Orelon logoOrelon
料金

Body Shape Prompts for AI Characters: A Consistency Guide

2026年11月5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。

Write body shape prompts that keep AI characters consistent across images and video, with structural ratios, negative cues, and a shot-by-shot QA loop.

Most character prompts fail on the body long before they fail on the face. You describe a tall, broad-shouldered dockworker, and what comes back is a generic figure with a jacket draped over it: shoulders narrowing in one frame and widening in the next, legs changing length between renders, the same character reading as two different people by the fourth shot. Faces get all the attention because they are easy to judge at a glance, but bodies carry the silhouette, the thumbnail read, and the continuity that makes a sequence feel deliberate rather than assembled.

This guide covers the practical craft of describing body shape in prompts: the vocabulary that actually changes a render, a structure that keeps build and proportion stable when you switch styles or cameras, and a workflow for holding one body consistent across stills and motion. It is written for people building characters for narrative shorts, product storytelling, and social video, where a character has to be recognizable in under a second.

Why a body reads as a different person between shots

Faces have a dense shared vocabulary. Everyone knows roughly what "high cheekbones" or "heavy brows" will do, and models have seen millions of labeled examples. Bodies are described more loosely, in words borrowed from fashion captions and fitness marketing, and those words overlap in ways models resolve by averaging. Write "slender, toned, curvy, petite" in one prompt and you have not described four traits; you have described a contradiction that the model solves by drifting toward a neutral middle.

The second problem is that body shape is relational, not absolute. Shoulder width only means something next to hip width. Leg length only reads against torso length. When you name a single attribute without the surrounding ratios, the model is free to satisfy it however it likes, and it will often satisfy it in a way that contradicts the previous shot.

The third problem is interaction with clothing and camera. A bulky jacket can erase a narrow waist. A wide-angle lens at close distance lengthens a torso and shrinks the head. If your prompt treats body shape as a fixed fact about the character rather than a fact about the frame, every change of wardrobe or lens becomes a change of physique.

The fix is not longer prompts. It is a small, disciplined vocabulary plus a written spec you reuse, so the body is defined once and then referenced instead of reinvented. Everything below builds toward that.

The three layers of body description

Describe bodies at three levels, in this order. Each layer survives a different kind of change, which is why the order matters.

Layer one: silhouette

The silhouette is the single shape you would recognize as a black cutout. "Broad, top-heavy frame" and "compact, even frame" are silhouette statements, and they survive style changes better than detail words because they describe mass distribution rather than measurement. When you only have one sentence to spend on a body, spend it here.

A useful test: squint at your reference until detail disappears. Whatever shape remains is the silhouette, and that is what a viewer processes first in a feed, a thumbnail, or the opening frame of a clip.

Layer two: structural ratio

This is where most consistency actually lives. Ratios are the thing that flips between shots and makes a character look recast, even when the face is identical. Name four of them and nothing else:

  • Shoulder width relative to hip width
  • Torso length relative to leg length
  • Limb thickness relative to torso mass
  • Head size relative to shoulder span

Phrase these comparatively rather than numerically, because models handle "shoulders clearly wider than hips" more reliably than "forty-four centimeter shoulder span." Measurements invite the model to guess at units; comparisons give it a direction.

Layer three: surface and fit

Surface cues describe how clothing tension, posture, and weight distribution reveal the frame underneath. "Fabric pulled tight across the shoulders" tells the model both the garment and the underlying build without adding another adjective. "Sleeves ending above the wrist" implies long arms without saying "long arms."

This layer is also where asymmetry lives, and asymmetry is what separates a photographed person from a mannequin. A slightly higher left shoulder, a stance that favors one leg, a neck carried a little forward — one such detail is worth five adjectives.

Words that quietly cancel each other

Some combinations cannot all be true. Slender plus muscular plus curvy plus petite will average out. "Delicate frame" plus "powerful build" will average out. "Realistic proportions" plus "fashion illustration elongation" will average out, usually in the direction of whichever phrase you wrote last.

