Orelon logoOrelon
料金

Consistent Character Prompts for AI Video: A Practical Guide

2026年11月7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 consistent character prompts for AI video, hold identity across every shot, and build a repeatable cinematic workflow that scales.

Character consistency is the difference between a clip that looks impressive for four seconds and a sequence a viewer will follow for two minutes. One face that holds across eight shots lets you tell a story. A face that becomes a different person every generation forces you to hide the seams behind fast cuts, heavy grading, and narration that apologizes for the visuals. The encouraging part is that consistency is rarely something you have to wait for a model release to fix. It is a prompt discipline you can build today, and it compounds across every project you ship.

This guide covers why generated faces drift, how to write a character sheet that survives scene changes, how to vary camera and emotion without changing identity, and how to run the whole thing as a repeatable workflow inside an AI video generator.

Why faces drift between generations

Every clip begins as noise that the model resolves toward the most plausible reading of your words. That resolution step is where identity leaks out. Three problems cause most of the damage.

The description is underspecified. A woman in her thirties with brown hair describes millions of people. The model picks one at random each run, and random is the opposite of continuity.

The attributes contradict each other. Sharp jawline, soft round face, childlike features, mature gaze. That is two incompatible blueprints in one sentence. The model resolves the conflict differently every time, so you get a family resemblance rather than a person.

You changed a variable you did not mean to change. Lighting, lens, weather, and time of day all push the model toward different reference material. Move a character from noon sun to candlelight and the model may quietly substitute a different face that fits the new lighting more naturally.

The three signals that control identity

Three inputs decide whether a face holds. A reference image is the strongest signal when your tool supports one. A written description is the portable signal that works everywhere, including tools that ignore reference images entirely. A seed or approved starting frame locks the noise pattern so a re-run stays close to the original.

Use all three when you can, but always keep the description strong enough to stand alone. It is the only part that transfers between tools, aspect ratios, model versions, and collaborators.

Order matters more than length

Text conditioning gives more weight to the early part of a prompt in most pipelines. Put identity in the first sentence, wardrobe next, then action, then environment, then light, then lens and grade. If you bury the face description behind three clauses about the weather, the model will build the weather first and fit a face into it afterward.

Build a character sheet before you generate a single clip

A character sheet is one small text block you write once and paste into every prompt. It removes the improvisation that causes drift. Keep it in a plain text file, not in your head.

The five blocks

  • Identity block. Age range, heritage if it matters to the story, face shape, eyes, nose, mouth, hair length and texture, skin tone, and two or three asymmetries.
  • Signature block. The details a viewer would use to recognize the character in silhouette: posture, a habitual gesture, a piece of jewelry, a scar, a specific jacket.
  • Wardrobe block. Named garments with material, color, and finish. Slate grey wool overcoat, slightly oversized, collar turned up beats nice coat.
  • Performance block. Resting expression, energy level, how they carry their shoulders, whether they meet the camera's eye.
  • Exclusion block. What must never appear: no beard, no glasses, no hat, no visible tattoos, no broad smile.

A worked identity block

ILSE — 36, Northern European, long narrow face with a high
forehead, straight dark brows, deep-set grey-green eyes, straight
nose with a slight bump at the bridge, thin upper lip, small scar
through the left brow tail, dark auburn hair cut to the jaw,
unstyled, faint freckling across the nose only. Resting expression:
watchful, unsmiling.

That paragraph is roughly eighty words and it will do more for continuity than any model upgrade. Notice how much of it is specific and slightly unflattering: the bump, the scar, the thin lip. Generic beauty is the enemy of recognition, because attractive faces converge toward an average in the model's learned space. A flaw gives the model something to hold onto.

Write for the tool you will use tomorrow

The identity block should read well in any engine. Avoid brand-specific syntax, avoid parameters that only one platform understands, and avoid references to a named actor or a real public figure. Describe the geometry of the face instead. When you want a visual anchor as well, generate a portrait first with the AI image generator, pick the frame that matches your mental image, and reuse it as a reference for every subsequent shot.

The master prompt template

Once the sheet exists, every shot becomes an assembly job rather than an act of creation.

[shot type and camera move] + [identity block] + [wardrobe block]
+ [action and performance] + [environment] + [lighting]
+ [lens and format] + [mood and grade]

A worked shot prompt

Medium close-up, slow push in. Ilse, 36, long narrow face, high
forehead, straight dark brows, deep-set grey-green eyes, straight
nose with a slight bump, thin upper lip, small scar through the left
brow tail, dark auburn jaw-length hair, faint freckles across the
nose, watchful and unsmiling. Slate grey wool overcoat, collar up,
black knit scarf. She turns her head slightly toward a sound
off-screen. Empty rural bus shelter at dusk, wet asphalt. Soft
overcast light, cool shadows, no direct sun. 50mm lens, shallow
depth of field, cinematic 2.39:1 framing, muted teal grade, fine
grain.

