Orelon logoOrelon
料金

AI Video Prompt Engineering: How to Direct Motion in Every Shot

2026年11月6日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

Learn how to write AI video prompts that direct motion, camera movement, and timing shot by shot, with worked examples and a repeatable review loop.

Most AI video clips fail for an unglamorous reason: the prompt described a photograph instead of a movement. A person was named, a mood was implied, a style was requested, and nothing was said about what should change between the first frame and the last. The model, with no instruction to follow, invents something. Usually it invents drift.

This guide treats a prompt as a short directing note: who is in frame, what they do, where the camera goes, how fast, and what the light is doing while all of it happens. The examples lean toward vertical short-form work, including a dating-profile clip, because that format punishes vague motion faster than almost any other.

Why motion is the missing instruction

A still-image prompt asks a model to resolve a scene. A video prompt asks it to resolve a scene and then change it. Those are different tasks, and the second one needs explicit instructions about change.

When a prompt contains no motion language, the model still has to produce movement, so it falls back on whatever it has learned most often: slow floating drift, a gentle zoom that goes nowhere, fabric and hair that ripple without wind, backgrounds that breathe. The result looks synthetic even when the render is technically clean, because nothing in the frame is motivated.

The three motion layers

Every useful video prompt specifies at least three things, and they are not interchangeable:

  • Subject motion: what the person or object does inside the frame. One action, with a visible beginning and end.
  • Camera motion: how the frame itself travels. Static, push-in, pan, orbit, tilt, handheld follow.
  • Environmental motion: the ambient life of the scene. Wind in fabric, rain on a window, steam off a cup, a crowd in soft focus, a flickering sign.

Most weak prompts contain only the first layer, and often only as a noun rather than a verb. A woman in a coat is a subject. A woman in a coat turning her head toward the window is a shot.

Think in start states and end states

A reliable habit is to describe the frame at second zero and the frame at the last second, then let the model fill the middle. If you cannot say what has changed by the end, the clip has no reason to exist and will read as a screensaver.

Compare these two prompts:

  • Weak: A barista making coffee, cinematic.
  • Strong: A barista in a black apron finishes tamping and pulls the portafilter up out of frame, medium close-up, static camera, warm morning light from a window behind her, steam rising, shallow depth of field.

The second prompt has a start state, an end state, one camera instruction, one light instruction, and one piece of environmental motion. That is the whole recipe. It also gives the model a physical event to solve, which is far easier than asking it to invent a mood.

A reusable prompt skeleton for AI video

Five slots cover most needs. Write them in this order and the prompt stays readable even at sixty words.

[SUBJECT ANCHOR] + [ACTION WITH AN ARC] + [CAMERA: SIZE, MOVE, SPEED]
+ [LIGHT: SOURCE, DIRECTION, PALETTE] + [TEXTURE AND MOOD]

Assembled:

A man in his forties, close-cropped grey hair, canvas work jacket, lifting a
wooden crate onto a truck bed, medium-wide shot, slow truck-right following
him, overcast late-afternoon light from behind camera-left, desaturated
blue-grey palette, light 16mm grain, unhurried and documentary

Three details are doing quiet work here. The action has a clear physical arc because the crate goes from ground to truck bed. The camera move mirrors the action rather than fighting it. And the grain plus palette give the model a texture target so it does not default to glossy.

Rules that survive contact with real models

  • One action per clip. Two actions in four seconds produce a smear. Split them into two shots.
  • Name the speed. Slow and fast are meaningful to a model. Dynamic is not.
  • Prefer positive description to prohibition. An empty street with wet asphalt works better than no cars. If you must negate, describe the desired state first and the exclusion second.
  • Translate emotion into physics. Nostalgic is unfilmable. Warm tungsten interior, soft halation, and slight handheld sway are a shot.
  • Never stack three adjectives from different visual worlds. Gritty dreamy minimalist brutalist averages into beige.

A prompt is not a poem and not a specification document. It is a compact set of production decisions. If a clause does not tell the model what to show or how to move, it is probably costing you clarity.

Camera motion that survives generation

You do not need a film-school vocabulary, but ten terms will cover nearly everything, and each one is a shortcut to a specific feeling.

  • Locked-off: no camera movement at all. Underrated. Makes subject motion read clearly and is the safest choice when hands or faces are involved.
  • Push-in or dolly in: camera physically moves toward the subject. Builds attention, intimacy, and pressure.
  • Pull-back or dolly out: camera retreats. Reveals context, ends a thought, and works beautifully as a final shot.
  • Pan: camera rotates horizontally from a fixed point. Use sparingly in vertical frames, where a pan empties the composition fast.
  • Tilt: camera rotates vertically. A slow tilt up is one of the most reliable moves in a 9:16 frame.
  • Truck or track: camera slides sideways. Great for following a walk, risky for wide shots.
  • Crane or jib: camera rises or descends through space. Expensive-looking and easy to overuse.
  • Orbit or arc: camera circles the subject. Strong hero-shot energy, but watch for face warping mid-arc.
  • Handheld follow: camera trails the subject with natural sway. Reads as documentary or urgent.
  • Whip pan: very fast pan. A transition tool more than a shot, and hard to control in generation.

