invent a face, which is exactly what you do not want. Name the relationship or the count, not the features.
What to change when the result is wrong
| Symptom | Likely slot | Fix |
|---|---|---|
| Face drifts or morphs | Slot 3 | Remove the camera move, keep body movement |
| Result ignores the scene | Slots 2 and 4 | Make the light direction explicit |
| Looks like a different couple | Photo, not prompt | Crop tighter before rerendering |
| Feels flat and dead | Slot 5 | Replace two mood words with one physical detail |
| Too clean, artificial | Slot 4 | Add a single hard source, remove fill light |
The pattern is consistent: identity problems are photo problems, and atmosphere problems are prompt problems. Diagnosing which one you have saves credits, which matters when a render costs 8 credits and the entry pack holds 10.
Frequently asked questions
Do longer ai video prompt examples work better?
No. Past roughly 30 words, additional clauses start competing for the model's attention rather than refining the scene. Cut adjectives before you add any.
Should the prompt mention the number of people?
Yes — "a couple" or "two people" is worth including, because it stops the model from adding a third figure in the background. Do not describe their faces.
Can I reuse the same prompt on different photos?
Yes, and it is a good diagnostic. If the same prompt produces a stable clip from photo A and a drifting clip from photo B, the difference is the photo, not the wording. Most ai video prompt examples fail for exactly that reason and get blamed on the model.
Does the prompt control the length of the clip?
Usually not. Duration is a product setting rather than a prompt parameter, and most outputs land in the 4-6 second range.
Take slot 1 and slot 2 from a template you already like, then rewrite slots 3 to 5 in your own words. Run both versions and keep the one where the eyes stay put.