Why your world changes when your character does not

You held the face and the coat still moved. A teardown of wardrobe, props and set drift, and how a sequence is written to hold all three.

6 September 2026 · Written by Fares Walid Mansour, FWM

The face finally holds. Then the coat changes colour, the car changes model, and the kitchen grows a second door in shot nine.

The symptom

You solved the face. You wrote the anchors, you repeated them, and she is the same woman in all eleven shots. Then you watch it back and the world is loose.

The overcoat is charcoal in shot two and black in shot five. The car is a saloon, then a hatchback. The kitchen has one door until shot nine, when it has two and the window has moved.

This drift is quieter than face drift and it costs more, because it makes the film feel careless rather than broken. A viewer will forgive a strange cut. They will not believe a world that keeps rearranging itself.

The prompt that drifted

Shot 9, as written

Mira, 34, oval face, wide-set dark brown eyes, aquiline
nose, scar through the left eyebrow, dark coat, standing
in her kitchen, morning, cinematic

The identity clause is doing its job. Everything after it is a placeholder. "Dark coat", "her kitchen" and "morning" carry no information the model can reproduce twice.

"Her kitchen" is the worst offender. It refers to a shot you generated an hour ago, and the model has never seen it. Possessive references do not travel between generations.

Why it drifts

A sequence has three layers that each need holding, and most filmmakers only write the first. The character, the wardrobe and props, and the set itself.

Wardrobe drifts because colour words are broad. Charcoal, slate, graphite and black all satisfy "dark", and the model has no reason to prefer the one you meant. A garment needs a cut, a material and a state.

Props drift because they are usually named as categories: a car, a mug, a phone. Each is a family of thousands, and each generation is a fresh draw from the family.

Sets drift because architecture is generated per frame. Unless you fix the count and position of the openings, the model will add a door, move a window or change a wall's material whenever the camera turns.

The prompt that holds

Shot 9, rewritten

Medium wide, 35mm, chest height, static.

SUBJECT: [identity clause, pasted unchanged from shot 1]

WARDROBE: the same oversized charcoal wool overcoat,
double-breasted, collar up, one button missing at the
waist, black crew-neck beneath, one silver ring on the
right index finger. The coat is damp at the shoulders.

SET: a narrow galley kitchen, exactly one doorway on
frame right, one window above the sink on frame left,
pale green tiles to shoulder height, a chipped enamel
sink, a single bare bulb on a black cord, no upper
cabinets.

PROPS: one white enamel mug with a blue rim on the
counter, a folded newspaper, nothing else on the
surfaces.

BLOCKING: standing at the sink, back three-quarters to
camera, looking out of the window.

LIGHT: [key light block, pasted unchanged from shot 1]

NEGATIVE: no second doorway, no upper cabinets, no
extra props on the counter, no coat change, no jewellery
beyond the ring, no tidy kitchen.

"Exactly one doorway on frame right" is the sentence that saves the scene. A count and a position are reproducible. "Her kitchen" is not.

The anchors, named

Write the World once, at the top of the bible, and paste it into every shot. It is three blocks, and it does not change unless the story changes it.

Hold these

Wardrobe with cut, material, colour and one flaw: oversized charcoal wool overcoat, double-breasted, one button missing at the waist.
One carried prop, described once: the white enamel mug with a blue rim.
Set openings by count and position: exactly one doorway frame right, one window above the sink frame left.
Set materials at eye level: pale green tiles to shoulder height, chipped enamel sink.
One light source, held from the lighting block.
The state of things: damp shoulders, nothing else on the counter.

A flaw is worth more than a colour. One missing button, one chip in the enamel, one bulb on a bare cord: these are details the model cannot average, so they come back.

When the world must change, write the change as a beat and carry the rest. The coat comes off in shot ten, and the ring, the mug and the single doorway stay exactly as written.

The same anchors, four dialects

Midjourney: Wardrobe and set will not both fit in one short prompt, so pick the one the shot is about and hold the other with a style reference. Keep the flaw in the words, always, because a reference image will not enforce a missing button.
Sora: Describe the room once in full sentences and re-use that paragraph verbatim in every shot set there. A paraphrase is a new room, in the same way a paraphrase is a new face.
Veo: Separate wardrobe, set and props into their own labelled lines. Veo holds labelled fields far better than one long paragraph, and props are the first thing a paragraph loses.
Kling: Use a start frame that already contains the world and write only what moves. Never ask a clip to introduce a new prop mid-motion, because it will re-draw the whole surface to fit it.

What to do tonight

Write three blocks before you generate: subject, wardrobe and props, set.
Replace every possessive reference: "her kitchen" becomes the kitchen described in full.
Give each garment a cut, a material, a colour and one flaw.
Count the openings in the room and fix their positions relative to frame.
Name what is not there: no upper cabinets, no extra props, no second doorway.
Paste all three blocks into every shot, and write only the beat that changes.

A sequence that holds is not eleven good prompts. It is one written World and eleven shots that obey it, which is why the fix lives in the document rather than in the settings.