Read it out loud. Every clause is doing a job, and nothing contradicts anything else. That is the whole test.

Fields you should never paraphrase

The identity block and the wardrobe block are frozen text. Copy and paste them, character for character. The moment someone on your team rewrites deep-set grey-green eyes as green eyes with heavy lids, continuity dies and nobody knows why. Everything else — action, environment, grade — is meant to change per shot.

Varying the shot without changing the person

A sequence needs variety. The common mistake is changing everything at once and then blaming the model when the face moves. Change one axis per generation and check the result against the previous shot.

Variable Safe to change freely Change with care
Shot size Wide, medium, close Extreme close-up
Camera move Static, push in, slow orbit Fast handheld, whip pans
Action Small gestures, turns, walking Running, fighting, falling
Environment Room to street to field Day to night in one jump
Light quality Soft to hard within one family Warm tungsten to cold moonlight
Wardrobe Removing a coat or scarf Full outfit replacement

Camera, lens, and shot size

Shot size is the cheapest variety you can add. A wide establishing shot with the character small in frame carries almost no identity risk, and it makes the following close-up feel earned. Keep the same lens family across a sequence so perspective does not lurch between shots; jumping from a 24mm to an 85mm and back reads as a different production, not a different angle.

Emotion without face meltdown

Asking for devastated, sobbing, screaming inside a four-second clip is a recipe for distortion. Describe the physical evidence of the emotion instead: jaw tight, eyes glassy, blinking slowly, breath visible in the cold. Subtle performance direction reads as acting. Extreme emotional labels read as a request to reshape the face, and the model will oblige.

Environment, weather, and time of day

If a scene must move from afternoon to night, add compensating language that tells the model only the source changed, not the visual grammar: same soft key light source, now a single practical lamp, same cool shadow bias, same contrast curve. Keep the lighting family constant across a sequence even when the location changes completely.

A repeatable workflow: from character sheet to a five-shot sequence

  1. Write the sheet. Five blocks, one page maximum, saved as plain text you can paste from.
  2. Test the identity block as a portrait. Use the AI image generator at a fixed focal length and iterate on wording until the face is unmistakably yours.
  3. Freeze the wording. From this point the identity and wardrobe blocks are copy and paste only.
  4. Build a shot list. Five to eight shots, each with one narrative purpose. Write every prompt before generating any of them.
  5. Generate the widest shot first. Wide shots are forgiving and establish the world. Work inward toward close-ups.
  6. Animate approved stills. Feed a still you have already approved into the AI video generator and animate it rather than describing the scene from nothing.
  7. Log every approved prompt. A running document of what worked becomes your own prompt library and cuts setup time on the next project by half.
  8. Reshoot single shots, not sequences. If shot four fails, regenerate shot four with the same frozen blocks. Do not rebuild the set.

For teams producing recurring content — a series, a channel, a brand character — step three is the one that decides whether the project stays coherent. Version your sheet the way you would version code: date it, note which model version you tuned it against, and keep the approved still beside it so you can rebuild the look months later.

The sixty-second QA pass before export

Run five checks on every clip. Does the face match the approved still at the first frame and the last frame? Is the wardrobe identical in material and color, not just similar? Is the lighting family the same as the previous shot? Does the performance match the performance block? Is anything from the exclusion block visible anywhere in frame?

A failure on any check means regenerating that single shot. A pass on all five means the sequence will cut together without the viewer noticing seams, which is the entire goal.

Multi-character scenes: keeping two people apart

Two characters in one frame is the hardest problem in the format, and the failure is always the same: the model blends them into a hybrid person in the middle of the shot. There are four practical countermeasures.

Separate them physically. Put one character in the foreground and one in the background, or split the frame with a doorway, a car, a window frame. Separation gives the model two distinct regions to resolve.

Contrast the silhouettes. Different heights, different hair volumes, different shoulder widths, different posture. Two people of similar build in similar coats will merge.

Contrast the palettes. Warm wardrobe against cool wardrobe, light against dark, textured against smooth.

Describe each character in its own sentence. Never blend adjectives into a shared clause. Write Ilse first with her full block, then write the second character with theirs, and explicitly state who is where in the frame.

Common failure modes and how to fix them

The face morphs mid-clip. Usually too much action in too short a window. Shorten the clip, simplify the motion, or split the action into two shots that cut together.

Wardrobe color shifts between shots. Vague color words drift. Use material plus color plus finish: matte charcoal wool, brushed steel grey, faded olive canvas.

The character ages between shots. Age is one of the most volatile attributes. Add an explicit anchor: early thirties, no visible lines, smooth skin, or late forties, deep nasolabial lines, greying temples.

Everyone looks like a stock photo. Add asymmetry and imperfection. A crooked nose, a chipped tooth, uneven eyebrows, a burn mark on one hand. Averages converge; character does not.

