Orelon logoOrelon
Pricing

AI OC Interaction Prompts for Cinematic Character Scenes

Nov 6, 2026 · By Orelon Team

Explore AI video templates

Browse a few community creations for inspiration, then open any template to continue creating in Orelon.

Build original characters that stay recognizable: a practical guide to writing AI OC interaction prompts, blocking two-character scenes, and fixing drift.

Two characters stand in a doorway. One of them is yours — an original character you have drawn, described, and thought about for months. In the render she is wearing someone else's jacket, her scar has migrated to the wrong eyebrow, and the person opposite her seems to have absorbed half of her face.

The prompt was not lazy. It was doing two jobs and only paying attention to one of them.

An OC interaction prompt is a scene brief, not a portrait brief. Its real work is holding a person's identity steady while that person wants something from someone else, in a specific space, seen through a specific lens. What follows is the craft of that: how to write the prompt, how to diagnose the failures that keep repeating, and how to run a loop that lands a usable shot in three passes instead of thirty.

What an OC interaction prompt actually controls

A single-character prompt describes a look. A two-character prompt describes a relationship, and relationships are made of space, gaze, and intent. The moment a second person enters the frame, the number of things that can go wrong multiplies: hairstyles blend, heights swap, one character inherits the other's coat, eye contact lands on a shoulder instead of a face.

Every workable interaction prompt answers five questions at once:

  • Who is in frame. A short, repeatable identity description for each character.
  • What each one wants. The intention behind the pose, not the pose itself.
  • Where everyone is. Distance, side, height difference, who sits nearer the lens.
  • How the camera sees it. Shot size, angle, lens feel, movement.
  • What cannot change. The continuity rules that carry into the next shot.

When a render feels wrong but you cannot name why, one of those five is missing. Most often it is the third. Writers produce gorgeous character descriptions and then forget to say which person is on the left, so the model flips them between takes and the scene stops cutting together. A close second is the fifth: nobody wrote down the light direction, so the reverse shot looks like a different day.

Build the character bible before the first render

Trying to establish a brand-new character directly in motion is the slowest route to a stable one. Do it in two stages: lock the look as a still, then animate from what you can actually see.

Start in the AI image generator and iterate on a single clean reference frame until you would recognize that face in a crowd. Then write the identity block by describing the still you kept — not the character you imagined. This is the step most people skip, and it is exactly why their prompts and their references quietly disagree.

What belongs in an identity block

Three or four visually load-bearing, slightly odd details survive rendering far better than a list of attractive adjectives. Beautiful woman with long dark hair is a category. Shaved head, scar through the left eyebrow, silver hoop in the right ear is a person.

Field What to write Weak version Strong version
Name A fixed label used in every prompt the girl Mira
Age band A range, not a precise number twenty-eight late twenties
Silhouette Hair, height, build slim short, square shoulders, cropped hair
Anchors Two or three unusual marks pretty scar through left eyebrow
Wardrobe One outfit per scene block jacket grey mechanic coverall, sleeves rolled
Palette Two dominant colors nice colors steel grey and oil black

Then obey one rule without exception: copy the identity block verbatim. Repetition is not laziness, it is the mechanism. Paraphrasing a scar through the left eyebrow into a scarred brow is how a character loses their face between shots.

Continuity is documentation, not memory

If you are producing more than a handful of clips with the same cast, write things down. A folder of renders with no notes becomes a folder you cannot reuse.

  • One identity block per character, stored as plain text, pasted rather than retyped.
  • A wardrobe list, one line per outfit, plus the scenes each outfit belongs to.
  • A light bible: time of day, key direction, color temperature, whether practical lights appear in frame.
  • A naming convention, for example mira_daylight_med_v03, so the version that worked stays findable.
  • A do-not-change note at the top of every scene file: props, hair state, injuries, weather.

That is admin work, and it is the difference between a character and a series of vaguely related strangers.

The anatomy of a two-character prompt

Instead of a paragraph of adjectives, build prompts in layers. Layers are reorderable, comparable, and debuggable, because you can change one without destroying the rest.

Layer Question it answers Example fragment
Identity Who is this, unmistakably? Mira, late twenties, shaved head, scar through left eyebrow, grey coverall
Intent What does she want right now? she is hiding the broken part behind her back
Blocking Where is everyone? Mira stands left facing right; Ilya sits at the bench, facing away
Camera How is it framed? medium shot, 35mm feel, slight low angle, slow push in
Continuity What must not drift? same coverall, same scar, no hat, warm tungsten from the left

Assembled, it should read like a shot note from a director rather than a poem:

Mira stands on the left in a grey coverall, shaved head, scar through the left eyebrow, weight on her back foot. Ilya stands on the right in a soaked navy coat, a head taller, hands still in his pockets. Twelve inches of air between them and neither is closing it. She looks at his hands, not his face. Medium shot, 35mm feel, cool daylight from behind him, slow push in. Keep the coverall, the scar, and the tungsten spill on the wall behind them.