Zoom and dolly are different animals

A zoom changes the lens. A dolly changes the position. Zooming flattens the background and compresses depth, and it can feel slightly artificial, like surveillance footage. Dollying preserves perspective and feels like a person walking closer. If a model supports both, ask for the one you want explicitly. Slow dolly in and slow zoom in produce visibly different clips, and mixing them up is a common reason a shot feels wrong without anyone knowing why.

Give the camera one primary move and one qualifier

Slow push-in from eye level, ending on a medium close-up, is a complete instruction. Dynamic camera swirling around the subject while zooming and panning is a request for chaos. If you want a second movement, make it a consequence of the first, not a parallel demand. A camera should have a reason to move: to reveal, to follow, to settle, or to abandon.

Directing human motion without artifacts

People are the hardest subject in generated video, and the failures are predictable. Hands merge. Fingers multiply. A mouth opens for a line that does not exist. A head turn smears the jawline for several frames. These are not random. They cluster where the model has to invent unseen geometry: small occluded parts, fast rotations, and anything crossing the face.

Practical constraints that reduce artifacts

  • Keep motion slow and single. A slow head turn reads as intentional. A fast one reads as a glitch.
  • Keep hands below the chin or out of frame. If a hand must rise, keep it out of the face's silhouette.
  • Avoid fine finger work. No typing, no lacing, no unbuttoning. If the story needs it, shoot it as an insert on an object instead.
  • Close the mouth. Unless the clip has real dialogue, a closed or barely parted mouth avoids the uncanny speaking look.
  • One body in focus. Crowds and background extras are where faces turn to wax.
  • Move the camera back before rewriting the character. If faces are failing, a wider shot often solves it faster than another prompt revision.

Walking is a special case. It is the single most requested and least reliable action. It works best when the subject moves toward or away from the camera at a slow pace, with a handheld or trucking camera, and with legs partly obscured by framing, clothing, or foreground. Walking across the frame at speed, in a crowd, in a wide shot, is the hardest version of the task and rarely survives. If walking is essential, design the shot around the failure mode instead of hoping the model has suddenly learned anatomy.

Vertical frames rewrite the rules

A 9:16 frame is not a cropped 16:9 frame. It has different physics, and prompts written for widescreen often fail in vertical for spatial reasons rather than model reasons.

Lateral motion exits instantly. A truck-right move in a vertical frame empties the subject out of the composition in under a second. Favor vertical-friendly moves: tilt up, push-in, crane down, slow pull-back.

Headroom becomes a character. In vertical, the space above a subject is a design element. Say where the subject sits, such as subject centered with head in the upper third, or the model will center the eyeline and leave an awkward gap.

Leave a caption lane. If text will be overlaid, describe the composition so the lower third stays uncluttered. Subject in the upper two-thirds with empty foreground below is a prompt instruction, not just an editing note.

Depth reads better than width. Because horizontal space is scarce, use foreground occlusion such as a blurred doorframe, a plant, or a shoulder to create depth the frame cannot get from width. This is especially useful in profile-style clips, where a little foreground can make a simple action feel like a directed scene rather than a phone test.

A vertical clip also benefits from a clear speed plan. Fast movement can work, but it usually needs to move toward or away from the camera rather than across it. Slow movement across the frame is possible if the subject enters and exits on purpose, but it should not be the default.

Worked example: a short dating-profile video

A dating-profile clip has a practical job: show personality in a few seconds without looking staged. The mistake is to prompt a pose. A pose is a still image with a pulse. A better approach is to prompt a small, characterful action with a clear beginning and end.

Here is a three-beat set for an eight-second vertical profile clip set in a print studio. The person is the subject, but the action tells the viewer something about her.

Beat 1 (establish, 3s):
A small print studio at golden hour, wide shot, static tripod, warm window
light raking from camera left across a paper cutter, dust in the air,
35mm grain, quiet

Beat 2 (action, 3s):
A printer in her late twenties, cropped dark hair, olive canvas apron,
sliding a fresh sheet onto the press bed, medium shot, slow dolly in,
warm window light from camera left, 35mm grain

Beat 3 (detail, 2s):
The press roller turning, extreme close-up, static, shallow depth of field,
warm palette, 35mm grain

All three prompts share a palette, a light direction, and a grain treatment, so any two cut together. The subject anchor sentence repeats word for word where the person appears. That repetition is what makes the sequence feel like one room. If you only generate one shot, Beat 2 is the strongest because it contains an action with an arc and a camera move that supports it.

The decision rule is simple: if the clip can be represented by a still image, it is not a video yet. Add an entrance, an exit, or a transformation. In a profile clip, that might be lifting a mug, turning toward a window, closing a notebook, or stepping into frame. The action does not need to be dramatic. It needs to be visible.