Negation inside a positive prompt is weak. "Not skinny, not heavy" is not a body description; it is noise the model has to interpret. If you need to exclude something, put it in a dedicated negative field where exclusion actually works, and keep the positive field purely additive.

Evaluative language is the other quiet failure. "Handsome build" and "good body" carry no geometry at all — they carry taste, and taste is not renderable. Replace them with shape words and let the viewer supply the judgment.

Build a reusable body spec before you open the prompt box

Write six lines about the body, in plain language, before generating anything. This takes four minutes and saves hours of re-rolling. Keep it in a notes file next to the character's face description and wardrobe list.

Attribute Write something like Why it holds up
Silhouette Broad upper body, straight fall to the hip Survives style and wardrobe changes
Shoulder to hip Shoulders clearly wider than hips Anchors the most drift-prone ratio
Limb proportion Long arms, hands past mid-thigh Prevents shrinking limbs in wide shots
Mass Dense, thick-set, low body fat Gives the model a material, not a mood
Posture Slight forward lean, shoulders held back Carries into motion and gait
Signature cue Thick neck, wide jaw-to-shoulder line One memorable anchor for quick checks

The spec exists so that every prompt you write later contains the same three or four structural phrases, word for word. Consistency across a sequence comes mostly from copy-paste discipline, not from clever synonyms.

Keep the spec under forty words. If it is longer, you will start trimming it under pressure and the character will drift. A body spec is a contract, not a description.

A prompt formula that survives wardrobe and lens changes

Order matters. Models weight early tokens more heavily, so put the subject and body before wardrobe, before lighting, before lens.

Block one — identity in one clause. Who this is: role, age band, silhouette, stance. "A dockworker in his fifties, broad and top-heavy, standing squarely." That clause already fixes mass distribution and posture.

Block two — proportion, verbatim. Add the ratios from your spec without paraphrasing. This is the block you never rewrite between shots. If you change one word here, expect the body to change.

Block three — fit, not fashion. Describe clothing by how it sits on the frame. "Heavy canvas jacket straining slightly at the shoulders, sleeves ending above the wrist" preserves the body while adding wardrobe. Compare that to "oversized jacket," which deletes the shoulders entirely.

A complete prompt built this way:

A dockworker in his fifties, broad and top-heavy, standing squarely.
Shoulders clearly wider than hips, thick neck, long arms with hands
past mid-thigh, short torso, dense build. Heavy canvas jacket
straining slightly at the shoulders, sleeves ending above the wrist.
Overcast side light, 50mm lens at chest height, documentary realism.

Three habits fall out of this structure. First, the body block stays identical across every shot; only the final line changes. Second, wardrobe is written as fit, so a costume change does not become a body change. Third, lens and camera height are explicit, because a low camera makes everyone look longer-legged and a high camera flattens the torso.

Worked example: one body, three shots

A wide establishing shot keeps the body block and adds "full body, camera at hip height, 35mm." A medium shot keeps the body block and changes only to "waist up, camera at chest height, 50mm." A close shot keeps the body block, adds "head and shoulders, neck and shoulder line visible," and accepts that the visible proportion cue is now the neck-to-shoulder relationship. The character stays the same person because the structural phrases never moved.

Negative cues that fix anatomy without flattening style

Negative prompts are where you delete failure modes, not where you restate the character. Keep the list short, stable, and specific to anatomy and rendering:

  • distorted proportions, elongated neck, uneven shoulders, mismatched limb length
  • extra fingers, fused limbs, floating anatomy
  • melting silhouette, collapsed torso, body horror
  • mannequin skin, plastic sheen (only when you want photographic realism)

Two cautions. A long negative list can flatten style, because many negatives also suppress legitimate stylization. And do not negate body types you simply did not ask for; that is what the positive block is for, and negating broad categories tends to produce stiff, defensive renders with flat lighting.

If anatomy keeps breaking, the problem is usually pose complexity rather than body description. Simplify the action, keep the camera near chest height, and describe one clear stance instead of motion blur or a dynamic jump. When in doubt, remove the action verb and re-render before touching the body text.