Four things are doing the heavy lifting there: separate sentences per character, named sides, a stated distance, and a gaze target. Vague togetherness such as standing close to each other is how you get fused bodies. Numbers give you blocking. And she looks at his hands changes posture, jaw tension, and framing more than any adjective in the paragraph.

Blocking: the layer everyone skips

Blocking is where interaction prompts collapse, and it is also where the cheapest wins live. You do not need better prose. You need geometry.

Left, right, and measurable distance

Assign a side to each character and never change it inside a scene. If Mira is camera-left in the wide, she stays camera-left in the reverse, unless you cross the line deliberately for disorientation. Then express the gap in body terms: a hand's width, an arm's length, twelve inches, three paces. Distance is the least expensive way to encode tension, and it is far more reliable than writing that the mood is tense.

Eye lines are the emotional spine

Where people look tells the audience what the scene is about. Say it explicitly, and say what the look avoids: she looks at his shoes, not his face. Eye-line mismatch is also the most common two-character failure, so naming the target is a practical fix as much as an artistic decision. In sequences with entrances and exits, keep gaze direction consistent across cuts so the geography stays readable for the viewer.

Decide who owns the frame

In any real interaction, one character dominates the composition. Say which one. Mira fills two thirds of the frame, Ilya soft in the background produces a legible power dynamic. Two equally weighted people in a medium shot tend to produce a flat, symmetrical image that reads as a stock photo rather than a moment.

Engineer silhouette contrast

Two characters with similar builds, hair, and clothing will blend no matter how carefully you describe them. Build contrast instead: a head of height difference, one heavy coat against one bare arm, one still figure against one in motion. Silhouette contrast is cheaper than detail and survives low light, fast movement, and small frame sizes on a phone screen.

Dialogue without lip-sync roulette

Short AI clips are unreliable at precise speech, and fighting that burns renders. Design interactions where speech is implied instead:

  • Pre-speech. Mouth closed, chin lifted, about to speak.
  • Post-speech. Mouth closed, the listener already reacting.
  • Averted speech. The speaker turns away mid-line, so the mouth is never the subject.
  • Reaction shots. Cut to the listener; reaction carries more story than articulation.
  • Off-frame speech. The line arrives from outside the shot and only the listener is seen.

Write the line into the prompt as context, never as a demand: she has just said something she cannot take back, and his face is still processing it. The model renders the aftermath, which is both easier and better acted.

If you genuinely need a spoken exchange, generate the two-shot, then a separate close-up of the listener, and cut them together. Two reliable shots beat one ambitious one. When a mouth does need to move, keep the action small — a single syllable, a breath, a half-laugh — and accept a slightly stylized delivery.

Camera grammar that serves the interaction

Camera choices are not decoration in a two-hander. They carry the power dynamic when dialogue cannot.

Story beat Camera choice Why it works
Two people avoiding each other Wide, locked off, deep space Space does the acting
One person studying the other Medium close-up, shallow depth Isolates the observer
A power shift Move from low angle to high angle across the cut Encodes hierarchy without words
A reconciliation Slow push in on a two-shot Moves the audience toward them
A refusal Static frame, subject exits Stillness against motion

Lens feel language beats focal-length precision. A 35mm feel with slight wide distortion changes the relationship between two faces more than an 85mm feel with a compressed background, and both phrases are easier for a model to honor than a number. Keep movement restrained: one slow push, one lateral drift, one handheld settle. Pick one. If the camera orbits, the two characters rotate through each other and identities smear.

Plan in pairs, too. Every two-shot wants a reverse. Generate the wide for geography, then a shot and reverse-shot pair for the exchange. That triptych cuts like a scene even when each clip is only four or five seconds long. If you are still choosing a starting point for shot grammar, the video templates collection is a faster on-ramp than inventing framing language from nothing.

A worked scene in four shots

The scene: a reunion that should not be warm. Mira and Ilya, a cardboard box, a doorway at dusk.

Shot 1 — establish the asymmetry. Wide, static. Prompt: Mira stands on the left, small in frame, grey coverall, scar through the left eyebrow, holding a box against her chest. Ilya stands on the right, tall in a navy coat, backlit by the doorway. Twenty feet apart, neither steps forward. Cool daylight behind him, tungsten spill on the wall behind her. The check here is simple: are they the right height, on the right sides, and is the distance readable?

Shot 2 — the gaze. Medium close on Mira. Prompt: she looks at his shoes, not his face, jaw tight, the box shifting in her grip. No camera movement. The gaze target does the acting, which is why this shot either works instantly or fails visibly.

