AI Video Prompts: A Practical Guide

Start with one subject, one action and one camera move, then add lighting and details that should stay consistent. Results also depend on source images, model capabilities and settings; a failed clip is not necessarily a prompt failure. Use the examples and diagnostic table below to change one variable at a time.

The six-part structure

Use these fields as a planning checklist, not a mandatory word order. Add only what the shot needs.

1. Subject — what is in frame, and what must stay unchanged.
2. Motion — what moves, and how fast.
3. Camera — the physical move, named.
4. Lighting — direction, quality, temperature.
5. Output requirements — select supported duration, aspect ratio and resolution in the generator. Numbers in a prompt do not override interface settings.
6. Constraints — what must not happen.

A prompt with all six:

A ceramic bowl of mango pomelo sago on a dark stone surface, spoon stirring slowly, camera pushes in gently, soft top-light with a warm rim, 4 seconds, 16:9, product shape unchanged, no camera shake, no objects added.

Choose camera motion by the shot’s job

Pan rotates the viewpoint from a fixed position; truck moves the viewpoint sideways. A push-in approaches the subject; zoom changes focal length. These are different visual requests, even when the subject becomes larger in both. Runway’s camera terminology guide documents these distinctions with Gen-4.5 examples; it does not establish compliance for every VidVid model.

Use the following editorial decision table to define a result you can inspect. It is a planning aid, not a tested success ranking.

Your shot needs	First request to consider	Inspect in the result	Change course when
A stable composition while fabric or steam moves	Locked-off framing; name the moving element	Compare fixed background edges near the start, middle and end	The whole scene drifts: simplify the action before adding more constraints
A reveal across a landscape	Pan from a named starting landmark to an ending landmark	Are the requested landmarks visible, in order?	The endpoint is outside the source image and must be invented
A sideways move past foreground detail	Truck left or right; keep subject motion separate	Does the foreground change position relative to the background without bending?	Geometry bends: reduce the move or use a fixed shot
Attention to one visible detail	Small push-in toward that detail	Does the detail remain identifiable and inside the delivery crop?	The source lacks enough detail: improve the input rather than ask for invented detail
A product seen from another angle	Small arc around the stationary product	Check silhouette, label and newly exposed surfaces	Only the front is documented: do not treat an invented rear view as product evidence

A product rotating under a fixed camera is subject motion. A camera moving around a stationary product is camera motion. State which one you need; asking for both makes a failed result harder to diagnose. For factual product work, use the product-video review workflow before accepting a new angle.

A motion brief before a prompt

Fill this in before generating. Example values below are untested editorial examples, not measured settings or guaranteed controls.

Purpose: draw attention to the visible embossed mark
Subject action: the object stays still
Camera action: a small, slow push-in toward the mark
Starting frame: mark visible, whole object inside the crop
Ending frame: mark larger; top and base still visible
Reject if: mark changes, object bends, or an unrequested cut occurs
Single next change: reduce camera travel; keep input and settings fixed

Select duration, resolution and aspect ratio from the chosen model’s actual controls. The camera line expresses intent through the prompt; it is not a promise of numerical trajectory control. In image-to-video, check whether blur, pose or directional cues in the input conflict with the requested motion. Runway’s image-to-video guidance discusses this for Gen-4.5 and recommends starting simply. Treat it as a diagnostic possibility, not a proven cause in another model.

Review the shot, not just the prompt

At the start, middle, end and first visible defect, record a timestamp and what changed. Separate three verdicts: camera path, subject preservation and delivery crop. A clip can execute a push-in while still failing because the label changes. Keep the input, model, prompt, selected settings, actual charged credits and output together in a run record.

For a retry, change only the camera line or only the input—not both—and record the remaining retry budget. An independent generation can vary even with the same prompt; one better result does not prove the change caused the improvement. If the needed view is undocumented, the same defect persists, or the next run exceeds the budget, switch to an approved input, a simpler shot or external editing. VidVid is not being presented here as an automatic editing or stabilization tool.

Text-to-video vs image-to-video prompts

Text-to-video has to build everything from words, so it needs more detail and benefits from a clear single beat.

Wide shot of a fog-covered valley at dawn, mist drifting slowly, camera pushes in gently, cool diffused light, no people, 5 seconds, 16:9, steady forward motion, documentary look.

Image-to-video already has the subject. The prompt should describe only what changes — motion, camera, light — and then constrain everything else.

Using the reference image, the same room, very slow push-in, morning window light unchanged, walls and window frames stay straight, nothing appears or is removed, 4 seconds, 16:9, steady forward motion.

Avoid repeating visual details unnecessarily. Describe the requested change; add visual detail when it clarifies a new element or transformation. Extra description is not, by itself, proof of what caused drift.

Reusable prompt set

PROMPT — Cinematic establishing
Wide aerial shot of a coastline at golden hour, waves moving slowly, camera drifts forward, warm low sun, 5 seconds, 16:9, no people, cinematic colour grade.

PROMPT — Macro detail
Extreme macro of woven fabric, very slow lateral drift, raking side light reveals texture, colour unchanged, 4 seconds, 1:1, steady lateral motion.

PROMPT — Character shot
Using the reference image, the same person turns slightly toward camera, subtle expression change, soft window light from the left, shallow depth of field, 4 seconds, 16:9, face unchanged, same clothing.

