This guide is the prompt structure we use when testing models in the studio. It is model-agnostic: the same skeleton works for Veo 3.1, Sora 2, Kling, Seedance and Wan 2.5, with small adjustments noted at the end.
Why most image-to-video prompts fail
The typical first prompt describes everything at once: a subject, a location, a mood, a style, and a camera move. The model has already been given the subject, the location and the style by your upload, so two-thirds of that prompt is redundant. Worse, the parts that are not redundant often contradict the image — asking for a sunset when the photo is a grey studio shot.
The second failure is motion ambiguity. "Make her smile" does not say how much, how fast, or whether the camera stays still. The model has to guess, and a guess on a moving subject usually lands on something uncanny: too much movement, or a face that drifts.
The five-part prompt skeleton
A prompt that holds together answers five questions in a fixed order. Short sentences beat poetic ones; the model is parsing intent, not grading prose.
| Part | Question it answers | Example |
|---|---|---|
| Subject action | Who or what moves, and how? | The woman turns her head slowly toward the window |
| Motion scale | How big is the change? | Small, subtle movement — no walking, no camera push |
| Camera | What does the camera do? | Static tripod shot, no zoom |
| Environment | What is already moving in the frame? | Curtains drift, dust in the light beam |
| Look | What should the footage feel like? | Soft daylight, shallow depth of field, natural grain |
Two lines of that table are doing most of the work. Motion scale and camera are the two levers that decide whether a clip looks shot or generated — and they are the two that beginners leave out.
Try it now
Open the image-to-video studio, upload one still and paste the prompt you build below. The credit cost is shown before the run, so a failed attempt costs one run, not a subscription.
Start from an example
Camera language the models actually understand
Video models are not cinematographers, but they do respond to a small vocabulary reliably. These are the terms that survived our tests without producing warped frames:
- Static tripod shot — the safest choice for portraits and product shots. Nothing moves but the subject.
- Slow push in — a gentle dolly toward the subject. Use it to add tension without redrawing the background.
- Slow pull back — reveals more of the scene. Works best when the upload has space around the subject.
- Pan left / pan right — moves the frame sideways. On a still it can expose invented detail at the edges, so keep the pan short.
- Handheld drift — a slight wobble that makes a perfect render feel like real footage. A little goes a long way.
- Orbit — circles the subject. Only reliable on product shots with a clean background.
Avoid camera language that implies cuts, zooms across rooms, or a change of lens. A single short clip is one continuous take; asking for a montage inside eight seconds gives you mush.
Motion strength: the dial nobody mentions
Every image-to-video model has an implicit motion-strength slider, and you control it with wording rather than a number. Words like subtle, slight, gently and barely nudge it down. Verbs like sprints, spins, whips and explodes push it up. The mistake is treating high motion as "more interesting". High motion is where faces melt and hands reproduce.
For portraits, aim low. A blink, a small turn, a breath, a strand of hair moving is enough for a believable clip. For product shots, medium: a slow rotate or a light sweep. For landscapes and abstract footage, high motion is safe because there are no faces or hands to get wrong.
Motion is a budget. Spend it on the one thing you want the viewer to notice, and keep everything else still.
Five failure modes and the prompt fix
| What you see | Why it happens | Prompt fix |
|---|---|---|
| Face warps or ages | Motion scale too high for a portrait crop | Add "subtle movement", "no head turn", keep camera static |
| Background breathes or slides | No camera instruction, model invents drift | Name the camera: "static tripod shot, no zoom" |
| Clip ends mid-action | Action needs more time than the clip length | Pick one small action that completes inside the clip |
| Text and logos morph | Generative models re-draw glyphs | Add "preserve all text and logos exactly" and keep motion low |
| Colour shifts versus the photo | Style words fight the upload | Drop style words entirely — the image already carries the look |
Two of those five fixes are deletions, not additions. When a clip drifts away from the source image, the first thing to remove is any sentence describing appearance. Colour, lighting, lens and mood are already fixed by the pixels you uploaded.
Model-specific notes
| Model | Strong at | Prompt quirk |
|---|---|---|
| Veo 3.1 | Photoreal motion, natural light | Rewards explicit camera language and short sentences |
| Sora 2 | Expressive subjects, longer beats | Handles a described action sequence better than a single verb |
| Kling | Dynamic body movement | Wants a clear subject action before any stylistic note |
| Seedance | Detailed motion on a proven shot | Keep prompts short; long prompts get partially ignored |
| Wan 2.5 | Image fidelity to the upload | Very responsive to "preserve" wording for text and faces |
The practical consequence: do not reuse one prompt across five models and compare outputs. Reuse the skeleton, then tune the two lines each model cares about. Otherwise you are measuring prompt-model mismatch, not model quality.
A checklist you can run in thirty seconds
- 1Does the prompt name one subject action and nothing else?
- 2Does it state the motion scale (subtle, moderate, strong)?
- 3Does it name the camera move, even if the answer is "static"?
- 4Have you deleted every sentence describing appearance?
- 5Does the described action finish inside the clip length?
- 6Would the same prompt still make sense on a different photo? If yes, rewrite it.
Run that list before you press generate and most of the retry budget disappears. The remaining failures are usually model choice, not prompt structure — which is a much cheaper problem to solve.
Compare models
Veo 3.1, Sora 2, Kling, Seedance and Wan 2.5 all run in the same studio. Each model page lists the modes it powers, its credit cost and what it is best at.
Frequently asked questions
- What makes a good image-to-video prompt?
- One subject action, an explicit motion scale, an explicit camera instruction, and nothing that re-describes the image. The upload already carries the subject, colour, lighting and style, so the prompt only needs to describe change.
- Should I mention the style in an image-to-video prompt?
- Usually no. Style words compete with the uploaded frame and cause colour shifts. If the clip drifts away from your photo, deleting the appearance sentences is the fastest fix.
- Why do faces warp in image-to-video?
- Almost always because the motion scale is too high for a portrait crop. Add wording like "subtle movement" and "static tripod shot", and keep the action small: a blink, a breath, a slight turn.
- How long are the clips?
- The models in the studio return short clips, typically five to ten seconds depending on the model. Write one action that completes inside that window instead of a sequence that gets cut off.
- How much does one image-to-video run cost?
- Video runs start at 60 credits for a standard ten-second Sora 2 clip, and Kling's v2-1 standard tier is 90. Veo 3.1 fast and Wan 2.5 image-to-video are 300, Seedance is 500, an eight-second 480p Wan Animate run is 600, and Kling's v2-1 master tier reaches 4,000. The studio shows the exact cost before each job, and failed runs on the provider side are refunded in credits.
Turn a still into a clip
Open the studio, upload an image, describe the motion in one line and pick a video model. The credit cost is shown before the job runs.