Shot 3 — the failed gesture. Medium two-shot. Prompt: Ilya lifts one hand as if to take the box, then stops it mid-air. Mira does not move. A blocked action is more interesting than a completed one, and it keeps both faces in frame for a beat without asking for complicated choreography.

Shot 4 — the aftermath. Close on his hand returning to his pocket. No faces, sound of rain, slight handheld drift. This is the shot that makes the sequence feel edited rather than generated, because it resolves on a detail instead of a reaction.

Four short, boring-to-describe shots produce one coherent scene. That is the trade: drama lives in the edit, not in a single heroic prompt.

Mistakes, decision criteria, and the loop that converges

Most bad two-character renders come from a small set of recurring causes.

Mistake What it produces Fix
Two characters in one clause Merged features, swapped clothing One sentence per character, fixed order
No spatial anchors Random placement, sides flipping between takes Name sides, heights, and a distance
Adjective stacking Over-styled, mushy frames Two or three precise visual facts
Contradictory light Washed-out or split lighting One named source and direction
Unnamed emotion Neutral mannequin energy State the cause of the feeling
Everything at once Busy frame with no subject One action, one dominant character
Identity paraphrased A different face every shot Paste the block verbatim

The last row is the most expensive and the easiest to fix. The everything-at-once row is the most common: a prompt asking for walking, talking, turning, and rain while the camera orbits has four chances to fail and no clear subject. Cut it in half.

Decision criteria before you re-render

  • Split the shot when two characters need to do two different things, or when they are more than a body length apart in importance.
  • Keep one shot when the whole point is co-presence — a standoff, a shared silence, a handshake.
  • Re-render when the composition and identities are right but the motion is awkward.
  • Rewrite the brief when the identities or blocking are wrong in every variant. That is a prompt problem, not a seed problem.
  • Change the wardrobe if two characters still blend after three attempts; the model is averaging similar colors, not ignoring you.

The iteration loop

  1. Write the moment in plain language first. One paragraph, no jargon. If you cannot explain it to a human, no prompt will rescue it.
  2. Translate into the five layers. Identity, intent, blocking, camera, continuity.
  3. Generate three variants, changing exactly one layer each. Changing everything teaches you nothing about what worked.
  4. Score against a fixed checklist rather than a feeling: recognizable identities, correct gaze, readable blocking, consistent light.
  5. Keep the failure notes. Knowing which phrasing breaks a model is as valuable as knowing which phrasing sings.

Three variants per round is enough. Beyond that you stop comparing and start hoping. Keep the winning prompts in a prompt library so the version that worked is never lost to a closed tab.

FAQ

Can a single prompt handle three or more characters?

It can, but expect drift. Above two characters, split into shots and lean on reaction framing. Three-person scenes work best when one character clearly dominates and the others hold background or edge positions, so the model has an obvious anchor.

Why does a character's face change between shots with the same prompt?

Usually because the identity block was paraphrased rather than repeated, or because the shot size changed dramatically. Copy the identity text exactly, keep the wardrobe identical, and change only the action and camera layers between shots.

How long should an interaction prompt be?

Long enough to cover all five layers, short enough that each sentence has one job. Most workable prompts land between 60 and 120 words. Past that, contradictions creep in and the model averages them into something bland.

Should I describe emotions or physical cues?

Both, in that order. Name the feeling, then give the body a task: she is embarrassed, so she keeps adjusting the strap of the bag. Emotion alone yields a blank face; physical cues alone yield an unexplained gesture.

What about negative prompts?

Use them for chronic, specific problems — duplicated limbs, text artifacts, extra fingers. Avoid using negatives to describe a style you dislike, because that often drags the unwanted idea into frame.

How many attempts before abandoning a prompt?

Three focused rounds with one variable changed each time. If it is still wrong, the brief is wrong rather than the seed. Rewrite the scene in plain language and translate it again from scratch.

Do I need reference images, or is text enough?

Text can hold a character for a short sequence, but a locked reference frame makes everything downstream cheaper — fewer re-rolls, faster identity matching, and a written description you can trust because you wrote it from what you actually see.

Put one character opposite one other person tonight

The temptation is to build an entire cast and a whole episode before proving a single shot. Resist it. Take one original character, write a ten-line identity block, stand them opposite one other person, and describe a moment where something is at stake and nothing is said. Render it. Then render the reaction.

That rhythm is what Orelon is built for — cinematic ideas in motion, one shot at a time. Lock the look in the AI image generator, animate it in the AI video generator with the identity block pasted verbatim, and keep your best phrasing in the prompt library so nothing good is ever rebuilt from memory. Your characters deserve to be recognized twice.