The background steals the shot. Reduce environmental detail in close-ups. A busy street with twenty readable objects competes with the face for the model's attention.

The prompt grows every generation. Someone keeps adding clauses to fix small problems until the prompt is three hundred words and nothing holds. When that happens, cut back to the frozen blocks and change one field only.

The reference image fights the description. If your anchor still shows a round face and your text says narrow face, the model splits the difference unpredictably. Pick one source of truth and delete the other.

Decision criteria for character-driven work

Model quality matters less than workflow fit. Judge any platform on six things before committing a project to it.

  • Reference image support. Can you anchor a character to a still, and how strongly does the anchor hold across a sequence?
  • Clip length versus stability. Longer clips drift more. A tool that produces clean five-second shots beats one that produces shaky ten-second shots for character work.
  • Iteration cost. Character work is iterative by nature, so measure how expensive a failed generation is in time and plan limits, not just in money.
  • Format control. Vertical, square, and widescreen should all be available without rebuilding the prompt, and switching ratio should not require a new anchor still from scratch.
  • Prompt fidelity. Does the model respect specific descriptive language, or does it flatten everything into a house style that overrides your scar and your nose bump?
  • Reusable starting points. Templates and presets remove an entire class of setup errors and let you drop a character into a proven shot structure.

Comparing engines side by side before you commit saves weeks; the alternatives overview shows how different tools handle reference images and motion, and the video templates library is useful when you want a proven shot structure for a new character.

Genre snippets you can adapt

Each snippet below follows the same order — identity, wardrobe, action, environment, light, lens, grade. That structural consistency is what makes output consistent.

Noir detective. Medium shot, static. Man in his fifties, heavy brow, broken nose set slightly left, grey stubble, thinning hair combed back. Dark navy trench coat, shoulders damp. Standing under a streetlamp in rain, wet brick behind him. Hard top light, deep shadows, no fill. 40mm lens, high contrast, desaturated grade.

Science fiction courier. Wide shot, slow orbit. Woman, late twenties, sharp cheekbones, shaved head with a faint scar above the right ear, dark brown skin, pale grey eyes. Weathered utility jacket, faded orange panels, scuffed shoulder plates. Crossing a hangar deck at night, blue practical lights. Cool ambient light with warm rim. 24mm lens, anamorphic flare, teal and amber grade.

Brand character. Close-up, static. Person in their thirties, warm round face, short curly black hair, small mole on the right cheek, easy half-smile. Soft oatmeal knit sweater, no logos. Seated near a window in a bright studio. Diffused daylight, gentle falloff. 85mm lens, shallow depth, clean neutral grade.

Documentary interview. Medium close-up, locked off. Woman, early forties, square jaw, wide-set hazel eyes, freckled shoulders, dark hair pulled back loosely. Charcoal linen shirt, sleeves rolled. Sitting slightly off-axis in a workshop, tools blurred behind her. Soft north window light, cool shadows. 65mm lens, shallow depth, natural grade with slight warmth.

FAQ

How many words should a character prompt be? Enough to be unmistakable, usually eighty to one hundred and fifty words for the identity and wardrobe blocks combined. Beyond that, extra adjectives dilute rather than sharpen, and the model starts averaging your details.

Do I need a reference image to stay consistent? It helps significantly, but a precise written description can carry a short sequence on its own. The strongest setup uses both, with the description doing the portable work across tools and formats.

Why does my character change when I switch aspect ratio? Cropping changes composition, and composition changes which training examples the model leans on. Generate a fresh anchor still in the new ratio before continuing the sequence, then freeze that still as the new reference.

Should I reuse the same seed for every shot? Same seed plus same prompt gives near-identical frames, which is not what a sequence needs. Instead keep the description stable, vary only the shot-specific fields, or animate an approved still.

How long can a single consistent clip be? Shorter than most people want. Four to six seconds is the reliable zone for character work, and drift rises sharply beyond it. Plan a sequence in short shots rather than one long take.

Can I keep several characters consistent at once? Yes, but give each its own sheet and keep them visually separated by height, silhouette, palette, and framing. Two similar-looking people in the same frame is the hardest case in the format.

What do I do when a shot fails and I cannot tell why? Revert to the last approved prompt, change exactly one field, and regenerate. Bisecting a prompt is faster than rewriting it, and it teaches you which words your model is actually sensitive to.

Start with one face, not ten

The instinct is to build an entire cast before shooting anything. Resist it. Pick one character, write one sheet, and produce a five-shot sequence. You will learn more about how your model handles age, wardrobe, and lighting in that single exercise than in a week of reading about it.

When you are ready to put the method into practice, Orelon gives you a cinematic space to move from idea to motion: generate an anchor portrait, animate it into a shot, and iterate on a single frame instead of rebuilding a scene. Bring your character sheet to the AI video generator, keep the identity block frozen, and let the camera do the changing. More workflow breakdowns like this one live on the Orelon blog.