Holding one body across a sequence

Reference images and fixed seeds

When the tool supports image references, generate one clean full-body reference at neutral pose and neutral light, then reuse it for every subsequent shot. That single image does more for continuity than any amount of adjective tuning. Where seeds are available, lock the seed for the reference pass, then vary only pose and camera.

Decision rule: use a reference image when the character must be recognizable across shots; use seed locking when you need repeatable composition or lighting; use both when the sequence will be cut quickly. If you can only keep one, keep the reference image — it encodes the body in a way text cannot.

A drift audit you can run in two minutes

After every four or five generations, place the newest render next to the reference and compare four things:

  1. Shoulder-to-hip ratio
  2. Head size relative to shoulder span
  3. Limb thickness
  4. Stance width

Drift almost always shows up in one of those four before it shows up anywhere else. If the body still matches but the character feels wrong, the problem has moved to wardrobe, lens, or lighting — not the body block.

A continuity checklist before approving a shot

  • Is the body block identical to the reference prompt, word for word?
  • Does the shoulder-to-hip relationship still hold at this camera distance?
  • Has wardrobe changed in a way that hides a structural cue?
  • Is the lens plausible for the shot size, or did a wide lens quietly restretch the body?
  • Does the silhouette still read at thumbnail size?

That last check matters most for video. A character who reads clearly as a cutout at 300 pixels will hold attention; one who reads as an average figure will not.

Camera, lens, and wardrobe: the hidden variables

Three variables silently rewrite the body without touching a single body word.

Camera height. A camera at hip height elongates legs and shrinks the head. A camera at head height compresses the torso. Pick a height per shot type and keep it consistent within a scene: hip height for establishing, chest height for medium, eye height for close work.

Focal length. A short lens close to the subject distorts — nearer features swell, the body stretches. A longer lens flattens and compresses. If you shot a character at 35mm and then switch to 85mm without saying so, the body will appear to change even though your text never moved.

Wardrobe volume. Every garment adds apparent mass. A puffer jacket on a narrow frame reads as a wider person; a tailored coat on a broad frame reads narrower. Write fit in the positive block and, when continuity matters, name one visible structural cue the garment cannot hide: the neck, the wrist, the shoulder line, the ankle.

A practical habit: describe the wardrobe's interaction with the frame in the same sentence as the garment. "Loose linen shirt falling from the shoulder line" keeps both pieces of information alive. "Loose shirt" alone lets the model decide whether shoulders exist.

Matching description to style and genre

Every style reinterprets anatomy. Animated and illustrative styles exaggerate head-to-body ratios, simplify limbs, and widen shoulders for readability. Fashion-editorial styles lengthen legs and narrow waists. Documentary realism keeps proportions close to life but adds asymmetry, which is what makes a person look photographed rather than modeled.

Decide the style before the body spec, because the spec has to be written in the style's own proportion language. A "compact, dense" character in a stylized animation style needs an explicit head-to-body ratio, or the style will pull it toward the genre default. If you are exploring stylizations, working from a prompt library of proven structural phrases is faster than inventing new ones per project.

Genre also sets expectations you can use deliberately. Thrillers reward heavy, grounded silhouettes. Sports and action read better with long limbs and clear shoulder lines. Quiet drama benefits from softer, less geometric descriptions. None of this is about beauty standards; it is about silhouette legibility, which is the only thing a viewer processes in the first half second.

One more decision: realism is a rendering choice, not a measurement choice. You can keep exaggerated proportions and still ask for photographic skin, believable lens behavior, and natural light. That combination is often more striking than either extreme.

Mistakes that cost the most time

Stacking near-synonyms. "Slim, slender, lean, lithe" is one instruction repeated four times. Fix: pick one term and add a ratio instead.

Describing the body only through clothing. "Wears a bulky sweater" tells you nothing about the person inside it. Fix: state the frame, then describe the fit.