PROMPT — Vertical hook
Vertical frame, subject centred and fully visible in frame one, slow push-in, clean background, bright key light, 3 seconds, 9:16, readable at small size, no camera shake.

PROMPT — Product turntable
Product rotates gently on a clean white surface, soft even studio light, subtle contact shadow, packaging text remains legible, 5 seconds, 1:1, shape unchanged.

Symptom → fix

What you see	Likely cause	Fix
Subject morphs	Too much motion in one request	Try a supported shorter duration or a smaller move, one change per run; a shape constraint is not a guarantee
Face changes between shots	Separate generations, no shared reference	One canonical image, image-to-video, face unchanged
Walls or windows warp	Large camera move on geometry	Push-in or slow traverse only, add lines stay straight
Hands break	Fine detail is not preserved in this result	Frame hands out, or cut away
Everything drifts or jitters	Vague or absent camera direction	Choose one physical move, or use locked-off for a static shot
Colours shift	No lighting instruction	State direction, quality and temperature
Output looks generic	Mood words instead of specifics	Replace adjectives with concrete nouns and moves
Text in frame is gibberish	Models do not render text reliably	Add all text in post

Iteration discipline

• Change one thing per run. If you change duration, camera and lighting together, you learn nothing.
• Use a supported draft setting. Check the selected model’s options and estimated credits; 480p is not available on every model. A lower-resolution draft does not establish final-detail quality.
• Keep a prompt log. Prompt, model, settings, result, verdict. It becomes your reusable library.
• Save the winners. A prompt that worked is worth more than a new one.

See the AI video pricing guide for what each iteration costs, and the model comparison for which model suits which prompt type.

Test cards

Three short drills built on the six-part structure above. One generation each, and you change only what the drill tells you to change.

Card A — structure test. Take the worst prompt you have and rewrite it against the six-part table. Same model, same settings — then compare what the rewrite actually bought you. Output: the full H3 run for this card — file, prompt and settings — is published below.

Card B — single-variable test. Take the output that came out almost right and change nothing but its camera line. Regenerate and compare the two side by side. Output: the full H3 run for this card — file, prompt and settings — is published below.

Card C — constraint test. Take an output where the subject deforms, add keep the exact shape, and halve the duration. Output: the full H3 run for this card — file, prompt and settings — is published below.

The three MiniMax H3 runs behind these cards

Same three cards, run once each on MiniMax H3 through the official API at 768P. Files, prompts and settings are published below so you can repeat them exactly. These are bench runs on a reference image and generic prompts, not on your catalogue — read them as evidence of turnaround and frame stability, not as a verdict on your own asset.

Open this clip in a new window →

Prompt we sent (as run): A slow macro push-in on a matte ceramic object, realistic reflections, black studio background, premium commercial lighting, the object keeps its exact shape, no deformation, 5 seconds.

Open this clip in a new window →

Prompt we sent (as run): Vertical 9:16 product hook, energetic but controlled camera move, subject fully visible in the first second, bright accent colour, shallow depth of field, object shape stays fixed.

Open this clip in a new window →

Prompt we sent (as run): Extreme macro of a brushed metal surface, slow lateral camera move, reflected light crawling across the material, no distortion, commercial product photography look.

Run	Mode	Settings	Output file	Submit → ready
A	image → video	5s · 768P	1344×768, 24 fps, 5.2 s, 726 KB	196 s
B	text → video	5s · 768P · 9:16	768×1344, 24 fps, 5.2 s, 1.0 MB	117 s
C	text → video	4s · 768P · 1:1	768×768, 24 fps, 4.5 s, 341 KB	83 s

Scope: three single runs on one model. We did not measure the credit deduction inside VidVid, we did not time any other model, and we did not score output quality — that stays a human call. Check the current credit quote before submitting; these recorded runs do not establish today’s price.

How to compare another model

Keep the input and shot goal unchanged. Match duration, resolution and aspect ratio only where both models support them; record any differences so the results are not presented as a controlled comparison. Check the current quote, set an attempt limit and review every result against the same criteria. One output does not establish a general quality ranking.

Choose supported settings in the VidVid studio. No other model was timed in the sample runs on this page.

FAQ

How long should a prompt be?

Cover the subject, action, camera, lighting and essential constraints. Remove conflicting instructions before adding detail; adjective count is not a quality measure.

Must I write in English?

There is no universal guarantee across models. Start in a language you can express precisely. To compare languages, keep the input and settings unchanged and log the results.

Do prompt numbers set duration and aspect ratio?

Select supported settings in the generator and confirm the output configuration and credit quote before submission.

Can negative prompts prevent deformation?

No guarantee. Use a negative-prompt field only if supported. Written constraints cannot ensure exact labels, faces or geometry; inspect the output.

Why do identical prompts produce different clips?

Generation can be stochastic. Save the model, input, settings, prompt and seed where available. Set an attempt limit before rerunning.

How should I handle exact text?

Add essential titles, prices and captions during editing and check them manually.

Where to go next

• Open the VidVid studio and run the reusable set on your own subject.
• Task-specific prompts: product video, real estate, vertical.
• Keeping one subject stable: the consistent character guide.
• Budget the iterations: the AI video pricing guide.

Troubleshoot the output

Use the AI video artifact checklist to isolate one defect before generating again.

For a sequence with multiple shots, use the storyboard workflow to assign each clip a purpose, source asset and transition before writing individual motion prompts.

Editorial motion framework reviewed 2026-10-10. No new paid generation test was performed for this update.