For a fifteen-second character intro, use three beats: establishing, action, detail. The establishing shot tells us where we are. The action shot tells us who the person is through behavior. The detail shot gives the sequence a texture and a closing punctuation. Three beats are more reliable than ten because each shot has one job and one motion instruction.

Consistency is a documentation habit

Sequence work collapses when documentation is sloppy. Keep a plain text file beside the project and fill it in before generating anything.

  • Anchor sentence: one exact sentence describing the subject. Copy and paste it into every prompt. Never retype it from memory.
  • Location lock: one sentence describing the room, street, or set, reused verbatim.
  • Light ledger: source, direction, time of day, and color temperature, written once.
  • Wardrobe and props: a list, not a description. Props that change between shots are the hardest continuity break to forgive in an edit.
  • Reference stills: approve a still frame first, then animate from it. Image-to-video removes most variance because composition and character are already locked. Building that still with an AI image generator is usually faster than three text-only video attempts.

Before generating shot four, compare it against shot one on eight items: wardrobe, hair, props, time of day, light direction, color temperature, lens character, and movement speed. Most failures are one of those eight drifting quietly while you were focused on the interesting part.

Consistency is not a magic setting. It is a stack of small contracts you make with yourself and then honor. The model can only be as consistent as the description it receives. If the description changes, the character changes.

Common mistakes, iteration, and tool choice

Most bad clips are not mysterious. They are the result of a prompt that asked for too much, too little, or the wrong thing at the wrong time.

Mistake Why it hurts Fix
Prompt describes a still Model invents drift to compensate Add a start state and an end state
Two actions in one clip Motion smears at the transition Split into two shots
No camera instruction Random framing, unusable in an edit Always name size and move
Three color palettes Model averages them into beige Pick one palette and delete the rest
Rewriting everything per attempt No learning and no repeatability Vary one element per generation
Aspect ratio decided last Reframing crops the composition you designed Set delivery ratio before the first prompt
Animating an unapproved still You spend time on a frame you will discard Lock the image, then animate
Fast head turns Jaw and ear artifacts for several frames Slow the rotation or cut before it

Iteration should be disciplined. If you rewrite the subject, the camera, and the light together, the result teaches you nothing. You cannot tell which change fixed the shot, so you cannot repeat the fix. Keep a log with four columns: prompt version, the single variable changed, verdict, and one sentence of reasoning. After twenty generations you own a personal phrasebook for your chosen model, and that phrasebook is worth more than any generic list of tips.

Review in batches and review muted. Judge ten clips in one sitting rather than one at a time, because isolated judgment drifts optimistic. Watch them all muted first. If a clip does not communicate with the sound off, the music was doing the work. Then watch at thumbnail size, which is how most viewers will meet it.

When choosing a tool, score options against the things that affect your week: prompt adherence, motion realism in the first two seconds, single-clip duration, image-to-video support, aspect ratio coverage, iteration speed, and predictable session cost. Compare a few AI video generator alternatives on those criteria rather than on a highlight reel. If you are learning structure, start from video templates because a template hands you a beat sheet and camera plan, which is most of the battle. A curated prompt library is useful for phrasing, but your own log is what makes results predictable rather than lucky.

FAQ

How long should an AI video prompt be?

Usually 25 to 60 words. Long enough for the five slots, short enough that no two clauses contradict each other. Past 80 words, look for something to cut rather than something to add.

Why does my character change between shots?

The description changed. Reuse the subject sentence verbatim, lock a reference still, and animate from that still instead of from fresh text. Consistency is a documentation habit, not a setting.

Can I use the same prompt across different models?

Keep the structure and the anchors, but adjust vocabulary. Some models respond to lens and film-stock shorthand. Others prefer plain physical description. Your log will tell you which is which within a dozen generations.

How many generations does a usable clip take?

For simple action in a forgiving frame, three to five. For faces, hands, or complex environments, ten or more. Plan the session around that reality instead of being surprised by it.

Is a static camera a waste of a video model?

No. A locked-off shot with strong subject motion is often the most usable clip you will get, because every unit of model attention goes into the action rather than into maintaining a camera path.

How do I plan a thirty-second piece?

Plan in beats, not shots. Four to six beats, one shot each, with one establishing frame per location. Solve continuity on paper first, then generate. Sequences that fail at the beat sheet stage fail harder once motion is added.

Put cinematic motion into your next clip

The pattern holds across every project that works: decide the shot, name the three motion layers, anchor the subject, generate deliberately, then edit with intent. None of it depends on a rare talent for wording. It depends on refusing to treat generation as a lottery.

Start small with one beat, one camera move, and one action with a clear end state. When that clip reads the way you imagined it, add a second shot and check the eight-point continuity list before you generate. That is how a folder of experiments turns into a sequence, and how a profile clip starts to feel directed rather than generated.

When you are ready to direct rather than gamble, open the AI video generator and begin with the shot you already know how to describe.