Paraphrasing the body block. Small rewording becomes large drift over eight shots. Fix: copy-paste the block, edit only wardrobe, pose, and camera.

Forgetting camera height and focal length. Identical prompts produce different bodies at different heights. Fix: always specify both in the final line.

Overloading the prompt. After roughly the fourth body clause, models start trading attributes against each other, and the last clause often wins. Fix: cut to three structural phrases and move the rest to negatives.

Using evaluative words. "Athletic" and "fit" mean ten different things to ten viewers. Fix: replace with shape and mass words the render can show.

Skipping asymmetry. Perfectly symmetrical bodies look synthetic. Fix: add one small real detail — a favored leg, a healed injury stance, one shoulder slightly forward.

Never testing at thumbnail size. Fix: shrink the render before approving it. If the silhouette disappears, the description is too vague.

Changing the style mid-sequence. Fix: lock style, lens family, and lighting direction before the first approved frame, then treat any change as a new character pass.

From a locked still to a moving shot

Motion adds variables that stills hide: gait, weight shift, stance width, and how fabric behaves when the body moves. The body block carries over almost unchanged, but the camera sentence needs more detail and one line about movement.

Write motion as mechanics rather than mood. "Walks with a slight forward lean, weight settling heavily on each step, arms swinging close to the body" gives the model something to render. "Walks confidently" does not.

Two practical rules for the first motion test:

  1. Keep the shot short and the action simple. One step, one turn, one gesture. Complex action multiplies anatomy errors faster than any body description can fix them.
  2. Compare frame one and the final frame against the still reference. If the shoulders narrow or the head grows between them, the motion prompt is fighting the body block — shorten it and try again.

A sensible order of operations: lock the full-body reference in the AI image generator, confirm the silhouette at thumbnail size, then push the character into motion in the AI video generator with the same structural phrases. Only after the first clip holds together should you build out the rest of the sequence.

FAQ

How many words should I spend on body shape?

Usually fifteen to thirty, arranged as three or four structural phrases plus one signature cue. Longer descriptions dilute rather than refine, because each added clause competes with the others for the model's attention.

Why does the same character change build between shots?

Almost always one of three causes: the body block was paraphrased, the camera distance or focal length changed, or wardrobe replaced the visible frame. Compare the prompts side by side and check those three before rewriting the description.

Can I describe several characters in one prompt?

You can, but attribute bleed rises sharply and each character loses specificity. Generate each character alone against a neutral background, then compose them together using references. That workflow keeps every body spec clean and reusable.

Do negative prompts fix bad anatomy?

They help with obvious failures such as fused limbs, extra fingers, or elongated necks, but they do not repair a contradictory positive description. Fix the additive block first, then use negatives to clean up edges and rendering artifacts.

How do I keep a child or older character's build consistent?

Anchor on head-to-body ratio and shoulder line rather than height words. State the ratio explicitly and keep that phrase frozen across shots, because models otherwise drift toward adult proportions over a sequence. For older characters, add one posture cue — a settled spine, a shorter stride — that persists in motion.

Does realism mean avoiding stylized proportions?

No. Realism refers to rendering, not measurement. Exaggerated proportions with photographic skin, believable lens behavior, and natural light is a legitimate and often memorable combination.

What if the face is consistent but the body keeps changing?

Treat the face reference as solved and stop editing it. Lock the seed, fix camera height, and shorten the body block to the three phrases that carry the most structure. In most cases the drift disappears once the prompt stops competing with itself.

Take the body from prompt to motion

A consistent character is a written spec, a frozen block of structural phrases, and a camera sentence you change deliberately. Do that once, and every later shot becomes a variation instead of a gamble: same shoulders, same silhouette, same person walking through the frame.

Start with a clean full-body reference in the AI image generator, lock the body block, and verify the silhouette at thumbnail size. Then move the character into motion with Orelon's AI video generator, where the same structural phrases carry from still to clip. If you want a head start, browse the Orelon blog for more prompt workflows, or open the studio and write your first body spec today.